WO2009063463A2 - Pain monitoring using multidimensional analysis of physiological signals - Google Patents

Pain monitoring using multidimensional analysis of physiological signals Download PDF

Info

Publication number
WO2009063463A2
WO2009063463A2 PCT/IL2008/001493 IL2008001493W WO2009063463A2 WO 2009063463 A2 WO2009063463 A2 WO 2009063463A2 IL 2008001493 W IL2008001493 W IL 2008001493W WO 2009063463 A2 WO2009063463 A2 WO 2009063463A2
Authority
WO
WIPO (PCT)
Prior art keywords
features
pain
vector
patient
classifier
Prior art date
Application number
PCT/IL2008/001493
Other languages
French (fr)
Other versions
WO2009063463A3 (en
Inventor
Galit Zuckerman
Mark Kliger
Original Assignee
Medasense Biometrics Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Medasense Biometrics Ltd filed Critical Medasense Biometrics Ltd
Publication of WO2009063463A2 publication Critical patent/WO2009063463A2/en
Publication of WO2009063463A3 publication Critical patent/WO2009063463A3/en
Priority to US12/779,963 priority Critical patent/US8512240B1/en
Priority to US13/945,657 priority patent/US9498138B2/en
Priority to US15/349,098 priority patent/US10743778B2/en
Priority to US16/983,466 priority patent/US11259708B2/en

Links

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/0002—Remote monitoring of patients using telemetry, e.g. transmission of vital signals via a communication network
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/48—Other medical applications
    • A61B5/4824—Touch or pain perception evaluation
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/68—Arrangements of detecting, measuring or recording means, e.g. sensors, in relation to patient
    • A61B5/6801—Arrangements of detecting, measuring or recording means, e.g. sensors, in relation to patient specially adapted to be attached to or worn on the body surface
    • A61B5/6813—Specially adapted to be attached to a specific body part
    • A61B5/6825—Hand
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G06F18/24137—Distances to cluster centroïds
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/40—Detecting, measuring or recording for evaluating the nervous system
    • A61B5/4029—Detecting, measuring or recording for evaluating the nervous system for evaluating the peripheral nervous systems
    • A61B5/4035—Evaluating the autonomic nervous system
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7203—Signal processing specially adapted for physiological signals or for diagnostic purposes for noise prevention, reduction or removal
    • A61B5/7207—Signal processing specially adapted for physiological signals or for diagnostic purposes for noise prevention, reduction or removal of noise induced by motion artifacts
    • A61B5/721—Signal processing specially adapted for physiological signals or for diagnostic purposes for noise prevention, reduction or removal of noise induced by motion artifacts using a separate sensor to detect motion or using motion information derived from signals other than the physiological signal to be measured
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients

Definitions

  • the field of the present invention relates to medical diagnostic tools. More particularly, the field of the present invention relates to systems and methods relating to measuring and reporting a subject's pain.
  • Pain is an unpleasant sensory and emotional experience associated with actual or potential tissue damage, or described in terms of such damage.
  • the inability to communicate verbally does not negate the possibility that an individual is experiencing pain and is in need of appropriate pain-relieving treatment (www. iasp-pain.org/ AM/). Pain is always subjective. Each individual learns the application of the word through experiences related to injury in early life.. Biologists recognize that those stimuli which cause pain are liable to damage tissue. Accordingly, pain is that experience we associate with actual or potential tissue damage. It is unquestionably a sensation in a part or parts of the body, but it is also always unpleasant and therefore also an emotional experience. Experiences which resemble pain but are not unpleasant, e.g., pricking, should not be called pain. Unpleasant abnormal experiences (dysesthesia) may also be pain but are not necessarily so because, subjectively, they may not have the usual sensory qualities of pain.
  • Pain Threshold is defined as the least experience of pain which a subject can recognize as pain. Traditionally, this threshold has been defined as the least stimulus intensity at which a subject perceives pain. Properly defined, however, the threshold should be related to the experience of the patient, whereas the measured intensity of the stimulus is an external event. Because the threshold stimulus can be recognized as such and measured objectively, it. has been common usage for most pain research workers to define the threshold in terms of the stimulus, even though it is preferable to avoid such a definition. In psychophysics, a threshold is defined as the level at which 50% of stimuli are recognized. Thus, the pain threshold would be the level at which 50% of stimuli would be recognized as painful. As the stimulus is only one aspect of pain, it cannot be a measure or a definition of pain.
  • Pain Tolerance Level is defined as the greatest level of pain which a subject is prepared to tolerate. As with pain threshold, the pain tolerance level is the subjective experience of the individual. The stimuli which are normally measured in relation to its production are the pain tolerance level stimuli and not the level itself. Thus, the same argument applies to pain tolerance level as to pain threshold, and it should not be defined in terms of the external stimulation as such.
  • the lowest level is the stimulus itself
  • DOA Monitoring Depth of Anesthesia monitoring
  • pain monitoring are two fields that use sympathetic signals from the sympathetic nervous system or brain signals for monitoring a certain state of a patient.
  • DOA monitoring is a general term for pain and awareness monitoring when a patient is under general anesthesia, a state in which pain and awareness cannot be distinguished since they both result in the same physiological symptoms. Pain monitoring follows only the sensation of physical discomfort while the subject can be in any condition including fully awake.
  • the DOA monitoring field is already saturated, with 30% of the market share being held by Aspect Medical (ASPM)'s BIS-Bispectral index. • BIS analyzes the patients' electroencephalograms during general anesthesia. Other predominant companies are GE Healthcare with their Entropy analysis and former Physiometrix (currently Hospira) with their PSA - Patient State Analyzer. These products apply the 'awareness part' of DOA while using the electroencephalogram signals.
  • Aspect Medical Aspect Medical
  • United States Patent 6,685,649 to Korhonen discloses a method for monitoring a condition of a patient under anesthesia or sedation by acquiring and analyzing signals representing the cardiovascular activity of the patient.
  • the anesthesia indicator is calculated by analysis of the acceleration trends of the cardiovascular activity (interval/rate or pressure) which result in an index correlated to the level of anesthesia
  • This patent uses a basic hard decision rule on each of the acceleration trends values. This is inadequate since there are cross-connections between the parameters that affect the threshold values, and as was described above the heart rate or the blood pressure separately depends on various other sources rather than the status of the analgesia alone.
  • Patents EP 1,495,715 and US 7,367,949 to Korhonen also disclose "method and apparatus based on combination of three physiological parameters for assessment of analgesia during anesthesia or sedation.” The method includes utilizing brain activity signals in addition to analyzing the cardiovascular activity. This patent refers only to the sedated patients and therefore requires a smaller number of parameters, and simpler methods for classifying the level of patient comfort during anesthesia.
  • United States Patent 7,215,994 to Huiku discloses a method for monitoring a state of anesthesia or sedation by comparing cortex related EEG biopotential signal data from the patient to subcortex-related biosignal data from the patient, the subcortex-related biosignal data including at least bioimpedance signal data.
  • Pain monitoring is more complicated, especially in patients who are fully awake, since the emotional state of the patient, the medication status, the environmental context and highly variable, often culturally determined, behavioral responses provide many signals that seriously affect specificity and sensitivity. Many more parameters therefore need to be measured and processed to achieve a meaningful pain monitoring method. Moreover, compared to anesthesia monitoring which is used only in operating rooms or in ICU when the patient is immobilized, not influenced from external stimuli and well controlled, pain monitoring needs to give pain indication in variant scenarios including when the patient might be in a movement and responsive to external stimuli. This scenario is far more complicated, and might require more information to be processed and more robust methods to handle the vast of information.
  • An unmet and long felt need remains to provide new algorithms for dealing with a large number of patient related parameters along with methods and devices to enable pain monitoring in subjects when awake, unanesthetized and unsedated.
  • a far more robust and comprehensive solution is needed.
  • a further long felt need is to enable pain detection and measurement in normal and impaired people under certain procedures in order to avoid unnecessary pain under surgical operations or medical procedures.
  • a further long felt need is to enable detection and measurement of sensation when this information is needed for a successful operation.
  • a further long felt need is to provide a method to differentiate between pain that is correlated to the stimuli and pain that is related to memory of pain so as to perform better diagnosis and treatment for pain clinics' patients.
  • a further long felt need is to provide a method to objectively measure pain as a function of given stimuli to neurological diagnosis , and to provide matching of applied sensors data with the pain level of stimulus input and a priori known data on the patient.
  • the combination of controlled stimuli with exact measurements will enable the objective pain measurement thereby fulfilling an unmet need.
  • the step of analyzing comprises steps of: a. acquiring a set of physiological signals from the body of a patient; b. processing the set of signals so as to extract a first vector of features representing the physiological status of the patient; wherein the first vector of features comprises a Great Plurality of Features (GPF); c. reducing of the dimensionality of the first vector of features by transforming the first vector to a second vector whose dimensions are lower by at least one order of magnitude compared to the first vector; d. classifying said second vector of features into at least two classes representing at least two conditions of pain; e. representing said classes of said pain level of said patient at a given time interval thereby establishing the pain level in an awake, semi- awake or sedated patient.
  • GPS Great Plurality of Features
  • the method additionally comprises a training step; said training step comprising; a. acquiring said set of physiological signals from said body of a patient or group of patients in a first non-pain state and a second pain state; b. processing said set of signals so as to extract a first vector of features representing the physiological status of said patient; wherein said first vector of features comprises a Great Plurality of Features (GPF); c. firstly, learning the parameters of said reducing of dimensionality by learning parameters of transformation of said first vector of features to a second vector whose dimensions are lower by at least one order of magnitude compared to said first vector; d.
  • GPF Great Plurality of Features
  • the step of acquiring comprises selecting data from the group consisting of data supplied by the physician's, environmental parameters, patient parameters or any combination thereof It is a further object of the invention to disclose the abovementioned method wherein the method wherein said extracted features are selected from Table 2. It is a further object of the invention to disclose the abovementinoed method wherein said physiological signals represent an activity selected from the group consisting of autonomic nervous system activity, muscular activity, and brain activity.
  • physiological signals are selected from a group comprising ECG, PPG, continuous blood pressure, respiration, internal or skin temperature, EOG or pupil diameter monitoring, GSR, EEG, EMG, EGG, LDV, capnograph and accelerometer or any combination thereof
  • step of processing comprises analyzing the artifacts occurrence in said acquired signals. It is a further object of the invention to disclose the abovementinoed method wherein said step of representing said pain level of said patient is provided continuously during at least one predetermined time interval. It is a further object of the invention to disclose the abovementinoed method wherein said step of representing a PAIN or NON-PAIN condition of said patient is provided in a graduated scale.
  • steps of reducing dimensionality or learning the parameters of reduction of dimensionality of said first vector of features further comprises the steps of: a. calculating extracted feature scores for each of said features or combination of features; b. filtering out said extracted low-score features thereby decreasing the number of said features to a predetermined number;
  • processing means for processing said set of signals so as to extract a first vector of features representing the physiological status of the patient; wherein the said first vector of features comprises a Great Plurality of Features (GPF), ii. reducing means for reducing the dimensionality of the said first vector of features by transforming to a second vector whose dimensions are lower by at least one order of magnitude compared to said first vector; iii. classifying means for classifying said second vector of features into at least two classes representing at least two conditions of pain.
  • GPF Great Plurality of Features
  • reducing means for reducing the dimensionality of the said first vector of features by transforming to a second vector whose dimensions are lower by at least one order of magnitude compared to said first vector
  • classifying means for classifying said second vector of features into at least two classes representing at least two conditions of pain.
  • processing means for processing said set of signals so as to extract a first vector of features representing the physiological status of the patient; wherein the said first vector of features comprises a Great Plurality of Features (GPF), ii. first learning means for training said reducing of dimensionality by learning parameters of transformation of the said first vector of features to a second vector whose dimensions are lower by at least one order of magnitude compared to said first vector iii. second learning means for learning the parameters of classifier that classifying said second vector of features into at least two classes representing at least two conditions of pain iv. setting means for setting the parameters of said classifier thereby establishing a classifier which classifies said pain level in an awake, semi- awake or sedated patient.
  • GPF Great Plurality of Features
  • said acquiring means comprises sensors attached to the body of said patient for detecting said physiological signals.
  • said representing means is selected from the group consisting of computer screen, PDA screen, TV screen , plasma screen, LCD screen, patient monitor or any means for displaying numbers or graphs in a continuous manner
  • GSR GSR, EEG, EMG, EGG, LDV, capnograph and accelerometer or any combination thereof.
  • classifying means or second learning means are adapted to apply statistical methods selected from the group consisting of Boosting, Linear classifier, Na ⁇ ve Bayes Classifier, k-nearest neighbor classifier, QDA classifier, RBF classifier, Multilayer Perceptron classifier, Bayesian Network classifier, Bagging classifier, SVM, NC, NCS, LDA, SCRLDA, Random Forest, or Committee of classifiers or any combination thereof
  • classifying means or second learning means are adapted to compute a confidence value of said vector.
  • Figure 1 Pain Monitoring - System Description
  • Figure 2 Pain Monitoring Sensors
  • Figure 3 Optional configuration of the pain monitoring system
  • Figure 4 Flow diagram of the pain monitoring system
  • Figure 5 ECG signal and its parameters
  • Figure 6 Blood pressure or PPG signal and its parameters
  • Figure 7 Parameter 'A' values on pain and not pain with two population
  • Figure 8 Parameter 'A 1 values on pain and not pain with two populations separated with parameter 'B' on z axis
  • Figure 10 FIRV from PPG, HRV from ECG and pain/no pain reports as function of time elapsed
  • Figure 11 FIRV from PPG, FIRV from ECG values when pain or no pain is reported
  • Figure 12 HRV from PPG, HRV from ECG values when pain or no pain is reported - linear classifier possible separation
  • Figure 13 Features reduction from 2 features (HRV from PPG, HRV from ECG) to one combined feature
  • Figure 15 HRV from PPG, HRV from ECG and R-R interval when pain or no pain is reported - 3D visualization
  • Figure 16 HRV from PPG, HRV from ECG and R-R interval when pain or no pain is reported - with a linear classifier possible separation
  • Huiku presents a pain monitoring that is based on one or more physiological parameter that are measured, normalized and then compared to 'a threshold surface'.
  • the rate (number of occurrences per unit of time) of crossing this threshold is considered the pain level.
  • the way to obtain the threshold which is the core of the classification, and therefore the most important part of it, is not specified
  • pain monitoring needs to give pain indication in variant scenarios including when the patient might be in motion and responsive to external stimuli. This scenario is far more complicated, and requires more information to be processed and more robust methods to handle the vast of information.
  • the method further describes the methods for feature selection, reduction of dimensionality of a feature space, and classification of pain level.
  • the heart of any pain/analgesia monitoring system is a classification algorithm for taxonomy of patterns founded in physiological signals into classes of different pain level.
  • classification algorithm is a mathematical engine which receives as an input a multidimensional vector of normalized features extracted from multiple physiological signals.
  • the algorithm receives (if exist) patient parameters (age, gender, weight, chronic diseases, historical measurements of physiological features, etc.), input from physician (diagnosis, receiving medicine, etc.) and environmental parameters (time, place, room temperature, accelerometer data, etc.).
  • the output of the algorithm is a number which symbolizing a strength or level of pain of patient.
  • a Training set denotes a data .available from a variety of sources: publicly available databases, records of proprietary clinical trials, on site recorded data from the same patient or group of patients etc.
  • the training set must be comprised from an input and output signals.
  • the input signal has to be similar to the expected input of a pain classifier, i.e. multiple physiological signals, input from physician, environmental parameters and patient parameters.
  • the output signal has to be similar to the expected output from a pain classifier, i.e. strength or level of pain of patient.
  • the training output signal is determined by the human operator (physician or other skilled personal) during a clinical trial with controlled pain stimuli.
  • Training a pain classifier on a training set means determining ("learning") the pain classifier parameters which will allow classifying of previously unseen input data (not from the training set) with sensitivity and specificity similar to or better than the performances of a human operator.
  • Patient related parameters, environmental parameters and input from physician may significantly improve performances of the pain classifier. They are essential in certain instances, for example in cases when a usage of medication affects the physiological response of the ANS to pain. However it may not influence the result when the information is less related to the physiological response to pain.
  • features Prior to input into the pain classifier, features should be appropriately normalized in order to remove patient baseline mean or/and normalize baseline variability or/and identify and remove outlier samples or/and normalize the features distribution etc.
  • a method of histogram normalization of feature probability distribution was proposed. The same normalization was proposed to all features in consideration. However, such normalization might eliminate valuable information which is often hidden in a shape of feature probability distribution.
  • the number of input features might be very large (hundreds or even thousands of features), they also should be preprocessed for feature selection and/or dimensionality reduction.
  • the Great Plurality of Features may be extracted from the received multiple physiological signals. Some of the features are directly related to painful response and some in an indirect way be related to painful response, or might be used to prevent misclassification. For example, a raise in blood pressure can result from painful stimuli, but also due to a change in position from sitting to standing which can be identified in the accelerometer. Some features are differentially expressed due to noise or inter patient variability.
  • Feature selection picking a subset of original features
  • Feature selection is a "straightforward" approach for dimensionality reduction problem.
  • Feature Selection i.e. picking a subset
  • “Dimensionality reduction” i.e. creating new features which are referred as 'meta features'.
  • Feature selection algorithms typically fall into two categories; Feature Ranking and Subset Selection.
  • Feature Ranking ranks the features by a specific metric, e.g. correlation with strength of pain stimuli, and eliminates all features that do not achieve an adequate score. However, simple ranking might eliminate some important features which by themselves are non good discriminants, but in combination with other features can play a vital role in a success of classification task.
  • Subset Selection searches the set of possible features for the optimal subset and evaluates a subset of features as a group for suitability.
  • Subset selection algorithms can be broken into Wrappers, Filters and Embedded ( Kohavi and John 1997 ). The Filter approach attempts to assess the quality of subset of features from the data ignoring specific classification algorithm.
  • the best subset of features is chosen to suite specific classification procedures.
  • Wrapper uses a search algorithm to search through the space of possible features and evaluate each subset by running a classifier on the subset. Wrappers can be computationally expensive and have a risk of over fitting to the model.
  • Embedded techniques are embedded in and specific to a classifier. Embedded methods will be mentioned later in section dedicated to classification.
  • the search for the best features set might be performed more than once in the system development lifetime, and may be used as a research tool for the physician. Therefore the correct automatic feature selection or similar procedure should be part of the system.
  • FLDA Fisher Linear Discriminant Analysis
  • the Great Plurality of Features extracted from physiological signals, environmental parameters and prior information after preprocessing by feature selection and/or dimensionality reduction algorithm are the input into the pain level classifier.
  • the pain classifier In order to design the pain classifier one should first determine the structure of the learning function and corresponding learning algorithm. There are many possible algorithms and approaches to choose from. For example, in US 7,367,949 B2 Korhonen et al. inventors specifically propose to use either Decision Tree classifier ("rule based reasoning") or Logistic Regression classifier. However, it is well kno ⁇ vn that in a case when number of input features is large, the logistic regression classifier suffers from so- called "curse of dimensionality" - exponentially grows of complexity of classifier training phase.
  • the present invention discloses a system that includes a plurality of sensors for acquisition of physiological signals that indicate sympathetic activity, parasympathetic activity brain activity, muscular activity, movements, environment parameters and prior information on the patient,
  • ECG and PPG sensors acquire sympathetic/parasympathetic signals and EEG and EMG sensors acquire brain and muscular activity signals.
  • Another component of the system disclosed herein is the processing unit, designed to process the signals in order to present them as features.
  • a further element of the invention disclosed herein is the feature extraction for extracting and filtering features describing the subject pain state.
  • the system further comprises artificial intelligent elements for defining the subject pain level and an output unit to present the results.
  • the present invention discloses a novel system for pain monitoring, which combines parameters derived from many sensors such as ECG, PPG, EGG, Laser Doppler Velocimetry, skin conductance measurements, blood pressure measurements, capnograph peripheral, internal temperature measurements, respiration measurements, PD (pupil diameter) monitors, EOG, EEG and EMG 5 movements of the patient from the accelerometer, and prior information on the patient.
  • sensors such as ECG, PPG, EGG, Laser Doppler Velocimetry, skin conductance measurements, blood pressure measurements, capnograph peripheral, internal temperature measurements, respiration measurements, PD (pupil diameter) monitors, EOG, EEG and EMG 5 movements of the patient from the accelerometer, and prior information on the patient.
  • the present invention discloses methods for dimensionality reduction and classification in order to deal with the large amount of information and parameters.
  • Fig. 1 Pain Monitoring - System Description that schematically represents embodiment of the invention wherein steps in the method for monitoring pain are depicted.
  • the patient is connected to appropriate sensors for a plurality of signal parameter acquisitions selected from a group comprising ECG, PPG, continuous blood pressure, respiration, internal or skin temperature, EOG or pupil diameter monitoring, GSR, EEG, EMG, EGG and accelerometer (1).
  • the acquired signals are then processed by a microprocessor ⁇ vhich performs siRnal pre-processing such as de-noising, filtering, and other functions used for clearing and preparing the signal for feature extraction (2).
  • siRnal pre-processing such as de-noising, filtering, and other functions used for clearing and preparing the signal for feature extraction
  • GSR Galvanic skin response
  • EDR electrodermal response
  • SCR skin conductance response
  • GSR is conducted by attaching two or three leads to the skin, and acquiring a base measure.
  • a base measure When an outgoing sympathetic nervous burst occurs, a wave of skin conductance will follow.
  • spontaneous skin conductance changes increased number and amplitude of the waves is interpreted as increased activity in this part of the sympathetic nervous system (Lidberg and Wallin 1981)
  • Electrogastrography is a noninvasive method for the measurement of gastric myoelectrical activity using abdominal surface electrodes
  • An electrogastrogram is similar in principle to an electrocardiogram (ECG) in that sensors on the skin detect electrical signals indicative of muscular activity within. Where the electrocardiogram detects muscular activity in various regions of the heart, the electrogastrogram detects the wave-like contractions of the stomach
  • Pupil size and movement can be measured by either infrared videography or computerized pupillometry.
  • Pupillometry has been used in a research setting to study the autonomic nervous system, drug metabolism, pain responses, psychology, fatigue and sleep disorders.
  • Infrared videography is used in order to detect magnified movement of both pupils at the same time. The technique allows the pupils to be visualized in the dark.
  • Infrared videography takes advantage of dark pigmentation, since melanin actually reflects the infrared light shone on to the iris. Therefore, pigmented irises appear light on the screen, and the black pupils stand out in contrast to the surrounding light-appearing iris.
  • Computerized pupillometry can record pupil size and movement in both the light and the dark. The typical instrument captures several frames per second over several seconds, and then averages the measurements. Such averaging compensates for the highly variable pupillary response that changes second to second.
  • Electromyography EMG
  • EMG is a technique for evaluating and recording physiologic properties of muscles at rest and while contracting. EMG is performed using an electromyography to produce a record called an electromyogram. An electromyograph detects the electrical potential generated by muscle cells when these cells contract, and also when the cells are at rest. A surface electrode may be used to monitor the general picture of muscle activation.
  • a motor unit is defined as one motor neuron and all of the muscle fibers- it innervates. When a motor unit fires, the impulse (called an action potential) is carried down the motor neuron to the muscle. The area where the nerve contacts the muscle is called the neuromuscular junction, or the motor end plate.
  • EMG signals are essentially made up of superimposed motor unit action potentials (MUAPs) from several motor units. For a thorough analysis, the measured EMG signals can be decomposed into their constituent MUAPs.
  • MUAPs from different motor units tend to have different characteristic shapes, while MUAPs recorded by the same electrode from the same motor unit are typically similar. Notably MUAP size and shape depend on where the electrode is located with respect to the fibers and so can appear to be different if the electrode moves position. EMG decomposition is non-trivial, although many methods have been proposed. Frontalis (scalp) electromyogram (FEMG) •
  • the frontalis muscle receives both visceral and somatic fibres from the facial nerve.
  • the dual nerve supply means that this muscle can be influenced by autonomic activity.
  • Two surface electrodes record compound action potentials from this muscle.
  • the amplitude of the EMG decreases with increasing depth of anaesthesia, but this cannot be used in the paralysed ⁇ patient.
  • FEMG has the advantages of being non-invasive and convenient, and it is easy to apply the electrodes.
  • EEG electroencephalography
  • PPG Photoplethysmograph
  • a finger photoplethysmograph is a non-invasive transducer to measure the relative changes of blood volume in a subject's .finger.
  • Photoplethysmography is based on the determination of the optical properties of a selected skin area. For this purpose non- visible infrared light is emitted into the skin. More or less light is absorbed, depending on the blood volume in the skin. Consequently, the backscattered light corresponds with the variation of the blood volume.
  • Blood volume changes can then be determined by measuring the reflected light and using the optical properties of tissue and blood. The measured signal records venous blood volume changes as well as the arterial blood pulsation in the arterioles.
  • the relative change in blood volume reflects also the cardiovascular system activity and is controlled by the ANS. It has been suggested to be used as part of anesthesia monitor.
  • Electrocardiogram ECG
  • An electrocardiogram is a graphic produced by an electrocardiograph, which records the electrical activity of the heart over time. Electrical impulses in the heart originate in the sinoatrial node and travel through the heart muscle where they cause contraction. The electrical waves can be measured at selectively placed electrodes (electrical contacts) on the skin. Electrodes on different sides of the heart measure the activity of different parts of the heart muscle. An ECG displays the voltage between pairs of these electrodes, and the muscle activity that they measure, from different directions, also understood as vectors. This display indicates the overall rhythm of the heart, and weaknesses in different parts of the heart muscle.
  • Electroencephalography ECG
  • Electroencephalography is the measurement of electrical activity produced by the brain as recorded from electrodes placed on the scalp.
  • Scalp EEG measures the summed activity of post-synaptic currents.
  • An action potential in a pre-synaptic axon causes the release of a neurotransmitter into the synapse that diffuses across the synaptic cleft and binds to receptors in a post-synaptic dendrite, resulting in a flow of ions into or out of the dendrite, which in turn results in compensatory currents in the extracellular space. It is these extracellular currents that generate EEG voltages.
  • EEG to determine the activity within a single dendrite or neuron. Rather, a surface EEG reading is the summation of the synchronous activity of thousands of neurons that have similar spatial orientation, radial to the scalp. Currents that are tangential to the scalp are not picked up by the EEG. The EEG therefore benefits from the parallel, radial arrangement of apical dendrites in the cortex. Because voltage fields fall off with the fourth power of the radius, activity from deep sources is more difficult to detect than currents near the skull.
  • Scalp EEG activity oscillates at multiple frequencies having different characteristic spatial distributions associated with different states of brain functioning such as waking and sleeping. These oscillations represent synchronized activity over a network of neurons. The neuronal networks underlying some of these oscillations are understood while many others are not.
  • temperature sensor is considered to be affected by a combination of different physiological processes, e.g., perspiration and vascular tone, which determine the response of thermoregulation. Temperature variations were thought to -reflect changes in sympathetic vasoconstrictive tone and in concentration of circulating vasoactive substances occurring during both relaxation and stress ( Guyton 1982 ). Changes of arterioles' smooth muscle tone regulated by the sympathetic nervous system are considered to be one of the origins of these temperature fluctuations ( Cohen and Sherman 1983 ). The fluctuation of this myogenic activity and its effect on the skin microcirculation has been studied by different methods( Fagrell 1984 ).
  • the respiration transducer directly measures the respiratory effort.
  • the transducer measures the changes in thoracic or abdominal circumference that occur as the subject breathes.
  • the design presents minimal resistance to movement and is extremely unobtrusive.
  • the transducer can measure arbitrarily slow to very fast respiration patterns with no loss in signal amplitude, while maintaining excellent linearity and minimal hysteresis.
  • Blood pressure refers to the force exerted by circulating blood on the walls of blood vessels, and constitutes one of the principal vital signs.
  • the term blood pressure generally refers to arterial pressure, i.e., the pressure in the larger arteries.
  • the systolic arterial pressure is defined as the peak pressure in the arteries, which occurs near the beginning of the cardiac cycle; the diastolic arterial pressure is the lowest pressure (at the resting phase of the cardiac cycle).
  • the average pressure throughout the cardiac cycle is reported as mean arterial pressure; the pulse pressure reflects the difference between the maximum and minimum pressures measured.
  • Rate of pumping the rate at which blood is pumped by the heart. The higher the heart rate, the higher the arterial pressure. • Volume of fluid or blood volume, the amount of blood that is present in the body. The more blood present in the body, the higher the rate of blood returns to the heart and the resulting cardiac output.
  • Resistance In the circulatory system, this is the resistance of the blood vessels. The higher the resistance, the higher the arterial pressure. Resistance is related to size (the larger the blood vessel, the lower the resistance), as well as the smoothness of the blood vessel walls.
  • each individual's autonomic nervous system responds to and regulates all these interacting factors so that, although the above issues are important, the actual arterial pressure response of a given individual varies widely because of both split-second and slow-moving responses of the nervous system and end organs.
  • the haemodynamic responses have been shown after noxious stimulation such as laryngoscopy or tracheal intubation. ( van den Berg, Sawa and Honjol 2006 )
  • the laser Doppler quantifies blood flow in human tissues such as skin and by that evaluates the skin vasomotor reflex (SVMR),
  • a monochromatic laser beam is directed at the skin surface.
  • Light that is reflected off stationary tissue undergoes no shift whilst light that is reflected off cells with velocity (like red blood cells) undergoes Doppler shift.
  • the degree of Doppler shift is proportional to the velocity of the cell into which it collided. This light is randomly reflected back out of the tissue and onto a photodetector which calculates the average velocity of cells within the tissue.
  • a capnograph is an instrument used to monitor the concentration or partial pressure of carbon dioxide (CO2) in the respiratory gases. It is usually presented as a graph of expiratory CO2 plotted against time, or, less commonly, but more usefully, expired volume. When expired CO2 is related to expired volume rather than time, the area beneath the curve represents the volume of CO2 in the breath, and thus over the course of a minute, this method can yield the CO2 minute elimination . , an important measure of metabolism. Sudden changes in CO2 elimination during lung or heart surgery usually imply important changes in cardiorespiratory function. During procedures done under sedation, capnography provides more useful information than pulse oximetry. Capnographs usually work on the principle that CO2 absorbs infra-red radiation. A beam of infra-red light is passed across the gas sample to fall on to a sensor. The presence of CO2 in the gas leads to a reduction in the amount of light falling on the sensor, which changes the voltage in a circuit. Acceleronieter
  • An accelerometer is a device for measuring acceleration and gravity induced reaction forces. Single and multi-axis models are available to detect magnitude and direction of the acceleration as a vector quantity. An accelerometer measures the acceleration and gravity it experiences. Both are typically expressed in SI units meters/second2 (m/s2) or popularly in terms of g-force. For the practical purpose of finding the acceleration of objects with respect to the earth, the correction due to gravity along the vertical axis is usually made automatically, e.g. by calibrating the device at rest.
  • Fig. 2 Pain Monitoring - Sensors.
  • Fig .2 is an example of some of the possible sensors which can be used in embodiments of the invention, and is intended to illustrate the present invention but should not be interpreted as a limitation upon the reasonable scope thereof.
  • the signal acquisition is performed by collecting data from some or all of the following noninvasive sensors.
  • the following sensors are preferably located on the patients' hand fingers and wrist (see Fig. 2):
  • ECG - PQRST signal (see Fig. 5) characterizes the heart activity
  • PPG waveform (from the pulse Oximeter) characterizes blood volume pulse (BVP)
  • PD Pupil Diameter - Infrared videography or computerized pupillometry - measuring the pupil size (dilation or erosion of pupil size can indicate on increase in sympathetic or parasympathetic activity accordingly)
  • the data thus acquired are decoded and transmitted via a transmission device to the local computer (monitor) that can present the data and perform all the data processing needed to compute the patient's pain level.
  • an accelerometer sensor is located near to other sensors (located on the patient's hand). This sensor does not acquire biological signals, but rather -signals generated by hand movements of the patient; such data is used to eliminate artifacts such as unexpected movements from signals due to sensitivity of the other sensors.
  • Fig. 3 schematically illustrating examples for possible configuration of the pain monitoring a.
  • the system includes processing means, data is acquired and represented by external system b.
  • the system includes acquiring means, processing means and displaying means, still some of the signal can be acquired by external means and the data can be also represented on external means. Communication between the sensors (acquiring means), the processing means and the displaying means can be wireless c.
  • the system includes acquiring means and processing means, still some of the signal can be acquired by external means, and the data is represented on external means d.
  • the system includes acquiring means, processing means and displaying means. Communication between the sensors (acquiring means), the processing means and the displaying means can be wireless.
  • the processing and displaying means can be on personal digital assistant (PDAs) e.
  • PDAs personal digital assistant
  • the system can be a stand alone systemj and includes all acquiring (physiological signal, and data input), processing and displaying means
  • FIG. 4 flow diagram of the pain monitoring system.
  • Fig. 4 is a flow diagram illustrating the method herein disclosed for pain monitoring using multidimensional analysis of physiological signals as follows;
  • Step 1 comprises acquiring the signal from the sensors.
  • Step 2 comprises separating out the artifacts in the signal from the signal of interest and defining the time resolution.
  • Step 3 Extracting features from the received signals.
  • steps 4-5 in Fig. 4 used for classification of features in the patient or subject to be monitored for pain
  • steps 4'-5' in Fig. 4 are used for forming a computer generated 'learning" or "training" profile from -which feature classifications are selected as follows: Learning or Training Steps:
  • Step 4' Processing vector of features by feature selection and/or dimensionality reduction methods in order to lower the dimension of the aforementioned vector.
  • Step 5' Classifier is trained on a set of training examples of manually labeled signals and classifier parameters are learned.
  • Step 4 Applying Feature selection and/or dimensionality reduction according to the learned parameters and function in the learning step
  • Step 5 Classification of the pain level according to classifier parameters learned in steps 4 -5'.
  • Steps 6 and 7 Presenting the pain level periodically both in training/classification; the refresh rate is defined according to the resolution time selected.
  • feature refers to an acquired signal describing a certain behavior of the signal, e.g. the amplitude of a band of frequency.
  • Some signals have many features, e.g. ECG can have over 50 because its structure includes patterns that have physiological significance, whereas other features have less, e.g. temperature, which does not have a consistent pattern and therefore can have less features.
  • artifact refers herein to unexpected behavior of physiological signals which may appear often due to local or global motion of the aware subject, for example in EEG this can be eyes blinking, in PPG the finger movement or change of body position etc.. In the present invention these artifacts may also be analyzed and machine interpreted to assist arriving at the correct pain classification in pain monitoring.
  • Fig. 5 schematically illustrating an ECG signal and its parameters as an example for one signal and description of some of its features.
  • Fig. 6 schematically illustrating a blood pressure or pulseplethysmograph signal and its parameters as an example for one signal and description of some of its features.
  • EGG Electrogastrogram - decrease in digestion activity Signals derived from the following are related to pain
  • Entropy Entropy is related to the amount of disorder, complexity, or unpredictability of the system. It is a property of a physical system or data string consisting of a great number of elements. The concept is used in physical sciences and information theory. By adding the measurement of the cortical electrical activity, the clinician can assess the effect of anesthetics more comprehensively. EEG recordings change from irregular to more regular patterns when anesthesia deepens. Similarly, FEMG quiets down as the deeper parts of the brain are increasingly saturated with anesthetics. Entropy measures the irregularity of EEG and FEMG signals.
  • the entropy of the EEG signal within a certain time window can be calculated from the signal itself or its spectrum. Entropy of the signal has been shown to drop when a patient falls asleep and increase again ⁇ vhen the patient wakes up.
  • the EEG frequency range is from about 0.5 Hz to 40 Hz depending on the state of mind a person is in, and the EMG frequency band is from about 20 Hz up to about 80 Hz
  • the EEG frequency domain can be further divided into the following frequency ranges that describe the following states of mind (see table below):
  • EEG and other neuron signals originating from the cortex hitherto have only been useful for DOA analysis. This is because during general anesthesia the functional activity of the neurons is decreased and synchronized and become more ordered and predictable. In awake subjects, the electrical activity of all cortical neurons working independently and recorded by the EEG results in a random, a-periodical and unpredictable signal, so they would not be useful on their own for pain monitoring.
  • Heart rate variability can be extracted from the signal(s) represent(s) the cardiovascular activity (ECG and PPG, Continuous BP).
  • the - sympathetic activity is measured (it increases in the case of pain) and the parasympathetic activity measured (it decreases in case of pain).
  • the aforementioned activities ratio can be extracted from the HRV and the peripheral blood pressure ( Deschamps, et al. 2004 ).
  • the high frequency (HF) peak located around the respiratory frequency, typically between 0.15-0.4 Hz, reflects primarily parasympathetic activity ( Akselrod, et al. 1981 )
  • the low frequency peak (LF) centered on 0.1 Hz content of HR fluctuations is an estimate of combined vagal and sympathetic activity ( Malik 1996 ).
  • the LF content fluctuations is an estimate of sympathetic activity ( Pagani, Rimoldi and Malliani 1992 ).
  • VLF very low frequency
  • the frequencies of 0.08-0.15 are defined as Medium Frequency (MF) and represent manly barrorecptor activity.
  • Table 2 List of possible extracted features Any subset of the features described above (signal, spectra, wavelets and statistical values and relations) can be combined into an N-dimensional vector which represents the patient's pain state at a certain time. Normalization per patient
  • each feature in a vector of features may be normalized in order to remove patient baseline mean or/and normalize patient baseline variability or/and identify and remove outliers or/and normalize the features distribution etc.
  • Normalization hereafter denotes any data normalization method known in the art. The normalization is performed with respect to either baseline record of a patient (removing baseline mean, normalizing variance, etc.) or training set (outliers detection, distribution normalization, etc.) or both.
  • the baseline may be recorded in the first few minutes, when the patient is in a constant position similar to the position of the treatment, with no pain stimuli. Alternatively, the baseline may be recorded at the first minutes, when the patient is in a constant position similar to the position of the treatment, and when a minimal pain stimulus such as infusion penetration has occurred.
  • Baseline records may be also obtained from a patient historical records if exist
  • the normalization is feature-specific, i.e. different features may be normalized in a different manner.
  • the normalization of a specific feature may be independent, i.e. performed for each feature independently from the other features, or the normalizations of a specific feature may depend on other features
  • the feature normalization which removes the patient's baseline feature mean is carried out, if needed, in the following manner: where X 1 -is the current feature that is processed and avg ⁇ X ⁇ aselme ) is the average of the feature values of the patient baseline record.
  • the feature normalization which normalize the patient's baseline feature variability is carried out, if needed, in the following manner: std(x; b sehm ) where std(XT''"') is the standard deviation of the feature values of the patient baseline record.
  • the features normalization which normalizes the feature value into a value between [0,1] is carried out in the following manner: max(X t ' rammg ) - min(X; ra "" ns ) where VUaX(X""'"'” 8 ) and mm(X" ammg ) are the maximum or the minimum values of the feature in the training set, respectively.
  • This normalization is prone to be affected by outliers in the training data set and can results in unreasonable max(X" a ' n '" g ) or min(Z; ra """ «) values.
  • a normalization which normalizes the feature value into a value between [0,1] is carried out in the following manner: where a is a factor used so that most of the population will be in the output range. For example, if the distribution is normal then the 2*STD value represents 68%, 4*STD represents more than 95% of the population and 6*STD represents -99% of the population. Therefore setting the value of a can increase/decrease the percentage of the extreme samples that are excluded from consideration. The extreme samples are declared as outliers and their values are set to be 0 or 1. Another possibility is to omit the outliers from consideration and to treat them as missing samples.
  • the autonomic tone is correlated with the conceptual pain (and might be also some of the context relevance). Since the subject that uses the pain monitoring system might not be a 'normal subject' in one of many terms, all relevant information that can affect the autonomic tone and known by the care provider, should be entered as parameters to the system. Some of these parameters are categorical: nominal (gender, type of medicine, diagnostics, etc.), ordinal (patient condition, patient definition of pain level, etc.), interval (age group etc.). Other parameters are numerical (weight, height, historical features data, etc.). The parameters might be continuous or discrete, quantitative or qualitative. These parameters are of high importance since for example, use of beta-blockers cause degradation of the sympathetic response, therefore, even though a pain stimulus has occurred, which usually causes high sympathetic tone in normal subjects, the sympathetic tone of a subject using beta-blockers does not change significantly.
  • the pain levels can be weighted. E.g., level 3 for a normal subject is considered as level 6 for subject with certain disorder or usage drug.
  • the system is trained to differ between the population by using feature that describe the prior information (as depicted in Fig 7 and Fig. 8)
  • PCA Principal Component Analysis
  • Sparsity of a principal component might significantly improve interpreterability of the result and provides valuable insights for physician.
  • Regular PCA is solved by finding an eigen decomposition of the covariance matrix, where the obtained eigenvectors (factor loadings) are used for projections of input variables into principal components.
  • Sparse PCA (SPCA) ( Zou, Hastie and Tibshirani 2006 ) seeks approximate sparse "eigenvectors" whose projections still capture the maximal variance of the data, but with only few input variables. SPCA is a regular eigen problem with cardinality constraints on eigenvectors.
  • SPCA is computationally intractable problem, recently few approximation techniques have been proposed: Lasso (elastic nets), Semi- Definite programming ( d'Aspremont, et al. 2005 ), and greedy approximation ( Moghaddam, Weiss and Avidan 2006 ). SPCA is intimately related to filter subset approach for feature selection.
  • Sparse LDA Moghaddam, Weiss and Avidan 2006a
  • FLDA is a dimensionality reduction technique, which aims to find a low-dimensional subspace of discriminant features where different classes linearly separated.
  • SLDA can be considered as an extension of SPCA.
  • Sparse LDA is intimately related to subset feature selection problem, and more specifically to Wrapper method. Roughly speaking, the solution of SLDA is an implementation of wrapper method for subset feature selection for a LDA classifier.
  • both methods, SPCA and SLDA perform simultaneous feature selection and dimensionality reduction.
  • Kernel PCA Scholkopf, Smola and Muller 1998
  • ISOMAP Teenbaum, de Silva and Langford 2000
  • LLE Locally Linear Embedding
  • Laplacian Eigenmap Belkin and Niyogi 2003
  • Diffusion maps Coifman, et al. 2005
  • Hessian eigenmaps Donoho and Grimes 2003
  • MDS Borg and Groenen 2005
  • the invention discloses an example of a reduction of dimensionality procedure as follows:
  • NSC Nearest Shrunken Centroid
  • NC Nearest Centroid
  • Nearest Shrunken Centroids classification The method Nearest Shrunken Centroids, also known by name Predictive Analysis of Microarrays (PAM), was first introduced for classification of genetic microarrays. It provides a list of significant features whose expression characterizes each class and estimates prediction error via cross-validation.
  • PAM Predictive Analysis of Microarrays
  • SUBSTITUTE SHEET (RULE 28) value of each feature in a class divided by the within-class standard deviation for that feature.
  • This standardization has the effect of giving higher weight to features whose expression is stable within samples of the same class.
  • Such standardization is inherent in other common statistical methods such as linear discriminant analysis.
  • Nearest Centroid classification takes a new sample, and compares it to each of these class centroids. The class whose centroid that it is closest to, in squared distance, is the predicted class for that new sample.
  • Nearest Shrunken Centroid classification "shrinks" each of the class centroids toward the overall centroid for all classes by an amount called the threshold.
  • the new sample is classified by the usual nearest centroid rule, but using the shrunken class centroids.
  • This shrinkage can make the classifier more accurate by reducing the effect of noisy features and provides an automatic feature selection.
  • a feature is shrunk to zero for all classes, then it is eliminated from the prediction rule. Alternatively, it may be set to zero for all classes except one, and it can be learned that the high or low value of that feature characterizes that class.
  • the user decides on the value to use for threshold. Typically one examines a number of different choices. To guide in this choice, NCS does K-fold cross-validation for a range of threshold values. The samples are divided up at random into K roughly equally sized parts.
  • the classifier is built on the other K-I parts then tested on the remaining part. This is done for a range of threshold values, and the cross-validated misclassification error rate is reported for each threshold value. Typically, the user would choose the threshold value giving the minimum cross-validated misclassification error rate.
  • NCS Given a dataset of n training samples distributed over k classes, NCS calculates a f - statistic d,(fe) of each feature / for each class k, xik ⁇ x i
  • s ⁇ is the pooled within-class standard deviation for feature i , k -J k where n k number of samples in class, and ⁇ o is a positive constant, usually equal to median value of s* .
  • ⁇ i00 compares the centroid x ⁇ tk of feature i of class & to the overall feature centroid *i .
  • the discriminant score for class k is defined as 2iOd ⁇ k where the first term is the standardized squared distance from new observation % * to &'th shrunken centroid and ⁇ r k j s simply the prior probability of class k.
  • the new observation will be classified into class c if S c C* * ) is the minimal among all classes.
  • Shrunken Centroid Regularized Linear Dirscrimination Analysis Closely related to NCS, but a more sophisticated algorithm, is Shrunken Centroid Reguralized Linear Dirscriminat Analysis (SCRLDA) ( Guo, Hastie and Tibshirani 2007 ).
  • Random Forrest (breiman 2001 ) is one of appealing alternatives when one deals with physiological parameters.
  • RF algorithm generates many random decision tree classifiers (splitting features chosen randomly) by bootstrapping (choosing with replacement) training samples. Final classification decision is calculated by majority voting of decision trees.
  • One of the major advantages of RF algorithm is it strong immunity against overfitting of training data. Moreover, as a sub-product it estimates the importance of variables in determining classification. Random forest is closely related to another method based on data bootstrapping called Bagging Classifier (Breiman 1996).
  • Bagging Classifier Breiman 1996
  • Each tree is fully grown and not pruned (as may be done in constructing a normal tree classifier).
  • This 0OB (out-of-bag) data is used to get a running unbiased estimate of the classification error as trees are added to the forest. It is also used to get estimates of variable importance.
  • SUBSTITUTE SHEET (RULb Ii) proximity is increased by one.
  • the proximities are normalized by dividing by the number of trees. Proximities are used in replacing missing data, locating outliers, and producing illuminating low-dimensional views of the data. Boosting
  • Boosting is similar to Random Forest approach as it works with multiple classifiers.
  • Boosting is a meta-classification paradigm which creates from plurality of weak classifiers (classifiers with classification performances only slightly better than random desicion) a strong classifier.
  • weak classifiers might be a simple threshold for single feature (decision stump) or decision tree with final depth (collection of thresholds for subset of features).
  • Boosting does not restrict the type of weak classifier.
  • weak classifiers might be a simple threshold for single feature (decision stump) or decision tree with final depth (collection of thresholds for subset of features).
  • Another difference between Boosting and Random Forrest is that weak classifiers are trained sequentially and classification is obtained by weighting average of weak classifier decisions, rather than by majority voting.
  • the major advantage of Boosting algorithms is their strong immunity against overfitting training data.
  • weights of each classifier provides indirect information about importance of associated with this classifier feature.
  • AdaBoost AdaBoost
  • LPBoost LPBoost
  • TotalBoost BrownBoost
  • MadaBoostm MadaBoostm
  • LogitBoost LogitBoost
  • GentleBoost SimpleBoost
  • SVM Support Vector Machine
  • Additional classifiers which might be used in a task of pain classification, include but are not limited to: Linear classifier, Na ⁇ ve Bayes Classifier, k-nearest neighbor, Quadratic Discriminant Analysis (QDA) classifier, Bagging Classifier, Radial Base Function (RBF) classifier, Multilayer Perceptron classifier, Bayesian Network (BN) classifier, etc. ( Hastie, Tibshirani and Friedman 2001 ) ( Bishop 2006 )
  • PPG envelope feature is extracted from PPG raw signal.
  • PPG signal envelope defined as PPG beat Peak amplitude minus beat Trough amplitude.
  • Two pain stimuli were applied. Each stimulus is 1 min long. "Start" and “End” point of each stimulus schematically depicted by red lines. 15 sec
  • Fig. 10 schematically illustrating HRV extracted from a PPG signal (HRV-PPG), HRV extracted from an ECG signal (HRV-ECG) and pain/no pain reports as a function of time elapsed
  • Fig. 1 1 represents the 2-dimensional scatter-plot of the HRV-PPG and the HRV-ECG.
  • Fig. 1 1 schematically illustrating HRV-PPG and HRV-ECG values when pain or no pain is reported
  • Fig 12 schematically illustrating HRV-PPG and HRV-ECG values when pain or no pain is reported - linear classifier possible separation Feature Reduction Example
  • the methods and algorithms of dimension reduction can find and perform such combinations and by that reduce the features space dimensionality without losing significant information.
  • the 2-d or 1-d line could't completely separate between the two classes; there were some error in each of the classes: pain occasions that were classified to the non-pain class (miss-detection) and non-pain occasions were classified to the pain class (false alarm).
  • Fig. 15 schematically illustrating HRV-PPG, HRV-ECG and
  • Fig. 16 schematically illustrating HRV-PPG, HRV-ECG and
  • Machine Learning 24 no. 2 ( 1996 ): 123-140 . Breiman, L. . " Random Forests .”
  • Machine Learning V45 no. 1 ( October 2001 ): 5-32 .

Abstract

The invention discloses a method and system for establishing the pain level in an awake, semi-awake or sedated patient. The method comprises steps of analyzing a multidimensional array of physiological signals to obtain the pain level of a patient. The signals are processed so as to extract a vector of Great Plurality of Features representing the patient's physiological status. The vector of the Great Plurality of Features is processed and classified into at least two classes for at least two conditions. These classes represent the pain level of the patient at a given time interval and thereby are used to establish the pain level of an awake, semi- awake or sedated patient. A system is provided for establishing the pain level in an awake, semi- awake or sedated patient.

Description

PAIN MONITORING USING MULTIDIMENSIONAL ANALYSIS OF PHYSIOLOGICAL SIGNALS
FIELD OF THE INVENTION
The field of the present invention relates to medical diagnostic tools. More particularly, the field of the present invention relates to systems and methods relating to measuring and reporting a subject's pain.
BACKGROUND OF THE INVENTION
Pain is an unpleasant sensory and emotional experience associated with actual or potential tissue damage, or described in terms of such damage. The inability to communicate verbally does not negate the possibility that an individual is experiencing pain and is in need of appropriate pain-relieving treatment (www. iasp-pain.org/ AM/). Pain is always subjective. Each individual learns the application of the word through experiences related to injury in early life.. Biologists recognize that those stimuli which cause pain are liable to damage tissue. Accordingly, pain is that experience we associate with actual or potential tissue damage. It is unquestionably a sensation in a part or parts of the body, but it is also always unpleasant and therefore also an emotional experience. Experiences which resemble pain but are not unpleasant, e.g., pricking, should not be called pain. Unpleasant abnormal experiences (dysesthesia) may also be pain but are not necessarily so because, subjectively, they may not have the usual sensory qualities of pain.
Many people report pain in the absence of tissue damage or any likely pathophysiological cause; usually this happens for psychological reasons. There is usually no way, based solely on the person's subjective report of his or her condition, to distinguish his or her experience from that due to tissue damage. If the experience is regarded as pain and if it is reported in the same way as pain caused by tissue damage, it should be accepted as pain. This definition avoids tying pain to stimulus.
"Pain Threshold" is defined as the least experience of pain which a subject can recognize as pain. Traditionally, this threshold has been defined as the least stimulus intensity at which a subject perceives pain. Properly defined, however, the threshold should be related to the experience of the patient, whereas the measured intensity of the stimulus is an external event. Because the threshold stimulus can be recognized as such and measured objectively, it. has been common usage for most pain research workers to define the threshold in terms of the stimulus, even though it is preferable to avoid such a definition. In psychophysics, a threshold is defined as the level at which 50% of stimuli are recognized. Thus, the pain threshold would be the level at which 50% of stimuli would be recognized as painful. As the stimulus is only one aspect of pain, it cannot be a measure or a definition of pain.
"Pain Tolerance Level" is defined as the greatest level of pain which a subject is prepared to tolerate. As with pain threshold, the pain tolerance level is the subjective experience of the individual. The stimuli which are normally measured in relation to its production are the pain tolerance level stimuli and not the level itself. Thus, the same argument applies to pain tolerance level as to pain threshold, and it should not be defined in terms of the external stimulation as such.
It can be said that there are four elements that could be used for observing and measuring pain in a subject: . . -
1. The lowest level is the stimulus itself;
2. Mental processing;
3. Environmental context of the experience (e.g. a visit to the dentist compared to playing football);
4. Behavioral response.
Depth of Anesthesia monitoring (DOA Monitoring) and pain monitoring are two fields that use sympathetic signals from the sympathetic nervous system or brain signals for monitoring a certain state of a patient. DOA monitoring is a general term for pain and awareness monitoring when a patient is under general anesthesia, a state in which pain and awareness cannot be distinguished since they both result in the same physiological symptoms. Pain monitoring follows only the sensation of physical discomfort while the subject can be in any condition including fully awake.
While depth of anesthesia monitoring has gained popularity during the last decade, due to publication of the terrible situation of "awareness during anesthesia," it is only in the last few years, since the Joint Commission on Accreditation of Healthcare Organizations (JCAHO) stated that (all) patients have a right to feel no pain, that pain monitoring has been of significant interest.
The DOA monitoring field is already saturated, with 30% of the market share being held by Aspect Medical (ASPM)'s BIS-Bispectral index. • BIS analyzes the patients' electroencephalograms during general anesthesia. Other predominant companies are GE Healthcare with their Entropy analysis and former Physiometrix (currently Hospira) with their PSA - Patient State Analyzer. These products apply the 'awareness part' of DOA while using the electroencephalogram signals. Other companies like Amtec Medical, MedSearch (US 6,117,075) and MedStorm (US 6,571,124), have tried to deal more with the analgesia component of the anesthesia using sympathetic and parasympathetic signal analysis such as Heart Rate Variability (Amtec), delay and correlation of signals in different locations (MedSearch) and skin conductance variability (Medstorm). However, it appears that using each parameter separately is dependent on various other sources other than the status of the analgesia level, such as temperature of the site, fluid balance, medication, etc. ( Guignard 2006 ), hence each of these parameters is unreliable. To summarize, most of the mentioned monitors that analyze Depth of Anesthesia (GE Healthcare, Physiomatrix, MedStorm, Med Search, Aspect Medical and others) use one type of signal such as EEG or one' of ANS signals, or at most a combination of two parameters in terms of delay between two signals (E. C. G + P.P. G by Aspect Medical app. 20040015091) or correlation between the two signals types or two signals sights (PPG + skin temperature by Med Search US 6,117,075).
The more advanced methods for monitoring the DOA are assigned by Instrumentarium (acquired by GE Healthcare) as the following:
United States Patent 6,685,649 to Korhonen discloses a method for monitoring a condition of a patient under anesthesia or sedation by acquiring and analyzing signals representing the cardiovascular activity of the patient. The anesthesia indicator is calculated by analysis of the acceleration trends of the cardiovascular activity (interval/rate or pressure) which result in an index correlated to the level of anesthesia This patent uses a basic hard decision rule on each of the acceleration trends values. This is inadequate since there are cross-connections between the parameters that affect the threshold values, and as was described above the heart rate or the blood pressure separately depends on various other sources rather than the status of the analgesia alone. Patents EP 1,495,715 and US 7,367,949 to Korhonen also disclose "method and apparatus based on combination of three physiological parameters for assessment of analgesia during anesthesia or sedation." The method includes utilizing brain activity signals in addition to analyzing the cardiovascular activity. This patent refers only to the sedated patients and therefore requires a smaller number of parameters, and simpler methods for classifying the level of patient comfort during anesthesia. United States Patent 7,215,994 to Huiku discloses a method for monitoring a state of anesthesia or sedation by comparing cortex related EEG biopotential signal data from the patient to subcortex-related biosignal data from the patient, the subcortex-related biosignal data including at least bioimpedance signal data. For each of the families of signals (EEG related, EMG related, Bioimpedance related and ECG related) an indicator is computed, and the four indicators are composed into a single number, but no method of combining is taught. Neither of the above prior art patents teach a method -of monitoring the pain of awake and aware patients.
Pain monitoring is more complicated, especially in patients who are fully awake, since the emotional state of the patient, the medication status, the environmental context and highly variable, often culturally determined, behavioral responses provide many signals that seriously affect specificity and sensitivity. Many more parameters therefore need to be measured and processed to achieve a meaningful pain monitoring method. Moreover, compared to anesthesia monitoring which is used only in operating rooms or in ICU when the patient is immobilized, not influenced from external stimuli and well controlled, pain monitoring needs to give pain indication in variant scenarios including when the patient might be in a movement and responsive to external stimuli. This scenario is far more complicated, and might require more information to be processed and more robust methods to handle the vast of information.
An unmet and long felt need remains to provide new algorithms for dealing with a large number of patient related parameters along with methods and devices to enable pain monitoring in subjects when awake, unanesthetized and unsedated. In order to apply all scenarios of pain including but not limited to acute pain under anesthesia, a far more robust and comprehensive solution is needed. A further long felt need is to enable pain detection and measurement in normal and impaired people under certain procedures in order to avoid unnecessary pain under surgical operations or medical procedures.
A further long felt need is to enable detection and measurement of sensation when this information is needed for a successful operation.
A further long felt need is to provide a method to differentiate between pain that is correlated to the stimuli and pain that is related to memory of pain so as to perform better diagnosis and treatment for pain clinics' patients.
A further long felt need is to provide a method to objectively measure pain as a function of given stimuli to neurological diagnosis , and to provide matching of applied sensors data with the pain level of stimulus input and a priori known data on the patient. The combination of controlled stimuli with exact measurements will enable the objective pain measurement thereby fulfilling an unmet need.
SUMMARY OF THE INVENTION
It is an object of the present invention to disclose a method for establishing the pain level in a patient, comprising analyzing a multidimensional array of physiological signals in order to obtain the pain level of a patient. The step of analyzing comprises steps of: a. acquiring a set of physiological signals from the body of a patient; b. processing the set of signals so as to extract a first vector of features representing the physiological status of the patient; wherein the first vector of features comprises a Great Plurality of Features (GPF); c. reducing of the dimensionality of the first vector of features by transforming the first vector to a second vector whose dimensions are lower by at least one order of magnitude compared to the first vector; d. classifying said second vector of features into at least two classes representing at least two conditions of pain; e. representing said classes of said pain level of said patient at a given time interval thereby establishing the pain level in an awake, semi- awake or sedated patient.
It is a further object of the invention to disclose the abovementioned method wherein the method additionally comprises a training step; said training step comprising; a. acquiring said set of physiological signals from said body of a patient or group of patients in a first non-pain state and a second pain state; b. processing said set of signals so as to extract a first vector of features representing the physiological status of said patient; wherein said first vector of features comprises a Great Plurality of Features (GPF); c. firstly, learning the parameters of said reducing of dimensionality by learning parameters of transformation of said first vector of features to a second vector whose dimensions are lower by at least one order of magnitude compared to said first vector; d. secondly, learning the parameters of a classifier that classifies said second vector of features into at least two classes representing at least two conditions of pain; e. setting the parameters of said classifier thereby establishing a classifier which classifies said pain level in an awake, semi- awake or sedated patient.
It is a further object of the invention to disclose the abovementoined method wherein the step of acquiring comprises selecting data from the group consisting of data supplied by the physician's, environmental parameters, patient parameters or any combination thereof It is a further object of the invention to disclose the abovementioned method wherein the method wherein said extracted features are selected from Table 2. It is a further object of the invention to disclose the abovementinoed method wherein said physiological signals represent an activity selected from the group consisting of autonomic nervous system activity, muscular activity, and brain activity. It is a further object of the invention to disclose the abovementinoed method wherein wherein said physiological signals are selected from a group comprising ECG, PPG, continuous blood pressure, respiration, internal or skin temperature, EOG or pupil diameter monitoring, GSR, EEG, EMG, EGG, LDV, capnograph and accelerometer or any combination thereof
It is a further object of the invention to disclose the abovementinoed method wherein said step of processing comprises analyzing the artifacts occurrence in said acquired signals. It is a further object of the invention to disclose the abovementinoed method wherein said step of representing said pain level of said patient is provided continuously during at least one predetermined time interval. It is a further object of the invention to disclose the abovementinoed method wherein said step of representing a PAIN or NON-PAIN condition of said patient is provided in a graduated scale.
It is a further object of the invention to disclose the abovementinoed method wherein said steps of reducing dimensionality or learning the parameters of reduction of dimensionality of said first vector of features further comprises the steps of: a. calculating extracted feature scores for each of said features or combination of features; b. filtering out said extracted low-score features thereby decreasing the number of said features to a predetermined number;
It is a further object of the invention to disclose the abovementinoed method wherein said calculating of said extracted features scores is based on activating said classifier processing and examining its results.
It is a further object of the invention to disclose the abovementinoed method wherein said steps of reducing dimensionality or learning the parameters of reduction of dimensionality of said first vector of features is achieved by obtaining sets of features and combining each of said sets into one meta feature activating a linear or non-linear function on said set, thereby reducing the dimensions of features vector.
It is a further object of the invention to disclose the abovementinoed method wherein said steps of combining and activating a linear or non-linear function is achieved by applying statistical methods selected from the group consisting of MDS, PCA, Sparse PCA,
FLDA, Sparse LDA, Kernel PCA, ISOMAP, LLE, Laplacian Eigenmaps, Diffusion
Maps, Hessian Eigenmaps or any combination thereof.
It is a further object of the invention to disclose the abovementinoed method wherein said classifying is obtained by applying a function on said second features vector which returns the pain class number.
It is a further object of the invention to disclose the abovementinoed method wherein said classifying is obtained by comparing said second features vector to at least one threshold surface which is defined by said learned parameters of classification in the space or subspace of said features. It is a further object of the invention to disclose the abovementinoed method wherein the step of comparing is performed by applying linear or non-linear operators or functions on said second features vector.
It is a further object of the invention to disclose the abovementinoed method wherein said steps of classifying or learning parameters of classification are achieved by applying statistical methods selected from a group consisting of Boosting, Linear classifier, Naϊve Bayes Classifier, k-nearest neighbor classifier, QDA classifier, RBF classifier, Multilayer Perceptron classifier, Bayesian Network classifier, Bagging classifier, SVM, NC, NCS, LDA, SCRLDA, Random Forest, or Committee of classifiers or any combination thereof. It is a further object of the invention to disclose the abovementinoed method wherein said step of classifying comprises the step of computing a confidence value of said second feature vector.
It is another object of the invention to provide a system for establishing the pain level in a patient wherein said system comprises; a. acquiring means for acquiring a set of physiological signals as input; b. memory means coupled to a microprocessor for analyzing . said set of physiological signals to obtain the pain level of the patient; c. representing means for representing said pain level of the patient as an output wherein said acquiring means communicates said physiological signals to said memory means coupled to said microprocessor. further wherein said memory means coupled to said microprocessor communicates said pain level of the patient to said representing means; further wherein said memory means coupled to said microprocessor is provided with; i. processing means for processing said set of signals so as to extract a first vector of features representing the physiological status of the patient; wherein the said first vector of features comprises a Great Plurality of Features (GPF), ii. reducing means for reducing the dimensionality of the said first vector of features by transforming to a second vector whose dimensions are lower by at least one order of magnitude compared to said first vector; iii. classifying means for classifying said second vector of features into at least two classes representing at least two conditions of pain. It is another object of the invention to provide the abovementioned system wherein said system additionally comprises a training subsystem; wherein said training subsystem comprises; a. acquiring means for acquiring a set of training physiological signals of a patient in a first non-pain state and a second pain state as training input; b. memory means coupled to a microprocessor for analyzing said set of training physiological signals to training and setting system parameters; c. representing means for representing said trained parameters and analysis results as an output; wherein said acquiring means communicates said physiological signals to said memory means coupled to said microprocessor further wherein said memory means coupled to said microprocessor communicates said trained parameters and analysis results to said representing means further wherein said memory means coupled to a microprocessor. means -is provided with; i. processing means for processing said set of signals so as to extract a first vector of features representing the physiological status of the patient; wherein the said first vector of features comprises a Great Plurality of Features (GPF), ii. first learning means for training said reducing of dimensionality by learning parameters of transformation of the said first vector of features to a second vector whose dimensions are lower by at least one order of magnitude compared to said first vector iii. second learning means for learning the parameters of classifier that classifying said second vector of features into at least two classes representing at least two conditions of pain iv. setting means for setting the parameters of said classifier thereby establishing a classifier which classifies said pain level in an awake, semi- awake or sedated patient.
It is another object of the invention to provide the abovementioned system wherein said acquiring means comprises sensors attached to the body of said patient for detecting said physiological signals. It is another object of the invention to provide the abovementioned system wherein said representing means is selected from the group consisting of computer screen, PDA screen, TV screen , plasma screen, LCD screen, patient monitor or any means for displaying numbers or graphs in a continuous manner
It is another object of the invention to provide the abovementioned system wherein said communication is by cable, wireless, blue tooth infra red, WI-FI or internet means.
It is another object of the invention to provide the system abovementioned wherein said acquiring means is provided with means for selecting data from the group consisting of data supplied by the physician's, environmental parameters, patient parameters or any combination thereof.
It is another object of the invention to provide the abovementioned system wherein said system further comprises input means for entering patient information and other data supplied by the physician.
It is another object of the invention to provide the abovementioned system wherein said physiological signals represent an activity selected from the group consisting of autonomic nervous system activity, muscular activity, and brain activity.
It is another object of the invention to provide the abovementioned system wherein said physiological signals are selected from a group comprising ECG, PPG, continuous blood pressure, respiration, internal or skin temperature, EOG or pupil diameter monitoring,
GSR, EEG, EMG, EGG, LDV, capnograph and accelerometer or any combination thereof.
It is another object of the invention to provide the abovementioned system wherein said processing means comprises means for analyzing the artifacts occurrence in said acquired signals.
It is another object of the invention to provide the abovementioned system wherein said representing means is adapted to represent said pain level of said patient continuously during at least one predetermined time interval.
It is another object of the invention to provide the abovementioned system wherein said representing means is adapted to provide the PAIN or NON-PAIN condition of said patient in a graduated scale.
It is another object of the invention to provide the abovementioned system wherein said reducing means or first learning means are adapted for; a. calculating said extracted feature scores for each of said features or combination of features and, b. filtering out said extracted low-score features thereby decreasing the number of said features to a predetermined number.
It is another object of the invention to provide the abovementioned system wherein said reducing means or first learning means are further adapted for calculating said extracted features score based on activating said classifier processing and examining its results. It is another object of the invention to provide the abovementioned system wherein said reducing means of said first vector of features or first learning means are provided with obtaining means for obtaining sets of features and combining each said set into one meta feature by activating a linear or non linear function on said set thereby reducing the dimensions of said first features vector.
It is another object of the invention to provide the abovementioned system wherein said obtaining means is adapted for combining and activating a linear or non linear function by applying statistical methods selected from the group consisting of MDS, PCA, Sparse PCA, FLDA, Sparse LDA5 Kernel PCA, ISOMAP, LLE, Laplacian Eigenmaps, Diffusion Maps, Hessian Eigenmaps or any combination thereof.
It is another object of the invention to provide the abovementioned system wherein said classifying means is adapted for classifying by applying a function on said second features vector which returns the pain class number.
It is another object of the invention to provide the abovementioned system wherein said classifying means is further adapted to classify by comparing said features to at least one threshold surface in the space or subspace of the said features.
It is another object of the invention to provide the abovementioned system wherein said classifying means is adapted to compare by applying linear or non linear operators or functions on said features.
It is yet another object of the invention to provide the abovementioned system wherein said classifying means or second learning means are adapted to apply statistical methods selected from the group consisting of Boosting, Linear classifier, Naϊve Bayes Classifier, k-nearest neighbor classifier, QDA classifier, RBF classifier, Multilayer Perceptron classifier, Bayesian Network classifier, Bagging classifier, SVM, NC, NCS, LDA, SCRLDA, Random Forest, or Committee of classifiers or any combination thereof Lastly, it is another object of the invention to provide the abovementioned system wherein said classifying means or second learning means are adapted to compute a confidence value of said vector.
In the following description, various aspects of the invention will be described. For the purposes of explanation, specific configurations and details are set forth in order to provide a thorough understanding of the invention. However, it will be also apparent to one skilled in the art that the invention may be practiced without specific details presented herein. Furthermore, well-known features may be omitted or simplified in order not to obscure the invention.
BRIEF DESCRIPTION OF THE DRAWINGS
Figure 1 : Pain Monitoring - System Description Figure 2: Pain Monitoring Sensors
Figure 3: Optional configuration of the pain monitoring system Figure 4: Flow diagram of the pain monitoring system Figure 5: ECG signal and its parameters Figure 6: Blood pressure or PPG signal and its parameters Figure 7: Parameter 'A' values on pain and not pain with two population Figure 8: Parameter 'A1 values on pain and not pain with two populations separated with parameter 'B' on z axis
Figure 9: PPG features changes during pain stimulus
Figure 10: FIRV from PPG, HRV from ECG and pain/no pain reports as function of time elapsed
Figure 11 : FIRV from PPG, FIRV from ECG values when pain or no pain is reported Figure 12: HRV from PPG, HRV from ECG values when pain or no pain is reported - linear classifier possible separation
Figure 13: Features reduction from 2 features (HRV from PPG, HRV from ECG) to one combined feature
Figure 14: HRV from ECG and R-R interval when pain or no pain is reported - with nonlinear classifier
Figure 15: HRV from PPG, HRV from ECG and R-R interval when pain or no pain is reported - 3D visualization
Figure 16: HRV from PPG, HRV from ECG and R-R interval when pain or no pain is reported - with a linear classifier possible separation
DETAILED DESCRIPTION OF THE PREFFERED EMBODIMENTS
The field of pain monitoring (distinct from depth of anesthesia), mostly developed by Algodyne (Patents US 6,826,426: to Lange et al, US 6,757,558: to Lange et ' al, US 6,654,632 to Lange et al), is only based on EEG tools, or as in the case of MedStorm, based on perspiration sensing. Instrumentarium Corporation (which was acquired by GE Healthcare) has also filed a patent in the field of pain monitoring: "Monitoring pain- related responses of a patient" (US 7,407,485). In this patent the inventor, Huiku, presents a pain monitoring that is based on one or more physiological parameter that are measured, normalized and then compared to 'a threshold surface'. The rate (number of occurrences per unit of time) of crossing this threshold is considered the pain level. In this patent the way to obtain the threshold which is the core of the classification, and therefore the most important part of it, is not specified
Moreover when using a plurality of physiological quantities there is a need to define a method for selecting a limited number of quantities, otherwise the computation complexity will explode and Huiku's method won't be applicable. Therefore we assume that either the number of quantities (or features) in the above system is very small (a few quantities), or there is a manual selection of features from a small number of expected features (a dozen of features).
Comparing to anesthesia monitoring which is used only in operating rooms or in ICU when the patient is immobilized, not influenced from external stimuli and well controlled, pain monitoring needs to give pain indication in variant scenarios including when the patient might be in motion and responsive to external stimuli. This scenario is far more complicated, and requires more information to be processed and more robust methods to handle the vast of information.
In this patent we suggest a new system and method to assess the pain level from a Great Plurality of Features (GPF) extracted from physiological signals, environmental parameters (temperature, sensor relative location) and prior information on the patient, so that all information that can affect the pain level classification will be included in the system. By Great Plurality we mean a large number (hundreds) of features or parameters. Therefore physiological parameters that are known to be related to pain and parameters that are allegedly less related to pain are included; as well additional parameters that can affect those physiological parameters are included. Also both relative (normalized) and
14 absolute values are calculated. All result in hundreds or even thousands of variables that cannot be compared or classified without prior selection or processing. The method further describes the methods for feature selection, reduction of dimensionality of a feature space, and classification of pain level.
Ultimately, the heart of any pain/analgesia monitoring system is a classification algorithm for taxonomy of patterns founded in physiological signals into classes of different pain level. Such classification algorithm is a mathematical engine which receives as an input a multidimensional vector of normalized features extracted from multiple physiological signals. In addition the algorithm receives (if exist) patient parameters (age, gender, weight, chronic diseases, historical measurements of physiological features, etc.), input from physician (diagnosis, receiving medicine, etc.) and environmental parameters (time, place, room temperature, accelerometer data, etc.). The output of the algorithm is a number which symbolizing a strength or level of pain of patient. In further we will denote this mathematical algorithm as a "Pain Classifier" or a "Classifier" The Pain classifier must be trained on a "training set". A Training set denotes a data .available from a variety of sources: publicly available databases, records of proprietary clinical trials, on site recorded data from the same patient or group of patients etc. The training set must be comprised from an input and output signals. The input signal has to be similar to the expected input of a pain classifier, i.e. multiple physiological signals, input from physician, environmental parameters and patient parameters. The output signal has to be similar to the expected output from a pain classifier, i.e. strength or level of pain of patient. Usually, the training output signal is determined by the human operator (physician or other skilled personal) during a clinical trial with controlled pain stimuli. Training a pain classifier on a training set means determining ("learning") the pain classifier parameters which will allow classifying of previously unseen input data (not from the training set) with sensitivity and specificity similar to or better than the performances of a human operator.
Patient related parameters, environmental parameters and input from physician may significantly improve performances of the pain classifier. They are essential in certain instances, for example in cases when a usage of medication affects the physiological response of the ANS to pain. However it may not influence the result when the information is less related to the physiological response to pain. Prior to input into the pain classifier, features should be appropriately normalized in order to remove patient baseline mean or/and normalize baseline variability or/and identify and remove outlier samples or/and normalize the features distribution etc. For example, in US 7,407,485 to Huiku and in US 7,367,949 to Korhonen a method of histogram normalization of feature probability distribution was proposed. The same normalization was proposed to all features in consideration. However, such normalization might eliminate valuable information which is often hidden in a shape of feature probability distribution.
In addition to normalization, since the number of input features might be very large (hundreds or even thousands of features), they also should be preprocessed for feature selection and/or dimensionality reduction. The Great Plurality of Features may be extracted from the received multiple physiological signals. Some of the features are directly related to painful response and some in an indirect way be related to painful response, or might be used to prevent misclassification. For example, a raise in blood pressure can result from painful stimuli, but also due to a change in position from sitting to standing which can be identified in the accelerometer. Some features are differentially expressed due to noise or inter patient variability.
Many of extracted variables may be redundant or at least strongly correlated (e.g. feature extracted from cardiovascular signals ECG, PPG and continuous blood pressure). Therefore efficient methods of features selection and dimensionality reduction should be applied as part of the training or learning of the pain monitoring system. Dimensionality reduction is defined as a mapping of the very high dimensional feature space into a space of fewer dimensions. Feature selection (picking a subset of original features) is a "straightforward" approach for dimensionality reduction problem. However, one can also combine different features in linear or non-linear way in order to create new features with good classification power. In the following we will differentiate between "Feature Selection" i.e. picking a subset and "Dimensionality reduction" i.e. creating new features which are referred as 'meta features'.
There are many different benefits of feature selection and dimensionality reduction in pain monitoring system: reducing the measurement and data storage requirements, reducing training and utilization times, defying the "curse of dimensionality" to improve classification performance, facilitating data visualization and data understanding for future research purposes. But the main purposes of the feature selection and dimensionality reduction are:
1. To evaluate the best features, and decrease the number of features used in the monitoring itself.
2. To reveal significant relation between pain and derived signals by indicating relations between parameters and pain stimuli.
3. To reveal specific combinations of parameters that together can result in one value relating to a pain state.
4. To overcome the risk of "overfitting" of the training data and improve performances of classification system for unseen data.
Feature selection algorithms typically fall into two categories; Feature Ranking and Subset Selection. Feature Ranking ranks the features by a specific metric, e.g. correlation with strength of pain stimuli, and eliminates all features that do not achieve an adequate score. However, simple ranking might eliminate some important features which by themselves are non good discriminants, but in combination with other features can play a vital role in a success of classification task. Subset Selection searches the set of possible features for the optimal subset and evaluates a subset of features as a group for suitability. Subset selection algorithms can be broken into Wrappers, Filters and Embedded ( Kohavi and John 1997 ). The Filter approach attempts to assess the quality of subset of features from the data ignoring specific classification algorithm. In contrast, in the Wrapper approach the best subset of features is chosen to suite specific classification procedures. Wrapper uses a search algorithm to search through the space of possible features and evaluate each subset by running a classifier on the subset. Wrappers can be computationally expensive and have a risk of over fitting to the model. Embedded techniques are embedded in and specific to a classifier. Embedded methods will be mentioned later in section dedicated to classification.
In patent US 7,367,949 B2 to Korhonen et al and patent US 7,407,485 to Huiku, the feature selection has been preformed, but the algorithm has not been disclosed. In simplistic scenario feature selection might be based either on some prior medical knowledge, exhaustive manual search or ad-hoc rule of thumb. However, manual selection is limited in the number of feature that can be searched, and using prior medical information is relevant when the feature are directly related to the result and limited in number, but less relevant when a hidden connection is searched or when the number of features is considerable high.
Moreover, in order to apply the pain-monitor into different scenarios: for example sedated person versus awake person, the search for the best features set, might be performed more than once in the system development lifetime, and may be used as a research tool for the physician. Therefore the correct automatic feature selection or similar procedure should be part of the system.
In ( Seitsonen, et al. 2005 ) on which partially based US 7,367,949 B2, a two-stage feature selection has been performed. More specifically, authors first applied Wilcoxon signed-ranked test to determine features with good classification power and later exhaustively searched over a grid of all combinations of remain features in order to find a best subset. The criterion of the search was performance of logistic regression classifier. In their work, 23 features were extracted from 3 sensors, after applying Wilcoxon signed- rank test 9 features were left, and after exhaustive search on the combination of features, 3 specific features were chosen. In other words, authors Stage-wise applied both Features ranking algorithm and Wrapper algorithm for logistic classifier. However, as we mentioned earlier, both Feature ranking and Wrapper methods have significant drawbacks in performances and computational complexity, and since Wilcoxon signed rank evaluates each feature alone, it might filter out features that differ in higher or different space (see the feature reduction example related to figure 13). In contrast to Feature Selection, Dimensionality Reduction methods in addition to reduction of complexity of the problem might also reveal an interesting interplay between features extracted from plurality of physiological signals. The main linear technique for dimensionality reduction is Principal Component Analysis (PCA) ( Hastie, Tibshirani and Friedman 2001 ). PCA chooses one or more linear combination of the original features which capture the maximal variance of the data. PCA is solved by finding an Eigen Value Decomposition (EVD) of the covariance matrix of available features. Unfortunately PCA does not incorporate any class information (pain levels) which can be obtained from the training data. Another example for dimensionality reduction is Fisher Linear Discriminant Analysis (FLDA) ( Hastie, Tibshirani and Friedman 2001 ), which can exploit class information. FLDA aims to find a low-dimensional subspace of discriminant features where different classes are linearly separated. FLDA is solved using the well known Generalized Eigen Value Decomposition (GEVD).
The Great Plurality of Features extracted from physiological signals, environmental parameters and prior information after preprocessing by feature selection and/or dimensionality reduction algorithm are the input into the pain level classifier. In order to design the pain classifier one should first determine the structure of the learning function and corresponding learning algorithm. There are many possible algorithms and approaches to choose from. For example, in US 7,367,949 B2 Korhonen et al. inventors specifically propose to use either Decision Tree classifier ("rule based reasoning") or Logistic Regression classifier. However, it is well knoλvn that in a case when number of input features is large, the logistic regression classifier suffers from so- called "curse of dimensionality" - exponentially grows of complexity of classifier training phase. Although Decision Trees are less affected by the "curse of dimensionality", in case of multiple features it tends to overfitt the training data. In US 2006/0217614 Al, US 2006/0217615 Al, US 2006/0217628 Al, and in US 2007/0010723 Al, inventors propose to calculate "weighted average'1 of few physiological features in order to calculate "Index of Nociception". Weighted average is a regression counterpart of simple linear classifier. It is well known that linear classifiers suffer from "curse of dimensionality" and perform poorly when the number of regressors is large.
In US 7,407,485 B2 by Huiku, the inventor suggests using "threshold surfaces" to determine the noxious event rate, without any specific reference to how to find those surfaces. Inventors of US 7,367,949 B2 have also mentioned the possible use of "some non-linear equation, logical rules and/or operators, artificial neural networks, fuzzy logic, etc." for calculating "probability of patient comfort", however, without specific details of designing and training of aforementioned algorithms those definitions are too vague to be performed or modified even by those skilled in the art.
In this invention we provide and disclose a strategy of using a classification algorithms specifically designed for classification in highly multidimensional spaces. The following description is provided, alongside all chapters of the present invention, so as to enable any person skilled in the art to make use of the invention and sets forth the best modes contemplated by the inventor of carrying out this invention. Various modifications, however, are adapted to remain apparent to those skilled in the art, since the generic principles of the present invention have been defined specifically to provide new methods and devices for a large number of patient related parameters to enable pain monitoring in subjects especially but not only when awake, unanesthetized and unsedated.
The present invention discloses a system that includes a plurality of sensors for acquisition of physiological signals that indicate sympathetic activity, parasympathetic activity brain activity, muscular activity, movements, environment parameters and prior information on the patient,
For example, ECG and PPG sensors acquire sympathetic/parasympathetic signals and EEG and EMG sensors acquire brain and muscular activity signals. Another component of the system disclosed herein is the processing unit, designed to process the signals in order to present them as features. A further element of the invention disclosed herein is the feature extraction for extracting and filtering features describing the subject pain state. The system further comprises artificial intelligent elements for defining the subject pain level and an output unit to present the results.
In more detail: the present invention discloses a novel system for pain monitoring, which combines parameters derived from many sensors such as ECG, PPG, EGG, Laser Doppler Velocimetry, skin conductance measurements, blood pressure measurements, capnograph peripheral, internal temperature measurements, respiration measurements, PD (pupil diameter) monitors, EOG, EEG and EMG5 movements of the patient from the accelerometer, and prior information on the patient.
Hundreds of features are extracted out of these sensors and the pain level is classified based on a multi-dimensional analysis. The use of the large number extracted from the above sensors describing the patient current state provides the patient with precise and efficient results and analysis. The present invention discloses methods for dimensionality reduction and classification in order to deal with the large amount of information and parameters.
Reference is now made to Fig. 1 : Pain Monitoring - System Description that schematically represents embodiment of the invention wherein steps in the method for monitoring pain are depicted. First, the patient is connected to appropriate sensors for a plurality of signal parameter acquisitions selected from a group comprising ECG, PPG, continuous blood pressure, respiration, internal or skin temperature, EOG or pupil diameter monitoring, GSR, EEG, EMG, EGG and accelerometer (1). The acquired signals are then processed by a microprocessor λvhich performs siRnal pre-processing such as de-noising, filtering, and other functions used for clearing and preparing the signal for feature extraction (2). Features are extracted and selected from the signals
(HRV, BPV, Freq analysis) and further features may be inputted by the clinician (3). The data is further processed by the algorithms processing and learning means (4). The output is then translated into pain level data (5).
Possible Sensors/Devices for Physiological Signals Acquisition
Galvanic skin response (GSR)
Galvanic skin response (GSR), also known as electrodermal response (EDR) or skin conductance response (SCR), is a method of measuring the electrical resistance of the skin. There is a relationship between sympathetic activity and emotional arousal, although one cannot identify the specific emotion being elicited.
GSR is conducted by attaching two or three leads to the skin, and acquiring a base measure. When an outgoing sympathetic nervous burst occurs, a wave of skin conductance will follow. During spontaneous skin conductance changes, increased number and amplitude of the waves is interpreted as increased activity in this part of the sympathetic nervous system (Lidberg and Wallin 1981)
Electrogastrogram (EGG)
Electrogastrography (EGG) is a noninvasive method for the measurement of gastric myoelectrical activity using abdominal surface electrodes
An electrogastrogram (EGG) is similar in principle to an electrocardiogram (ECG) in that sensors on the skin detect electrical signals indicative of muscular activity within. Where the electrocardiogram detects muscular activity in various regions of the heart, the electrogastrogram detects the wave-like contractions of the stomach
Pupil Diameter (PD) Measurement
Pupil size and movement can be measured by either infrared videography or computerized pupillometry. Pupillometry has been used in a research setting to study the autonomic nervous system, drug metabolism, pain responses, psychology, fatigue and sleep disorders. Infrared videography is used in order to detect magnified movement of both pupils at the same time. The technique allows the pupils to be visualized in the dark. Infrared videography takes advantage of dark pigmentation, since melanin actually reflects the infrared light shone on to the iris. Therefore, pigmented irises appear light on the screen, and the black pupils stand out in contrast to the surrounding light-appearing iris. Computerized pupillometry can record pupil size and movement in both the light and the dark. The typical instrument captures several frames per second over several seconds, and then averages the measurements. Such averaging compensates for the highly variable pupillary response that changes second to second. Electromyography (EMG)
EMG is a technique for evaluating and recording physiologic properties of muscles at rest and while contracting. EMG is performed using an electromyography to produce a record called an electromyogram. An electromyograph detects the electrical potential generated by muscle cells when these cells contract, and also when the cells are at rest. A surface electrode may be used to monitor the general picture of muscle activation. A motor unit is defined as one motor neuron and all of the muscle fibers- it innervates. When a motor unit fires, the impulse (called an action potential) is carried down the motor neuron to the muscle. The area where the nerve contacts the muscle is called the neuromuscular junction, or the motor end plate. After the action potential is transmitted across the neuromuscular junction, an action potential is elicited in all of the innervated muscle fibres of that particular motor unit. The sum of all this electrical activity is known as a motor unit action potential (MUAP). This electrophysiologic activity from multiple motor units is typically evaluated during an EMG. The composition of the motor unit, the number of muscle fibres per motor unit, the metabolic type of muscle fibres and many other factors affect the shape of the motor unit potentials in the myogram. EMG signals are essentially made up of superimposed motor unit action potentials (MUAPs) from several motor units. For a thorough analysis, the measured EMG signals can be decomposed into their constituent MUAPs. MUAPs from different motor units tend to have different characteristic shapes, while MUAPs recorded by the same electrode from the same motor unit are typically similar. Notably MUAP size and shape depend on where the electrode is located with respect to the fibers and so can appear to be different if the electrode moves position. EMG decomposition is non-trivial, although many methods have been proposed. Frontalis (scalp) electromyogram (FEMG) •
The frontalis muscle receives both visceral and somatic fibres from the facial nerve. The dual nerve supply means that this muscle can be influenced by autonomic activity. Two surface electrodes record compound action potentials from this muscle. The amplitude of the EMG decreases with increasing depth of anaesthesia, but this cannot be used in the paralysed ■ patient.
FEMG has the advantages of being non-invasive and convenient, and it is easy to apply the electrodes. However, the simultaneous recording of electroencephalography (EEG) signals poses technical problems due to low amplitude and interference. There is a wide inter-individual variability in measured FEMG. Photoplethysmograph (PPG)
A finger photoplethysmograph (PPG) is a non-invasive transducer to measure the relative changes of blood volume in a subject's .finger. Photoplethysmography is based on the determination of the optical properties of a selected skin area. For this purpose non- visible infrared light is emitted into the skin. More or less light is absorbed, depending on the blood volume in the skin. Consequently, the backscattered light corresponds with the variation of the blood volume. Blood volume changes can then be determined by measuring the reflected light and using the optical properties of tissue and blood. The measured signal records venous blood volume changes as well as the arterial blood pulsation in the arterioles. The relative change in blood volume reflects also the cardiovascular system activity and is controlled by the ANS. It has been suggested to be used as part of anesthesia monitor. Electrocardiogram (ECG)
An electrocardiogram is a graphic produced by an electrocardiograph, which records the electrical activity of the heart over time. Electrical impulses in the heart originate in the sinoatrial node and travel through the heart muscle where they cause contraction. The electrical waves can be measured at selectively placed electrodes (electrical contacts) on the skin. Electrodes on different sides of the heart measure the activity of different parts of the heart muscle. An ECG displays the voltage between pairs of these electrodes, and the muscle activity that they measure, from different directions, also understood as vectors. This display indicates the overall rhythm of the heart, and weaknesses in different parts of the heart muscle.
Electroencephalography (EEG)
Electroencephalography (EEG) is the measurement of electrical activity produced by the brain as recorded from electrodes placed on the scalp.
Scalp EEG measures the summed activity of post-synaptic currents. An action potential in a pre-synaptic axon causes the release of a neurotransmitter into the synapse that diffuses across the synaptic cleft and binds to receptors in a post-synaptic dendrite, resulting in a flow of ions into or out of the dendrite, which in turn results in compensatory currents in the extracellular space. It is these extracellular currents that generate EEG voltages.
Although post-synaptic potentials generate the EEG signal, it is not possible for a scalp
EEG to determine the activity within a single dendrite or neuron. Rather, a surface EEG reading is the summation of the synchronous activity of thousands of neurons that have similar spatial orientation, radial to the scalp. Currents that are tangential to the scalp are not picked up by the EEG. The EEG therefore benefits from the parallel, radial arrangement of apical dendrites in the cortex. Because voltage fields fall off with the fourth power of the radius, activity from deep sources is more difficult to detect than currents near the skull.
Scalp EEG activity oscillates at multiple frequencies having different characteristic spatial distributions associated with different states of brain functioning such as waking and sleeping. These oscillations represent synchronized activity over a network of neurons. The neuronal networks underlying some of these oscillations are understood while many others are not.
Skin/Peripheral temperature
Not much has to be talked about temperature sensor. The only consideration should be is the location of the probe. Once the probe is located near or on the fingers, the temperature is influenced from the air temperature (which should be therefore controlled, or removed according to the base recording), and from the body (blood fluent) which is what is considered as parameter. Skin temperature is considered to be affected by a combination of different physiological processes, e.g., perspiration and vascular tone, which determine the response of thermoregulation. Temperature variations were thought to -reflect changes in sympathetic vasoconstrictive tone and in concentration of circulating vasoactive substances occurring during both relaxation and stress ( Guyton 1982 ). Changes of arterioles' smooth muscle tone regulated by the sympathetic nervous system are considered to be one of the origins of these temperature fluctuations ( Cohen and Sherman 1983 ). The fluctuation of this myogenic activity and its effect on the skin microcirculation has been studied by different methods( Fagrell 1984 ).
Respiration
The respiration transducer directly measures the respiratory effort. The transducer measures the changes in thoracic or abdominal circumference that occur as the subject breathes. The design presents minimal resistance to movement and is extremely unobtrusive. The transducer can measure arbitrarily slow to very fast respiration patterns with no loss in signal amplitude, while maintaining excellent linearity and minimal hysteresis.
Blood pressure
Blood pressure refers to the force exerted by circulating blood on the walls of blood vessels, and constitutes one of the principal vital signs. The term blood pressure generally refers to arterial pressure, i.e., the pressure in the larger arteries.
The systolic arterial pressure is defined as the peak pressure in the arteries, which occurs near the beginning of the cardiac cycle; the diastolic arterial pressure is the lowest pressure (at the resting phase of the cardiac cycle). The average pressure throughout the cardiac cycle is reported as mean arterial pressure; the pulse pressure reflects the difference between the maximum and minimum pressures measured.
There are many physical factors that influence arterial pressure. Each of these may in turn be influenced by physiological factors, such as diet, exercise, disease, drugs or alcohol, obesity, excess weight and so-forth. Some physical factors are:
• Rate of pumping: the rate at which blood is pumped by the heart. The higher the heart rate, the higher the arterial pressure. • Volume of fluid or blood volume, the amount of blood that is present in the body. The more blood present in the body, the higher the rate of blood returns to the heart and the resulting cardiac output.
• Resistance. In the circulatory system, this is the resistance of the blood vessels. The higher the resistance, the higher the arterial pressure. Resistance is related to size (the larger the blood vessel, the lower the resistance), as well as the smoothness of the blood vessel walls.
• Viscosity or thickness of the fluid. If the blood gets thicker, the result is an increase in arterial pressure.
In practice, each individual's autonomic nervous system responds to and regulates all these interacting factors so that, although the above issues are important, the actual arterial pressure response of a given individual varies widely because of both split-second and slow-moving responses of the nervous system and end organs. The haemodynamic responses have been shown after noxious stimulation such as laryngoscopy or tracheal intubation. ( van den Berg, Sawa and Honjol 2006 )
Laser Doppler Velocimetry (LDV)
The laser Doppler quantifies blood flow in human tissues such as skin and by that evaluates the skin vasomotor reflex (SVMR),
In principle a monochromatic laser beam is directed at the skin surface. Light that is reflected off stationary tissue undergoes no shift whilst light that is reflected off cells with velocity (like red blood cells) undergoes Doppler shift. The degree of Doppler shift is proportional to the velocity of the cell into which it collided. This light is randomly reflected back out of the tissue and onto a photodetector which calculates the average velocity of cells within the tissue.
Capnograph
A capnograph is an instrument used to monitor the concentration or partial pressure of carbon dioxide (CO2) in the respiratory gases. It is usually presented as a graph of expiratory CO2 plotted against time, or, less commonly, but more usefully, expired volume. When expired CO2 is related to expired volume rather than time, the area beneath the curve represents the volume of CO2 in the breath, and thus over the course of a minute, this method can yield the CO2 minute elimination., an important measure of metabolism. Sudden changes in CO2 elimination during lung or heart surgery usually imply important changes in cardiorespiratory function. During procedures done under sedation, capnography provides more useful information than pulse oximetry. Capnographs usually work on the principle that CO2 absorbs infra-red radiation. A beam of infra-red light is passed across the gas sample to fall on to a sensor. The presence of CO2 in the gas leads to a reduction in the amount of light falling on the sensor, which changes the voltage in a circuit. Acceleronieter
An accelerometer is a device for measuring acceleration and gravity induced reaction forces. Single and multi-axis models are available to detect magnitude and direction of the acceleration as a vector quantity. An accelerometer measures the acceleration and gravity it experiences. Both are typically expressed in SI units meters/second2 (m/s2) or popularly in terms of g-force. For the practical purpose of finding the acceleration of objects with respect to the earth, the correction due to gravity along the vertical axis is usually made automatically, e.g. by calibrating the device at rest.
Reference is now made to Fig. 2: Pain Monitoring - Sensors. Fig .2 is an example of some of the possible sensors which can be used in embodiments of the invention, and is intended to illustrate the present invention but should not be interpreted as a limitation upon the reasonable scope thereof.
The signal acquisition is performed by collecting data from some or all of the following noninvasive sensors. For example the following sensors are preferably located on the patients' hand fingers and wrist (see Fig. 2):
• ECG - PQRST signal (see Fig. 5) characterizes the heart activity
• PPG waveform (from the pulse Oximeter) characterizes blood volume pulse (BVP)
• Blood Saturation (from the pulse Oximeter) characterizes oxygen levels in the blood
• Continuous Non-Invasive Blood Pressure -systolic, diastolic, median and internal quantities (see Fig. 6)
• Peripheral Temperature
• GSR - galvanic skin response - skin perspiration
• Accelerometer - for detecting change in positioning that need to be considered when analyzing the signal states
The following sensors are located wherever appropriate on the patient's body. • EOG - electrooculography - eye activity
• EGG - electrogastrogram - gastromydeletric activity
• EEG — electroencephalography - brain activity
• EMG - electromyography - muscles activity
• ECG - respiration via impedance characterizes.breathing
• Internal Temperature usually located on intubation - relevant only in case of anesthesia
• Capnograph - measures the carbon dioxide (CO2) concentration in exhaled and inhaled air.
• PD - Pupil Diameter - Infrared videography or computerized pupillometry - measuring the pupil size (dilation or erosion of pupil size can indicate on increase in sympathetic or parasympathetic activity accordingly)
The data thus acquired are decoded and transmitted via a transmission device to the local computer (monitor) that can present the data and perform all the data processing needed to compute the patient's pain level.
As part of the system, an accelerometer sensor is located near to other sensors (located on the patient's hand). This sensor does not acquire biological signals, but rather -signals generated by hand movements of the patient; such data is used to eliminate artifacts such as unexpected movements from signals due to sensitivity of the other sensors. Reference is now made to Fig. 3, schematically illustrating examples for possible configuration of the pain monitoring a. The system includes processing means, data is acquired and represented by external system b. The system includes acquiring means, processing means and displaying means, still some of the signal can be acquired by external means and the data can be also represented on external means. Communication between the sensors (acquiring means), the processing means and the displaying means can be wireless c. The system includes acquiring means and processing means, still some of the signal can be acquired by external means, and the data is represented on external means d. The system includes acquiring means, processing means and displaying means. Communication between the sensors (acquiring means), the processing means and the displaying means can be wireless. The processing and displaying means can be on personal digital assistant (PDAs) e. The system can be a stand alone systemj and includes all acquiring (physiological signal, and data input), processing and displaying means
Reference is now made to Fig. 4: flow diagram of the pain monitoring system. Fig. 4 is a flow diagram illustrating the method herein disclosed for pain monitoring using multidimensional analysis of physiological signals as follows;
• Step 1 comprises acquiring the signal from the sensors.
• Step 2 comprises separating out the artifacts in the signal from the signal of interest and defining the time resolution.
• Step 3: Extracting features from the received signals.
The steps are then divided into steps 4-5 in Fig. 4 used for classification of features in the patient or subject to be monitored for pain, and the steps 4'-5' in Fig. 4 are used for forming a computer generated 'learning" or "training" profile from -which feature classifications are selected as follows: Learning or Training Steps:
• Step 4': Processing vector of features by feature selection and/or dimensionality reduction methods in order to lower the dimension of the aforementioned vector.
• Step 5': Classifier is trained on a set of training examples of manually labeled signals and classifier parameters are learned.
Classification Steps:
• Step 4: Applying Feature selection and/or dimensionality reduction according to the learned parameters and function in the learning step
• Step 5: Classification of the pain level according to classifier parameters learned in steps 4 -5'.
• Steps 6 and 7: Presenting the pain level periodically both in training/classification; the refresh rate is defined according to the resolution time selected.
Features Description
The term "feature" hereinafter refers to an acquired signal describing a certain behavior of the signal, e.g. the amplitude of a band of frequency. Some signals have many features, e.g. ECG can have over 50 because its structure includes patterns that have physiological significance, whereas other features have less, e.g. temperature, which does not have a consistent pattern and therefore can have less features.
The term "artifact" refers herein to unexpected behavior of physiological signals which may appear often due to local or global motion of the aware subject, for example in EEG this can be eyes blinking, in PPG the finger movement or change of body position etc.. In the present invention these artifacts may also be analyzed and machine interpreted to assist arriving at the correct pain classification in pain monitoring.
Reference is now made to Fig. 5: schematically illustrating an ECG signal and its parameters as an example for one signal and description of some of its features.
Reference is now made to Fig. 6: schematically illustrating a blood pressure or pulseplethysmograph signal and its parameters as an example for one signal and description of some of its features.
As an example one can expect the following responses to acute pain stimuli within healthy subjects:
1. ECG - heart rate variability decreases, heart rate increases
2. ECG - respiration via impedance - increase in number of respirations, decrease, in ' respiration variability
3. Blood volume waveform - decrease in blood volume
4. Blood Pressure - increase in blood pressure, decrease in blood pressure variability
5. Temperature - increase in internal Temperature - decrease of external temperature (due to perspiration)
6. GSR - increase in GSR activity
7. EGG - Electrogastrogram - decrease in digestion activity Signals derived from the following are related to pain
8. EEG - electroencephalography - increase in EEG entropy
9. EMG - Electromyography - increase in FEMG entropy
Examples for common signal analysis algorithms that are used as part of this invention:
Entropy Entropy is related to the amount of disorder, complexity, or unpredictability of the system. It is a property of a physical system or data string consisting of a great number of elements. The concept is used in physical sciences and information theory. By adding the measurement of the cortical electrical activity, the clinician can assess the effect of anesthetics more comprehensively. EEG recordings change from irregular to more regular patterns when anesthesia deepens. Similarly, FEMG quiets down as the deeper parts of the brain are increasingly saturated with anesthetics. Entropy measures the irregularity of EEG and FEMG signals.
The entropy of the EEG signal within a certain time window can be calculated from the signal itself or its spectrum. Entropy of the signal has been shown to drop when a patient falls asleep and increase again λvhen the patient wakes up.
FEMG and EEG frequency analysis
When analyzing the frequency domain of EMG and EEG it can be shown-that the EEG frequency range is from about 0.5 Hz to 40 Hz depending on the state of mind a person is in, and the EMG frequency band is from about 20 Hz up to about 80 Hz
The EEG frequency domain can be further divided into the following frequency ranges that describe the following states of mind (see table below):
Figure imgf000032_0001
Table 1 : different state of mind frequency ranges
The usefulness of the present invention in pain monitoring is clear when consideration is given to the following: EEG and other neuron signals originating from the cortex hitherto have only been useful for DOA analysis. This is because during general anesthesia the functional activity of the neurons is decreased and synchronized and become more ordered and predictable. In awake subjects, the electrical activity of all cortical neurons working independently and recorded by the EEG results in a random, a-periodical and unpredictable signal, so they would not be useful on their own for pain monitoring. Methods disclosed by Algodyne, in the patents US 6,826,426: to Lange et al., US 6,757,558: to Lange et al., and US 6,654,632 to Lange et al.) in fact relies on EEG alone for pain monitoring and does not supply adequate data for sufficiently sensitive and specific pain monitoring in practice. In contrast, the present invention discloses a method wherein these signals can indeed be used for pain monitoring, when the data is combined with data derived from sympathetic/parasympathetic signals and other signals such as environmental signals. Heart rate variability (HRV)
Heart rate variability can be extracted from the signal(s) represent(s) the cardiovascular activity (ECG and PPG, Continuous BP). The - sympathetic activity is measured (it increases in the case of pain) and the parasympathetic activity measured (it decreases in case of pain). The aforementioned activities ratio can be extracted from the HRV and the peripheral blood pressure ( Deschamps, et al. 2004 ).
With heart rate variability, the high frequency (HF) peak, located around the respiratory frequency, typically between 0.15-0.4 Hz, reflects primarily parasympathetic activity ( Akselrod, et al. 1981 ) The low frequency peak (LF) centered on 0.1 Hz content of HR fluctuations is an estimate of combined vagal and sympathetic activity ( Malik 1996 ). In continuous peripheral blood pressure the LF content fluctuations is an estimate of sympathetic activity ( Pagani, Rimoldi and Malliani 1992 ).
The frequencies of 0.0033-0.04Hz are defined as very low frequency (VLF) and are associated with sympathetic activity.
The frequencies of 0.08-0.15 are defined as Medium Frequency (MF) and represent manly barrorecptor activity.
The following table summarizes the features are extracted from the acquired signals:
Figure imgf000033_0001
32
SUBSTiTuTE SHEET (RULE 28)
Figure imgf000034_0001
33
SUBSTITUTE SHEET (RULE 2E)
Figure imgf000035_0001
Figure imgf000036_0001
35
SUBSTITUTE SHEET (RULE 28)
Figure imgf000037_0001
Figure imgf000038_0001
37
SUBSTITUTE SHEET (RULE 28)
Figure imgf000039_0001
Figure imgf000040_0001
Figure imgf000041_0001
Table 2: List of possible extracted features Any subset of the features described above (signal, spectra, wavelets and statistical values and relations) can be combined into an N-dimensional vector which represents the patient's pain state at a certain time. Normalization per patient
It is herein acknowledged that each feature in a vector of features may be normalized in order to remove patient baseline mean or/and normalize patient baseline variability or/and identify and remove outliers or/and normalize the features distribution etc. Normalization hereafter denotes any data normalization method known in the art. The normalization is performed with respect to either baseline record of a patient (removing baseline mean, normalizing variance, etc.) or training set (outliers detection, distribution normalization, etc.) or both. The baseline may be recorded in the first few minutes, when the patient is in a constant position similar to the position of the treatment, with no pain stimuli. Alternatively, the baseline may be recorded at the first minutes, when the patient is in a constant position similar to the position of the treatment, and when a minimal pain stimulus such as infusion penetration has occurred. Baseline records may be also obtained from a patient historical records if exist
The normalization is feature-specific, i.e. different features may be normalized in a different manner. The normalization of a specific feature may be independent, i.e. performed for each feature independently from the other features, or the normalizations of a specific feature may depend on other features
The following describe some examples for normalization process in a non limiting manner:
1. The feature normalization which removes the patient's baseline feature mean is carried out, if needed, in the following manner:
Figure imgf000042_0001
where X1 -is the current feature that is processed and avg{X^aselme) is the average of the feature values of the patient baseline record.
2. The feature normalization which normalize the patient's baseline feature variability is carried out, if needed, in the following manner:
Figure imgf000042_0002
std(x;b sehm) where std(XT''"') is the standard deviation of the feature values of the patient baseline record.
3. The feature normalization which removes the patient's baseline feature mean and normalize the patient's baseline feature variability is carried out, if needed, in the following manner:
Figure imgf000043_0001
4. The feature normalization which removes the feature mean and normalize the feature variability with respect to the training set is carried out, if needed, in the following manner:
X, - avg(X™""g)
Figure imgf000043_0002
where OVg(X'""""* ) and std(X"am'"g ) are average and standard deviation of a feature values in the training data set, respectively.
5. The features normalization which normalizes the feature value into a value between [0,1] is carried out in the following manner:
Figure imgf000043_0003
max(Xt'rammg) - min(X;ra""ns) where VUaX(X""'"'"8 ) and mm(X"ammg) are the maximum or the minimum values of the feature in the training set, respectively. This normalization is prone to be affected by outliers in the training data set and can results in unreasonable max(X"a'n'"g) or min(Z;ra"""«) values.
6. Alternative a normalization which normalizes the feature value into a value between [0,1] is carried out in the following manner:
Figure imgf000043_0004
where a is a factor used so that most of the population will be in the output range. For example, if the distribution is normal then the 2*STD value represents 68%, 4*STD represents more than 95% of the population and 6*STD represents -99% of the population. Therefore setting the value of a can increase/decrease the percentage of the extreme samples that are excluded from consideration. The extreme samples are declared as outliers and their values are set to be 0 or 1. Another possibility is to omit the outliers from consideration and to treat them as missing samples. The normalizations methods that are described above normalize either the features range of values or the features first and second moments with respect to either the patient baseline record or to the training data set. The normalizations of other features parameters such as kurtosis, skewness, higher order moments and cummulants or features probability distribution function are also applicable if needed. For example, in US 7,407,485 to Huiku and in US7367949 to Korhonen histogram normalization method of all features is proposed. However, such normalization might delete valuable information which is often hidden in the shape of the feature probability distribution. A priori data: The pain response is known to be a complicated sensation resulting from:
• Nociception response
• Conceptual response: how the brain processes incoming stimuli
• Context relevance response: the context in which the pain was caused.
• Behavioral response: individual responses to pain - based on cultural and personal behavior.
In a normal subject (healthy subject without known disorders, chronic diseases or receiving medication), it is assumed that the autonomic tone is correlated with the conceptual pain (and might be also some of the context relevance). Since the subject that uses the pain monitoring system might not be a 'normal subject' in one of many terms, all relevant information that can affect the autonomic tone and known by the care provider, should be entered as parameters to the system. Some of these parameters are categorical: nominal (gender, type of medicine, diagnostics, etc.), ordinal (patient condition, patient definition of pain level, etc.), interval (age group etc.). Other parameters are numerical (weight, height, historical features data, etc.). The parameters might be continuous or discrete, quantitative or qualitative. These parameters are of high importance since for example, use of beta-blockers cause degradation of the sympathetic response, therefore, even though a pain stimulus has occurred, which usually causes high sympathetic tone in normal subjects, the sympathetic tone of a subject using beta-blockers does not change significantly.
It is a core principle of the invention that this prior data is utilized in some embodiments of the invention in at least one of the many following options:
1. Features from normal subject and subject known to have a certain disorder, or that use a certain drug (for example the beta blocker) will be differently normalized.
2. The pain levels can be weighted. E.g., level 3 for a normal subject is considered as level 6 for subject with certain disorder or usage drug.
3. By inserting to the training set examples of both normal subjects and subjects known to have a certain disorder, or that use a certain drug, the system is trained to differ between the population by using feature that describe the prior information (as depicted in Fig 7 and Fig. 8)
4. A differently trained classifier for certain disorders/drug usage users is employed. The following example schematically demonstrates an exemplary use of the a-priori data. Reference is made to Fig. 7. Two populations that have different specific prior parameter are depicted by small or large symbols. Small/Large '+' symbols and small/large O' symbols represent non-pain and pain events respectively. Using parameter 'A' (y-axis of the figure) it is difficult to distinguish between the pain reported event and the no pain reported event, since parameter 'A' differs between the two populations. In Fig. 8 it can be seen that once the populations are separated, the values of parameter 'A' do differ between the pain reports and no pain reports.
Reference is now made to Fig. 7, schematically representing parameter 'A' values on pain and no pain with two populations
Reference is now made to Fig. 8, schematically representing Parameter 'A' values on pain and no pain with two populations separated with parameter 'B' on the z axis
Dimensionality Reduction and Feature Selection
Processing vector of features in order to reduce its dimension enables the system to evaluate the best features and decrease the number of features used in the monitoring itself. It also reveals specific combinations of parameters that together can result in one value with significant relation to a pain state, can overcome the risk of "overfitting" of the training data and can improve the overall performances of classification system for unseen data. Principal Component Analysis (PCA) was and remains the most popular technique for dimensionality reduction. Recall that PCA chooses one or more linear combination of the original features which capture the maximal variance of the data. However, those linear combinations are usually combined of all the input features. In a pain monitoring system, one wishes to have more sparse principal components, i.e. linear combination of only few input features. Sparsity of a principal component might significantly improve interpreterability of the result and provides valuable insights for physician. Regular PCA is solved by finding an eigen decomposition of the covariance matrix, where the obtained eigenvectors (factor loadings) are used for projections of input variables into principal components. Sparse PCA (SPCA) ( Zou, Hastie and Tibshirani 2006 ) seeks approximate sparse "eigenvectors" whose projections still capture the maximal variance of the data, but with only few input variables. SPCA is a regular eigen problem with cardinality constraints on eigenvectors. Although SPCA is computationally intractable problem, recently few approximation techniques have been proposed: Lasso (elastic nets), Semi- Definite programming ( d'Aspremont, et al. 2005 ), and greedy approximation ( Moghaddam, Weiss and Avidan 2006 ). SPCA is intimately related to filter subset approach for feature selection.
Similarly to SPCA, Sparse LDA ( Moghaddam, Weiss and Avidan 2006a ), attempts to solve Fisher LDA with a cardinality constraint. Recall that FLDA is a dimensionality reduction technique, which aims to find a low-dimensional subspace of discriminant features where different classes linearly separated. SLDA can be considered as an extension of SPCA. Moreover, Sparse LDA is intimately related to subset feature selection problem, and more specifically to Wrapper method. Roughly speaking, the solution of SLDA is an implementation of wrapper method for subset feature selection for a LDA classifier. Thus, both methods, SPCA and SLDA, perform simultaneous feature selection and dimensionality reduction.
All of the above mentioned techniques for dimensionality reduction of feature vector are linear methods. However, it is well known that sometimes non-linear relationship between input features may play great role in pain and DOA monitoring. For example, well known BIS feature, which is a core of Aspect Medical DOA monitor, is based on spectral bicoherence. Therefore, non-linear dimensionality reduction methods should be considered. During the last decade few very powerful non-linear methods for dimensionality reduction were proposed. Among them: Kernel PCA ( Scholkopf, Smola and Muller 1998 ), ISOMAP (Tenenbaum, de Silva and Langford 2000 ), Locally Linear Embedding (LLE) (Roweis and Saul 2000 ), Laplacian Eigenmap (Belkin and Niyogi 2003 ), Diffusion maps (Coifman, et al. 2005 ) Hessian eigenmaps ( Donoho and Grimes 2003 ), MDS ( Borg and Groenen 2005 ) etc.
The invention discloses an example of a reduction of dimensionality procedure as follows:
In this invention we propose to use an automated feature selection and dimensionality reduction procedure of pain related features as a data processing step followed by training of classifier. Linear methods (e.g. PCA, SPCA, FLDA, SLDA, MDS etc.) and/or nonlinear methods (e.g. KPCA,ISOMAP, LLE, Laplacian Eigenmaps, Diffusion maps, Hessian eigenmaps etc.) applied to Great Plurality of Features allow to reduce training time of classifier, improve classification performances, and reveal an interesting information regarding interplay of different physiological features. Classification
In this invention we propose to use in a non limiting manner one from a few possible classifier which were specifically designed for classification tasks in Great Plurality of Features cases.
Nearest Shrunken Centroids
First family of classifiers we propose to use is Nearest Shrunken Centroid (NSC) classifier ( Tibshirani, et al. 2002 ) which is a variation of well known Nearest Centroid (NC) classifier ( Hastie, Tibshirani and Friedman 2001 ) adapted for large number of features. The algorithm "shrinks" unimportant features which are not significantly differentiate between classes. Amount of shrinkage is chosen using cross validation. This part of the algorithm can be considered as a feature selection step embedded in NC classifier.
General description of Nearest Shrunken Centroids classification: The method Nearest Shrunken Centroids, also known by name Predictive Analysis of Microarrays (PAM), was first introduced for classification of genetic microarrays. It provides a list of significant features whose expression characterizes each class and estimates prediction error via cross-validation.
In this method, a standardized centroid is computed for each class. This is the average
46
SUBSTITUTE SHEET (RULE 28) value of each feature in a class divided by the within-class standard deviation for that feature. This standardization has the effect of giving higher weight to features whose expression is stable within samples of the same class. Such standardization is inherent in other common statistical methods such as linear discriminant analysis. In general, Nearest Centroid classification takes a new sample, and compares it to each of these class centroids. The class whose centroid that it is closest to, in squared distance, is the predicted class for that new sample. Nearest Shrunken Centroid classification "shrinks" each of the class centroids toward the overall centroid for all classes by an amount called the threshold.
After shrinking the centroids, the new sample is classified by the usual nearest centroid rule, but using the shrunken class centroids. This shrinkage can make the classifier more accurate by reducing the effect of noisy features and provides an automatic feature selection. In particular, if a feature is shrunk to zero for all classes, then it is eliminated from the prediction rule. Alternatively, it may be set to zero for all classes except one, and it can be learned that the high or low value of that feature characterizes that class. The user decides on the value to use for threshold. Typically one examines a number of different choices. To guide in this choice, NCS does K-fold cross-validation for a range of threshold values. The samples are divided up at random into K roughly equally sized parts. For each part in turn, the classifier is built on the other K-I parts then tested on the remaining part. This is done for a range of threshold values, and the cross-validated misclassification error rate is reported for each threshold value. Typically, the user would choose the threshold value giving the minimum cross-validated misclassification error rate.
Short mathematical description of Nearest Shrunken Centroids: Given a dataset of n training samples distributed over k classes, NCS calculates a f - statistic d,(fe) of each feature / for each class k, xik ~ xi
<*,(*) = m k (β, + S0)
where sι is the pooled within-class standard deviation for feature i , k -J k where nk number of samples in class, and ^o is a positive constant, usually equal to median value of s* . Thus άi00 compares the centroid x~tk of feature i of class & to the overall feature centroid *i . By further soft thresholding on d-i VO as
47
SUBSTITUTE SHEET (RULE 28)
Figure imgf000049_0001
where + mean positive part and Δ is a threshold choosen by cross validation as explained earlier. The shrunken centroid of feature i is xik = Xi + mk Q>i +■ sB)di'(k)
Thus the model only selects features that are truly distant from the overall centroid to be included in class centroid computation.
For a given new observation of feature vector x* = (%£,x|, ...x*) the discriminant score for class k is defined as
Figure imgf000049_0002
2iOdπk where the first term is the standardized squared distance from new observation %* to &'th shrunken centroid and τrk js simply the prior probability of class k. The new observation will be classified into class c if Sc C**) is the minimal among all classes. Shrunken Centroid Regularized Linear Dirscrimination Analysis Closely related to NCS, but a more sophisticated algorithm, is Shrunken Centroid Reguralized Linear Dirscriminat Analysis (SCRLDA) ( Guo, Hastie and Tibshirani 2007 ). This method generalizes the idea of the nearest shrunken centroids into the classical discriminant analysis. As in NCS, algorithm shrinks unimportant features and uses Regularized LDA as a classification algorithm. This algorithm can be useful even if a number of features is greater than a number of training data points. This situation may arise if we don't have enough training data for specific patient or pain-related scenario. Short mathematical description of SCRLDA:
Recall that class discriminant scores used in classical Linear Discriminant Analysis (LDA) ( Hastie, Tibshirani and Friedman 2001 ) are (in vector notation): δk CO = -Oc* - xύτW-ι&- - X31)- 2/σ5τrk where W is a pooled within-class covariance matrix. This is eventually a Mahanlobis distance from a new sample to class centroid. In case when a number of features is larger than number of data samples pooled within-class covariance matrix W; is rank deficient. Therefore, one should regularize it by W = aW +■ (1 - ay where 0 < α < 1 and / is an identity matrix. Shrinkage procedure of class centroids is performed similarly to NSC. Modified class score is
48
SUBSTITUTE SHEET (RULE 28) SkCO = ~(x' - xkf - xk)~ 2logπk Another possibility instead to shrink centroid itself is to shrink normalized class centroid
Figure imgf000050_0001
Random Forest
Random Forrest (RF) ( Breiman 2001 ) is one of appealing alternatives when one deals with physiological parameters. RF algorithm generates many random decision tree classifiers (splitting features chosen randomly) by bootstrapping (choosing with replacement) training samples. Final classification decision is calculated by majority voting of decision trees. One of the major advantages of RF algorithm is it strong immunity against overfitting of training data. Moreover, as a sub-product it estimates the importance of variables in determining classification. Random forest is closely related to another method based on data bootstrapping called Bagging Classifier (Breiman 1996). General description of Random Forest classification: Each tree is constructed using the following algorithm:
• Let the number of training cases be N, and the number of variables in the classifier be M.
• We are told the number m of input variables to be used to determine the decision at a node of the tree; m should be much less than M.
• Choose a training set for this tree by choosing N times with replacement from all N available training cases (i.e. take a bootstrap sample). Use the rest of the cases to estimate the error of the tree, by predicting their classes.
• For each node of the tree, randomly choose m variables on which to base the decision at that node. Calculate the best split based on these m variables in the training set.
• Each tree is fully grown and not pruned (as may be done in constructing a normal tree classifier).
When the training set for the current tree is drawn by sampling with replacement, about one-third of the cases are left out of the sample. This 0OB (out-of-bag) data is used to get a running unbiased estimate of the classification error as trees are added to the forest. It is also used to get estimates of variable importance.
After each tree is built, all of the data are run down the tree, and proximities are computed for each pair of cases. If two cases occupy the same terminal node, their
49
SUBSTITUTE SHEET (RULb Ii) proximity is increased by one. At the end of the run, the proximities are normalized by dividing by the number of trees. Proximities are used in replacing missing data, locating outliers, and producing illuminating low-dimensional views of the data. Boosting
Another possible approach for pain classification is to use a Boosting framework ( Bishop 2006 ). Boosting is similar to Random Forest approach as it works with multiple classifiers. Boosting is a meta-classification paradigm which creates from plurality of weak classifiers (classifiers with classification performances only slightly better than random desicion) a strong classifier. However, in contrast to Random Forest, Boosting does not restrict the type of weak classifier. For example weak classifiers might be a simple threshold for single feature (decision stump) or decision tree with final depth (collection of thresholds for subset of features). Another difference between Boosting and Random Forrest is that weak classifiers are trained sequentially and classification is obtained by weighting average of weak classifier decisions, rather than by majority voting. Similarly to RF, the major advantage of Boosting algorithms is their strong immunity against overfitting training data. Moreover, if we use decision stumps as a weak classifiers, weights of each classifier provides indirect information about importance of associated with this classifier feature.
Algorithms that were developed using Boosting framework: AdaBoost, LPBoost, TotalBoost, BrownBoost, MadaBoostm, LogitBoost, GentleBoost and many others. Support Vector Machine
We propose to use Support Vector Machine (SVM) as a pain classifier ( Vapnik 1998 ). Using simple mathematical apparatus of kernel ization and convex optimization, training of SVM is able to produce complicated non-linear classification surfaces. The major disadvantage of SVMs is that it black-box procedure with little interpretative value. Moreover, kernelized SVM is strongly dependent on choosing appropriate kernel (we propose to use a family of universal kernels, for example the RBF kernel) ( Chapelle, et al. 2000 ). However flexibility of classification surfaces obtained by SVM and its computational efficiency provide great advantage against other linear and non-linear classifiers.
50
SUBSTITUTE SHEET (RULE 28) Additional Classifiers
Additional classifiers, which might be used in a task of pain classification, include but are not limited to: Linear classifier, Naϊve Bayes Classifier, k-nearest neighbor, Quadratic Discriminant Analysis (QDA) classifier, Bagging Classifier, Radial Base Function (RBF) classifier, Multilayer Perceptron classifier, Bayesian Network (BN) classifier, etc. ( Hastie, Tibshirani and Friedman 2001 ) ( Bishop 2006 )
Combination of Pain Classifier with Feature Selection and cascading of Classifiers Some classifiers do required to use feature selection and dimensionality reduction methods as a preprocessing step of training and classifications tasks, when other classifiers do not. As we have mentioned earlier, Nearest Shrunken Centroid classifier and Shrunken Centroid Reguralized Linear Dirscriminat Analysis classifier use feature selection as an intermediate step in classification procedure. Hence feature selection is already embedded in these classifiers. Random Forest provides feature importance index, which can be used for feature selection. However these and other aforementioned classifiers might benefit of preprocessing feature selection, and even more important dimensionality reduction step.
When working on real life data, sometime it is not sufficient to use one classifier. One type of classification which is appropriate to specific signal can have worse performances for other signals. Hence, a combination of classifiers of different nature is sometimes required. For example, hierarchical cascade of classifiers will reduce the computational complexity of problem and will provide flexibility to classify different physiological features using different methods. Another possibility is tree of classifiers. Committee of classifiers is a third approach. In this case classifiers are not worked not sequentially, rather in parallel. The classification is performed using majority voting of all classifiers or weighting average of voters. Feature extraction examples PPG Envelope Feature:
Reference is now made to Fig. 9. In the following example the PPG envelope feature is extracted from PPG raw signal. PPG signal envelope defined as PPG beat Peak amplitude minus beat Trough amplitude. Two pain stimuli were applied. Each stimulus is 1 min long. "Start" and "End" point of each stimulus schematically depicted by red lines. 15 sec
51
SUBSTITUTE SHEET (RULE 2E) sliding window was applied, and Mean Value and Standard deviation of each interval were calculated.
It can be observed that regions of pain are characterized by stable low mean amplitude of PPG envelope, and low amplitude variability (standard deviation). It demonstrates the high importance of a PPG signal envelope feature. However, non-pain regions with either low mean amplitude or low amplitude variability or both exist as well. Erratic behavior of a feature can be explained due to noise or movement artifact. Therefore, combination of a PPG envelope feature with other features is necessary. Classification example
Classification based on HRV extracted from PPG and ECG signals Reference is now made to Fig. 10: schematically illustrating HRV extracted from a PPG signal (HRV-PPG), HRV extracted from an ECG signal (HRV-ECG) and pain/no pain reports as a function of time elapsed
The HRV is affected by a real change in the autonomic tone. It can be seen in this example that the pain occurrences marked by the thin line sign occur near local minima both in the HRV-PPG and the HRV-ECG. This is more easily noticed on the HRV-ECG graph but it is still very difficult to be noticed in this representation. Fig. 1 1 represents the 2-dimensional scatter-plot of the HRV-PPG and the HRV-ECG.
Reference is now made to Fig. 1 1 : schematically illustrating HRV-PPG and HRV-ECG values when pain or no pain is reported
In the scatter-plot shown in Fig.1 1 it can be seen that the two populations (pain/circle- no pain/plus) can be easily separated by a line dividing the space as is depicted in Fig. 12. For such a division even a simple linear classifier can result in high sensitivity and specificity results as follows.
Reference is now made to Fig 12: schematically illustrating HRV-PPG and HRV-ECG values when pain or no pain is reported - linear classifier possible separation Feature Reduction Example
Reference is now made to Figure 13: Features reduction from 2 features (HRV-PPG and HRV-ECG) to one combined feature
In this example the schematic example of figure 12 is considered. As can be seen, had one tried to classify pain level using only either the HRV-PPG (X dimension) or HRV- ECG (Y dimension) alone no reliable classification could be made. Only the combination
52
SUBSTITUTE SHEET (RULE 28) of the two features can result in a good-separation. Had we activated a per feature score test, both features might have been filtered out. However, if one combines them into one feature z = ax -y+b one would find that this feature is a significant feature that can differ between two pain/no-pain classed.
The methods and algorithms of dimension reduction can find and perform such combinations and by that reduce the features space dimensionality without losing significant information.
Non Linear Classification Example
In the above examples with or without features reduction, the 2-d or 1-d line couldn't completely separate between the two classes; there were some error in each of the classes: pain occasions that were classified to the non-pain class (miss-detection) and non-pain occasions were classified to the pain class (false alarm).
Reference is now made to Figure 14: HRV-ECG and R-R interval when pain or no pain is reported - with non-linear classifier
In this example it can be shown that had we had a classifier with a non linear function to differ between the pain and non-pain classes, the number of miss-detection and false alarms would have been dramatically decreased.
Classification based on 3 Dimension Example
In the following schematic example, 3 parameters are extracted: HRV-PPG, HRV-ECG and R-R interval values. Separation of the populations can be made by dividing the space by a surface as shown in Figs. 15 and 16. Nevertheless, better results would be achieved by separating the space with a non linear surface or number of non linear surfaces.
Reference is now made to Fig. 15: schematically illustrating HRV-PPG, HRV-ECG and
R-R interval when pain or no pain is reported - 3D visualization
Reference is now made to Fig. 16: schematically illustrating HRV-PPG, HRV-ECG and
R-R interval when pain or no pain is reported - with a linear classifier separation
As was shown in the above examples, despite the fact that in certain and limited scenarios small number of features can differ between pain and non pain assuming that a non-linear classification is used, still in the non-controlled environment and even in this limited and controlled environment there is a need for a better description of the pain state and a way to decrease the false alarm and miss detection. This is where the Great Plurality of Features is used, and this is why advanced methods of classification and feature selection need to be applied.
While a number of exemplary aspects and embodiments have been discussed above, those of skill in the art will recognize certain modifications, permutations,additions and sub-combinations thereof. It is therefore intended that the following appended claim as and claims hereafter introduced be interpreted to include all such modifications, permutations, additions and sub-combinations as are within their true spirit and scope.
BIBLIOGRAPHY
Akselrod, S. , D. Gordon, F. A. Ubel, D. C. Shannon, A. C. Berger, and R. J. Cohen. " Power spectrum analysis of heart rate fluctuation: a quantitative probe of beat-to-beat cardiovascular control. ." Science 213 , no. 4504 ( July 1981 ): 220-222 .
Belkin, M. , and P. Niyogi. " Laplacian eigenmaps for dimensionality reduction and data representation ." Neural Computation 15 ( 2003 ): 13731396 .
Bishop, CM. Pattern recognition and machine learning . New York : Springer , 2006 .
Borg, I. , and P. Groenen. Modern Multidimensional Scaling: theory and applications . New York : Springer- Verlag , 2005 .
Breiman, L. . " Bagging Predictors ." Machine Learning 24 , no. 2 ( 1996 ): 123-140 . Breiman, L. . " Random Forests ." Machine Learning V45 , no. 1 ( October 2001 ): 5-32 .
Bruhn, J. , H. Ropcke, and A. Hoeft. " Approximate entropy as an electroencephalograph ic measure of anesthetic drug effect during desflurane anesthesia. ." Anesthesiology 92 , no. 3 ( March 2000 ): 715-726 .
Chapelle, O. , V. Vapnik, O. Bousquet, and S. Mukherjee. " Choosing Kernel Parameters for Support Vector Machines ." Machine Learning , 2000 .
Cohen, D. H. , and S. M. Sherman. " The autonomic nervous system and its central control ." Edited by R. Berne, and M. Levy,, 322 . St. Louis : CV. Mosby , 1983 .
Coifman, R. R. , et al. " Geometric diffusions as a tool for harmonic analysis and structure definition of data ." Proceedings of the National Academy of Sciences 102 ( 2005 ): 7426-7438 . d'Aspremont, A. , L. El Ghaoui, M. I. Jordan, and G. R. G. Lanckriet. " A direct formulation for sparse PCA using semidefinite programming ." Cambridge, MA : MIT Press , 2005 .
Deschamps, A. , I. Kaufman, S. B. Backman, and G. Plourde. " Autonomic nervous system response to epidural analgesia in laboring patients by wavelet transform of heart rate and blood pressure variability. ." Anesthesiology 101 , no. 1 ( July 2004 ): 21-27 .
Donoho, D. L. , and C Grimes. " Hessian eigenmaps: Locally linear embedding techniques for high-dimensional data ." PNAS 100 , no. 10 ( May 2003 ): 5591-5596 .
Goncharova, 1. 1. , and J. S. Barlow. " Changes in EEG mean frequency and spectral purity during spontaneous alpha blocking. ." Electroencephalography and clinical neurophysiology 76 , no. 3 ( September 1990 ): 197-204 .
55
SJBSTiT UTE SHEET (RULE 2&) Guignard, B. . " Monitoring analgesia. ." Best practice and research. Clinical anaesthesiology 20 , no. 1 ( March 2006 ): 161-180 .
Guo, Y. , T. Hastie, and R. Tibshirani. " Regularized linear discriminant analysis and its application in microarrays ." Biostatistics ( Oxford University Press ) 8 , no. 1 ( January 2007 ): 86-100 .
Guyton, A. C. Human Physiology and Mechanisms ofDisease . Philadelphia : Saunders , 1982 .
Fagrell, B. Microcirculation of the skin. The physiology and pharmacology of the microcirculation . Vol. 2 . New York : Academic Press , 1984 .
Hastie, T. , R. Tibshirani, and J. Friedman. The elements of statistical learning: data mining, inference, and prediction . New York : Springer , 2001 .
Hjorth, B. . " The physical significance of time domain descriptors in EEG analysis. ." Electroencephalography and clinical neurophysiology 34 , no. 3 ( March 1973 ): 321-
325.
Kohavi, R. , and G. H. John. " Wrappers for Feature Subset Selection ." Artificial Intelligence 97 , no. 1-2 ( 1997 ): 273324 .
Lidberg, L., and G. B. Wallin. "Sympathetic Skin Nerve Discharges in Relation to Amplitude of Skin Resistance Responses." Psychophysiology 18 , no. 3 (1981): 268-270.
Malik, M. " Heart rate variability: standards of measurement, physiological interpretation and clinical use. Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. ." Circulation 93 , no. 5 ( March 1996 ): 1043-1065 .
Moghaddam, B. , Y. Weiss, and S. Avidan. " Spectral Bounds for Sparse PCA: Exact and Greedy Algorithms ." Edited by Weiss,Y., Scholkopf,B. and Platt,J., 915-922 . Cambridge, MA : MIT Press , 2006 .
Moghaddam, B. , Y. Weiss, and S. Avidan. " Generalized Spectral Bounds for Sparse LDA ." ICML . 2006a . 641-648 .
Pagani, M. , O. Rimoldi, and A. Malliani. " Low-frequency components of cardiovascular variabilities as markers of sympathetic modulation. ." Trends in pharmacological sciences 13 , no. 2 ( February 1992 ): 50-54 .
Pan, J. , and W. J. Tompkins. " A real-time QRS detection algorithm. ." IEEE Trans BiomedEng 32 , no. 3 ( March 1985 ): 230-236 .
Pop-Jordanova, N. , and J. Pop-Jordanov. " Spectrum-weighted EEG frequency ("brain- rate") as a quantitative indicator of mental arousal. ." Prilozi Makedonska akademija na
56
dϋϋS liϊ UTE SHEET (RULE 2δ) naukite i umetnostite, Oddelenie za biolovski i medicinski nauki 26 , no. 2 ( December 2005 ): 35-42 .
Schlogl, Alois . " A comparison of multivariate autoregressive estimators ." Signal Process. ( Elsevier North-Holland, Inc. ) 86 , no. 9 ( September 2006 ): 2426-2429 .
Scholkopf, B. , A. Smola, and K.-R. Muller. " Nonlinear component analysis as a kernel eigenvalue problem ." Neural Comput. ( MIT Press ) 10 , no. 5 ( 1998 ): 1299-1319 .
Seitsonen, E. R. , I. K. Korhonen, M. J. Van Gils, J. M. Lotjonen, K. T. Kortilla, and A. M. Yli-Hankala. " EEG spectral entropy, heart rate, photoplethysmography and motor responses to skin incision during sevoflurane anaesthesia ." Acta Anaesthesiol Scand , 49, ( 2005 ): 284292 .
Roweis, S. , and L. Saul. " Nonlinear dimensionality reduction by locally linear embedding ." Science 290. ( 2000 ): 2323-2326 .
Tenenbaum, J. , V. de Silva, and J. C. Langford. " A global geometric framework for nonlinear dimensionality reduction ." Science 290 ( 2000 ): 23192323 .
Thomas, D. W. , and J. M. Evans. " Lower oesophageal contractility monitoring during anaesthesia for cardiac surgery: preliminary observations. ." Annals of the Royal College of Surgeons of England 1\ , no. 5 ( September 1989 ): 311-315 .
Tibshirani, R. , T. Hastie, B. Narasimhan, and G. Chu. " Diagnosis of multiple cancer types by shrunken centroids of gene expression ." PNAS 99 , no. 10 ( May 2002 ): 6567- 6572 . van den Berg, A. A. , D. Sawa, and N. M. Honjol. " Attenuation of the haemodynamic responses to noxious stimuli in patients undergoing cataract surgery. A comparison of magnesium sulphate, esmolol, lignocaine, nitroglycerine and placebo given i.v. with induction of anaesthesia ." European Journal of Anesthesiology 14 , no. 02 ( 2006 ): 134-147 .
Vapnik, V. N. Statistical Learning Theory . Edited by Haykin S. New York : Wiley , 1998 .
Wackermann, J. . " Towards a quantitative characterisation of functional states of the brain: from the non-linear methodology to the global linear description ." International Journal of P sychophysiology 34 , no. 1 ( October 1999 ): 65-80 .
Weiss, T. , A. Del Bo, N. Reichek, and K. Engelman. " Pulse transit time in the analysis of autonomic nervous system effects on the cardiovascular system. ." P sychophysiology 17 , no. 2 ( March 1980 ): 202-207 .
Zou, H. , T. Hastie, and R. Tibshirani. " Sparse Principal Component Analysis ." Journal of Computational Graphical Statistics 15 ( 2006 ): 265286.
57
SUBSTITUTE SHEET(RULE2δ)

Claims

CLAIMS:
1. A method for establishing the pain level in a patient, comprising analyzing a multidimensional array of physiological signals in order to obtain the pain level of a patient; wherein said step of analyzing comprises the steps of: a. acquiring a set of physiological signals from the body of a patient; b. processing said set of signals so as to extract a first vector of features representing the physiological status of the patient; wherein the said first vector of features comprises a Great Plurality of Features (GPF); c. reducing the dimensionality of said first vector of features by transforming the said first vector to a second vector whose dimensions are lower by at least one order of magnitude compared to said first vector; d. classifying said second vector of features into at least two classes representing at least two conditions of pain; e. representing said classes of said pain level of said patient at a given time interval thereby establishing the pain level in an awake, semi- awake or sedated patient.
2. A method according to claim 1, wherein said method additionally comprises a training step; said training step comprising; a. acquiring said set of physiological signals from said body of a patient or group of patients in a first non-pain state and a second pain state; b. processing said set of signals so as to extract a first vector of features representing the physiological status of said patient; wherein said first vector of features comprises a Great Plurality of Features (GPF); c. firstly, learning the parameters of said reducing of dimensionality by learning parameters of transformation of said first vector of features to a second vector whose dimensions are lower by at least one order of magnitude compared to said first vector; d. secondly, learning the parameters of a classifier that classifies said second vector of features into at least two classes representing at least two conditions of pain; e. setting the parameters of said classifier thereby establishing a classifier which classifies said pain level in an awake, semi- awake or sedated patient.
3. The method according to claim 1 or 2, wherein said step of acquiring comprises selecting data from the group consisting of data supplied by the physician's, environmental parameters, patient parameters or any combination thereof
4. The method according to claim 1, 2 or 3, wherein said extracted features are selected from Table 2.
5. The method according to claim 1,2 or 3 wherein said physiological signals represent an activity selected from the group consisting of autonomic nervous system activity, muscular activity, and brain activity.
6. The method according to claim 1, 2 or 3, wherein said physiological signals are selected from a group comprising ECG, PPG, continuous blood pressure, respiration, internal or skin temperature, EOG or pupil diameter monitoring, GSR, EEG, EMG, EGG, LDV, capnograph and accelerometer or any combination thereof
7. The method according to claim 1, 2 or 3 wherein said step of processing comprises analyzing the artifacts occurrence in said acquired signals.
8. The method according to claim 1, 2 or 3 wherein said step of representing said pain level of said patient is provided continuously during at least one predetermined time interval.
9. The method according to claim 1, 2 or 3 wherein said step of representing a PAIN or NON-PAIN condition of said patient is provided in a graduated scale.
10. The method according to claim 1, 2 or 3 wherein said steps of reducing dimensionality or learning the parameters of reduction of dimensionality of said first vector of features further comprises the steps of: a. calculating extracted feature scores for each of said features or combination of features; b. filtering out said extracted low-score features thereby decreasing the number of said features to a predetermined number;
11. The method according to claim 10 wherein said calculating of said extracted features scores is based on activating said classifier processing and examining its results.
12. The method according to claim 1, 2 or 3 wherein said steps of reducing dimensionality or learning the parameters of reduction of dimensionality of said first vector of features is achieved by obtaining sets of features and combining each of said
59
OiJCSIiTUTE SHEET(RULE 2E) sets into one meta feature activating a linear or non-linear function on said set, thereby reducing the dimensions of features vector.
13. The method according to claim 12 wherein said steps of combining and activating a linear or non-linear function is achieved by applying statistical methods selected from the group consisting of MDS, PCA, Sparse PCA, FLDA, Sparse LDA, Kernel PCA, ISOMAP, LLE, Laplacian Eigenmaps, Diffusion Maps, Hessian Eigenmaps or any combination thereof.
14. The method according to claim 1, 2 or 3 wherein said classifying is obtained by applying a function on said second features vector which returns the pain class number.
15. The method according to claim 1, 2 or 3, wherein said classifying is obtained by comparing said second features vector to at least one threshold surface which is defined by said learned parameters of classification in the space or subspace of said features.
16. The method according to claim 15, wherein the step of comparing is performed by applying linear or non-linear operators or functions on said second features vector.
17. The method of according to claim 1, 2 or 3 wherein said steps of classifying or learning parameters of classification are achieved by applying statistical methods selected from a group consisting of Boosting, Linear classifier, Naϊve Bayes Classifier, k-nearest neighbor classifier, QDA classifier, RBF classifier, Multilayer Perceptron classifier, Bayesian Network classifier, Bagging classifier, SVM, NC, NCS, LDA, SCRLDA, Random Forest, or Committee of classifiers or any combination thereof.
18. The method according to claim 1, 2 or 3 wherein said steps of classifying comprises the step of computing a confidence value of said second feature vector.
19. A system for establishing the pain level in a patient wherein said system comprises; a. acquiring means for acquiring a set of physiological signals as input; b. memory means coupled to a microprocessor for analyzing said set of physiological signals to obtain the pain level of the patient; c. representing means for representing said pain level of the patient as an output wherein said acquiring means communicates said physiological signals to said memory means coupled to said microprocessor. further wherein said memory means coupled to said microprocessor communicates said pain level of the patient to said representing means; further wherein said memory means coupled to said microprocessor is provided with; i. processing means for processing said set of signals so as to extract a first vector of features representing the physiological status of the patient; wherein the said first vector of features comprises a Great Plurality of Features (GPF), ii. reducing means for reducing the dimensionality of the said first vector of features by transforming to a second vector whose dimensions are lower by at least one order of magnitude compared to said first vector; iii. classifying means for classifying said second vector of features into at least two classes representing at least two conditions of pain.
20. A system according to claim 19, wherein said system additionally comprises a training subsystem; wherein said training subsystem comprises; a. acquiring means for acquiring a set of training physiological signals of a patient in a first non-pain state and a second pain state as training input; b. memory means coupled to a microprocessor for analyzing said set of training physiological signals to training and setting system parameters; c. representing means for representing said trained parameters and analysis results as an output; wherein said acquiring means communicates said physiological signals to said memory means coupled to said microprocessor further wherein said memory means coupled to said microprocessor communicates said trained parameters and analysis results to said representing means further wherein said memory means coupled to a microprocessor means is provided with; i. processing means for processing said set of signals so as to extract a first vector of features representing the physiological status of the patient; wherein the said first vector of features comprises a Great Plurality of Features (GPF),
61
SUBSTITUTE SHEET (RULE 2&) ii. first learning means for training said reducing of dimensionality by learning parameters of transformation of the said first vector of features to a second vector whose dimensions are lower by at least one order of magnitude compared to said first vector iii. second learning means for learning the parameters of classifier that classifying said second vector of features into at least two classes representing at least two conditions of pain iv. setting means for setting the parameters of said classifier thereby establishing a classifier which classifies said pain level in an awake, semi- awake or sedated patient.
21. The system according to claim 19 or 20, wherein said acquiring means comprises sensors attached to the body of said patient for detecting said physiological signals.
22. The system according to claim 19 or 20 wherein said representing means is selected from the group consisting of computer screen, PDA screen, TV screen , plasma screen, LCD screen, patient monitor or any means for displaying numbers or graphs in a continuous manner
23. The system according to claim 19 or 20 wherein said communication is by cable, wireless, blue tooth infra red, WI-FI or internet means.
24. The system according to claim 19 or 20 wherein said acquiring means is provided with means for selecting data from the group consisting of data supplied by the physician's, environmental parameters, patient parameters or any combination thereof.
25. The system according to claim 19, 20 or 24 wherein said system further comprises input means for entering patient information and other data supplied by the physician.
26. The system according to claim 19, 20, 24, or 25 wherein said physiological signals represent an activity selected from the group consisting of autonomic nervous system activity, muscular activity, and brain activity.
27. The system according to claim 19, 20, 24, or 25 wherein said physiological signals are selected from a group comprising ECG, PPG, continuous blood pressure, respiration, internal or skin temperature, EOG or pupil diameter monitoring, GSR, EEG, EMG, EGG, LDV, capnograph and accelerometer or any combination thereof.
62
SUBSTITUTE SHEET (RULE 22)
28. The system according to claim 19, 20, 24, or 25 wherein said processing means comprises means for analyzing the artifacts occurrence in said acquired signals.
29. The system according to claim 19, 20, 24, or 25 wherein said representing means is adapted to represent said pain level of said patient continuously during at least one predetermined time interval.
30. The system according to claim 19, 20, 24, or 25 wherein said representing means is adapted to provide the PAIN or NON-PAIN condition of said patient in a graduated scale.
31. The system according to claim 19, 20, 24, or 25 wherein said reducing means or first learning means are adapted for; a. calculating said extracted feature scores for each of said features or combination of features and, b. filtering out said extracted low-score features thereby decreasing the number of said features to a predetermined number.
32. The system according to claim 19, 20, 24, or 25 wherein said reducing means or first learning means are further adapted for calculating said extracted features score based on activating said classifier processing and examining its results.
33. The system according to claim 19, 20, 24, or 25 wherein said reducing means of said first vector of features or first learning means are provided with obtaining means for obtaining sets of features and combining each said set into one meta feature by activating a linear or non linear function on said set thereby reducing the dimensions of said first features vector.
34. The system according to claim 33 wherein said obtaining means is adapted for combining and activating a linear or non linear function by applying statistical methods selected from the group consisting of MDS, PCA, Sparse PCA, FLDA, Sparse LDA, Kernel PCA, ISOMAP, LLE, Laplacian Eigenmaps, Diffusion Maps, Hessian Eigenmaps or any combination thereof.
35. The system according to claim 19, 20, 24, or 25 wherein said classifying means is adapted for classifying by applying a function on said second features vector which returns the pain class number.
36. The system according to claim 19, 20, 24, or 25 wherein said classifying means is further adapted to classify by comparing said features to at least one threshold surface in the space or subspace of the said features.
37. The system according to claim 36 wherein said classifying means is adapted to compare by applying linear or non linear operators or functions on said features.
38. The system according to claim 19, 20, 24, or 25 wherein said classifying means or second learning means are adapted to apply statistical methods selected from the group consisting of Boosting, Linear classifier, Naϊve Bayes Classifier, k-nearest neighbor classifier, QDA classifier, RBF classifier, Multilayer Perceptron classifier, Bayesian Network classifier, Bagging classifier, SVM, NC, NCS, LDA, SCRLDA, Random Forest, or Committee of classifiers or any combination thereof
39. The system according to claim 19, 20, 24, or 25 wherein said classifying means or second learning means are adapted to compute a confidence value of said vector.
PCT/IL2008/001493 2007-11-14 2008-11-13 Pain monitoring using multidimensional analysis of physiological signals WO2009063463A2 (en)

Priority Applications (4)

Application Number Priority Date Filing Date Title
US12/779,963 US8512240B1 (en) 2007-11-14 2010-05-14 System and method for pain monitoring using a multidimensional analysis of physiological signals
US13/945,657 US9498138B2 (en) 2007-11-14 2013-07-18 System and method for pain monitoring using a multidimensional analysis of physiological signals
US15/349,098 US10743778B2 (en) 2007-11-14 2016-11-11 System and method for pain monitoring using a multidimensional analysis of physiological signals
US16/983,466 US11259708B2 (en) 2007-11-14 2020-08-03 System and method for pain monitoring using a multidimensional analysis of physiological signals

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US98778207P 2007-11-14 2007-11-14
US60/987,782 2007-11-14

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US12/779,963 Continuation-In-Part US8512240B1 (en) 2007-11-14 2010-05-14 System and method for pain monitoring using a multidimensional analysis of physiological signals

Publications (2)

Publication Number Publication Date
WO2009063463A2 true WO2009063463A2 (en) 2009-05-22
WO2009063463A3 WO2009063463A3 (en) 2010-01-28

Family

ID=40639263

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/IL2008/001493 WO2009063463A2 (en) 2007-11-14 2008-11-13 Pain monitoring using multidimensional analysis of physiological signals

Country Status (1)

Country Link
WO (1) WO2009063463A2 (en)

Cited By (30)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2011044228A1 (en) * 2009-10-09 2011-04-14 Brainscope Company, Inc. System and method for pain monitoring at the point-of-care
WO2012104853A3 (en) * 2011-02-03 2013-01-03 The Medical Research, Infrastructure, And Health Services Fund Of The Tel Aviv Medical Center Method and system for use in monitoring neural activity in a subject's brain
US8364254B2 (en) 2009-01-28 2013-01-29 Brainscope Company, Inc. Method and device for probabilistic objective assessment of brain function
WO2014199093A1 (en) * 2013-06-14 2014-12-18 Centre Hospitalier Regional Universitaire De Lille Device for assessing medication and/or paramedical care requirements, assessment method for implementing the assessment device, and associated delivery device
CN104887198A (en) * 2014-03-06 2015-09-09 中国科学院沈阳自动化研究所 Pain quantitative analysis system and method based on human body physiological signal multi-parameter fusion
US20150265208A1 (en) * 2014-03-18 2015-09-24 Oridion Medical 1987 Ltd. Systems and methods for pain measurement
EP2658437A4 (en) * 2010-12-29 2016-05-18 Intel Corp Integrated biometric sensing and display device
CN106510625A (en) * 2016-10-27 2017-03-22 广州医科大学附属肿瘤医院 Pain feedback analyzing and calculating method and system
CN106951711A (en) * 2017-03-22 2017-07-14 中国科学院深圳先进技术研究院 A kind of appraisal procedure and system of patients with lumbar disc herniation massage effect
WO2017157746A1 (en) 2016-03-15 2017-09-21 Koninklijke Philips N.V. A method and apparatus for determining a baseline for one or more physiological characteristics of a subject
CN107205676A (en) * 2015-01-08 2017-09-26 美达森斯生物测定有限公司 Equipment for the electrod-array of physiology monitoring and including or using electrod-array
CN107315479A (en) * 2017-07-06 2017-11-03 哈尔滨工业大学 Myoelectricity real-time operation device based on laser projection
CN107440687A (en) * 2017-08-09 2017-12-08 中国科学院深圳先进技术研究院 A kind of pain grade appraisal procedure and use its pain grade apparatus for evaluating
WO2018170792A1 (en) * 2017-03-22 2018-09-27 中国科学院深圳先进技术研究院 Method and system for assessing effectiveness of tui na on lumbar disc herniation
US10321840B2 (en) 2009-08-14 2019-06-18 Brainscope Company, Inc. Development of fully-automated classifier builders for neurodiagnostic applications
EP3498169A1 (en) * 2017-12-13 2019-06-19 Tata Consultancy Services Limited System and method for classification and quantitative estimation of cognitive stress
CN110025322A (en) * 2019-04-11 2019-07-19 辽宁师范大学 Multi-modal physiological signal sensibility classification method based on filtering with integrated classifier
CN110200617A (en) * 2019-06-26 2019-09-06 慧言科技(天津)有限公司 A kind of system using heart rate variability quantitative assessment pain
CN110516762A (en) * 2019-10-10 2019-11-29 深圳大学 A kind of muscular states quantization assessment method, apparatus, storage medium and intelligent terminal
WO2020089576A1 (en) * 2018-10-30 2020-05-07 Oxford University Innovation Limited Method and apparatus for monitoring a patient
CN111133526A (en) * 2017-07-18 2020-05-08 生命分析有限公司 Discovering novel features that can be used in machine learning techniques, e.g. for diagnosing medical conditions
CN111110221A (en) * 2019-12-26 2020-05-08 天津大学 Time-frequency-nonlinear multidimensional body surface gastric electrical feature extraction method
CN111345803A (en) * 2020-03-20 2020-06-30 浙江大学城市学院 Heart rate variability measuring method based on mobile device camera
WO2021005271A1 (en) * 2019-07-10 2021-01-14 Kipuwex Oy Apparatus for determining indication of pain
US11039792B2 (en) 2017-05-25 2021-06-22 Tata Consultancy Services Limited System and method for heart rate estimation
US11062792B2 (en) 2017-07-18 2021-07-13 Analytics For Life Inc. Discovering genomes to use in machine learning techniques
EP3936175A1 (en) 2014-06-30 2022-01-12 Syqe Medical Ltd. Methods, devices and systems for pulmonary delivery of active agents
IT202000017212A1 (en) * 2020-07-15 2022-01-15 Andrea Chellini EQUIPMENT FOR THE MEASUREMENT OF PHYSIOLOGICAL PARAMETERS.
US11304663B2 (en) 2017-12-29 2022-04-19 Tata Consultancy Services Limited Systems and methods for detecting anomaly in a cardiovascular signal using hierarchical extremas and repetitions
RU2774207C1 (en) * 2021-09-27 2022-06-16 федеральное государственное автономное образовательное учреждение высшего образования "Санкт-Петербургский политехнический университет Петра Великого" (ФГАОУ ВО "СПбПУ") Method for research and objective registration of pain syndrome

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070010723A1 (en) * 2005-07-05 2007-01-11 Kimmo Uutela Determination of the clinical state of a subject
US20070219433A1 (en) * 2004-07-10 2007-09-20 Stupp Steven E Apparatus for providing information based on association variables

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070219433A1 (en) * 2004-07-10 2007-09-20 Stupp Steven E Apparatus for providing information based on association variables
US20070010723A1 (en) * 2005-07-05 2007-01-11 Kimmo Uutela Determination of the clinical state of a subject

Cited By (44)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8364254B2 (en) 2009-01-28 2013-01-29 Brainscope Company, Inc. Method and device for probabilistic objective assessment of brain function
US10321840B2 (en) 2009-08-14 2019-06-18 Brainscope Company, Inc. Development of fully-automated classifier builders for neurodiagnostic applications
WO2011044228A1 (en) * 2009-10-09 2011-04-14 Brainscope Company, Inc. System and method for pain monitoring at the point-of-care
EP2658437A4 (en) * 2010-12-29 2016-05-18 Intel Corp Integrated biometric sensing and display device
WO2012104853A3 (en) * 2011-02-03 2013-01-03 The Medical Research, Infrastructure, And Health Services Fund Of The Tel Aviv Medical Center Method and system for use in monitoring neural activity in a subject's brain
US10893822B2 (en) 2011-02-03 2021-01-19 The Medical Research, Infrastructure and Health Services Fund of the Tel Aviv Medical Center Method and system for use in monitoring neural activity in a subject's brain
JP2016530594A (en) * 2013-06-14 2016-09-29 セントル ホスピタリエ リージョナル ユニヴェルシテール ドゥ リールＣｅｎｔｒｅ Ｈｏｓｐｉｔａｌｉｅｒ Ｒｅｇｉｏｎａｌ Ｕｎｉｖｅｒｓｉｔａｉｒｅ ｄｅ Ｌｉｌｌｅ Evaluation device for essential requirements of pharmacotherapy and / or paramedical care, evaluation method for executing the evaluation device, and related delivery device
FR3006880A1 (en) * 2013-06-14 2014-12-19 Ct Hospitalier Regional Et Universitaire De Lille DEVICE FOR ASSESSING NEEDS IN MEDICAMENTS AND / OR PARAMEDICAL CARE, EVALUATION METHOD FOR IMPLEMENTING THE EVALUATION DEVICE AND ASSOCIATED DELIVERY DEVICE.
WO2014199093A1 (en) * 2013-06-14 2014-12-18 Centre Hospitalier Regional Universitaire De Lille Device for assessing medication and/or paramedical care requirements, assessment method for implementing the assessment device, and associated delivery device
FR3006879A1 (en) * 2013-06-14 2014-12-19 Chru Lille DEVICE AND METHOD FOR EVALUATING THE NEEDS OF ANALGESIC AND / OR HYPNOTIC COMPOUNDS OF A PATIENT PLACED UNDER ANESTHESIA OR SEDATION AND ASSOCIATED DELIVERY DEVICE
CN104887198A (en) * 2014-03-06 2015-09-09 中国科学院沈阳自动化研究所 Pain quantitative analysis system and method based on human body physiological signal multi-parameter fusion
US20150265208A1 (en) * 2014-03-18 2015-09-24 Oridion Medical 1987 Ltd. Systems and methods for pain measurement
WO2015140777A1 (en) * 2014-03-18 2015-09-24 Oridion Medical 1987 Ltd. Systems and methods for pain measurement
EP3936175A1 (en) 2014-06-30 2022-01-12 Syqe Medical Ltd. Methods, devices and systems for pulmonary delivery of active agents
CN107205676A (en) * 2015-01-08 2017-09-26 美达森斯生物测定有限公司 Equipment for the electrod-array of physiology monitoring and including or using electrod-array
CN107205676B (en) * 2015-01-08 2021-06-18 美达森斯生物测定有限公司 Electrode array for physiological monitoring and device comprising or utilizing the electrode array
WO2017157746A1 (en) 2016-03-15 2017-09-21 Koninklijke Philips N.V. A method and apparatus for determining a baseline for one or more physiological characteristics of a subject
CN108778099A (en) * 2016-03-15 2018-11-09 皇家飞利浦有限公司 The method and apparatus of the baseline of one or more physiological properties for determining object
CN106510625A (en) * 2016-10-27 2017-03-22 广州医科大学附属肿瘤医院 Pain feedback analyzing and calculating method and system
CN106951711A (en) * 2017-03-22 2017-07-14 中国科学院深圳先进技术研究院 A kind of appraisal procedure and system of patients with lumbar disc herniation massage effect
WO2018170792A1 (en) * 2017-03-22 2018-09-27 中国科学院深圳先进技术研究院 Method and system for assessing effectiveness of tui na on lumbar disc herniation
CN106951711B (en) * 2017-03-22 2019-04-26 中国科学院深圳先进技术研究院 A kind of appraisal procedure and system of patients with lumbar disc herniation massage effect
US11039792B2 (en) 2017-05-25 2021-06-22 Tata Consultancy Services Limited System and method for heart rate estimation
CN107315479A (en) * 2017-07-06 2017-11-03 哈尔滨工业大学 Myoelectricity real-time operation device based on laser projection
US11139048B2 (en) 2017-07-18 2021-10-05 Analytics For Life Inc. Discovering novel features to use in machine learning techniques, such as machine learning techniques for diagnosing medical conditions
EP3655895A4 (en) * 2017-07-18 2021-04-21 Analytics For Life Inc. Discovering novel features to use in machine learning techniques, such as machine learning techniques for diagnosing medical conditions
CN111133526A (en) * 2017-07-18 2020-05-08 生命分析有限公司 Discovering novel features that can be used in machine learning techniques, e.g. for diagnosing medical conditions
CN111133526B (en) * 2017-07-18 2024-04-12 生命分析有限公司 Novel features useful in machine learning techniques, such as machine learning techniques for diagnosing medical conditions
US11062792B2 (en) 2017-07-18 2021-07-13 Analytics For Life Inc. Discovering genomes to use in machine learning techniques
CN107440687A (en) * 2017-08-09 2017-12-08 中国科学院深圳先进技术研究院 A kind of pain grade appraisal procedure and use its pain grade apparatus for evaluating
EP3498169A1 (en) * 2017-12-13 2019-06-19 Tata Consultancy Services Limited System and method for classification and quantitative estimation of cognitive stress
US11304663B2 (en) 2017-12-29 2022-04-19 Tata Consultancy Services Limited Systems and methods for detecting anomaly in a cardiovascular signal using hierarchical extremas and repetitions
WO2020089576A1 (en) * 2018-10-30 2020-05-07 Oxford University Innovation Limited Method and apparatus for monitoring a patient
CN110025322A (en) * 2019-04-11 2019-07-19 辽宁师范大学 Multi-modal physiological signal sensibility classification method based on filtering with integrated classifier
CN110200617A (en) * 2019-06-26 2019-09-06 慧言科技(天津)有限公司 A kind of system using heart rate variability quantitative assessment pain
WO2021005271A1 (en) * 2019-07-10 2021-01-14 Kipuwex Oy Apparatus for determining indication of pain
CN110516762A (en) * 2019-10-10 2019-11-29 深圳大学 A kind of muscular states quantization assessment method, apparatus, storage medium and intelligent terminal
CN110516762B (en) * 2019-10-10 2022-11-15 深圳大学 Muscle state quantitative evaluation method and device, storage medium and intelligent terminal
CN111110221A (en) * 2019-12-26 2020-05-08 天津大学 Time-frequency-nonlinear multidimensional body surface gastric electrical feature extraction method
CN111345803A (en) * 2020-03-20 2020-06-30 浙江大学城市学院 Heart rate variability measuring method based on mobile device camera
CN111345803B (en) * 2020-03-20 2022-04-12 浙江大学城市学院 Heart rate variability measuring method based on mobile device camera
WO2022013901A1 (en) * 2020-07-15 2022-01-20 Chellini Andrea Apparatus and method for measuring physiological parameters
IT202000017212A1 (en) * 2020-07-15 2022-01-15 Andrea Chellini EQUIPMENT FOR THE MEASUREMENT OF PHYSIOLOGICAL PARAMETERS.
RU2774207C1 (en) * 2021-09-27 2022-06-16 федеральное государственное автономное образовательное учреждение высшего образования "Санкт-Петербургский политехнический университет Петра Великого" (ФГАОУ ВО "СПбПУ") Method for research and objective registration of pain syndrome

Also Published As

Publication number Publication date
WO2009063463A3 (en) 2010-01-28

Similar Documents

Publication Publication Date Title
US20220167858A1 (en) System and method for pain monitoring using a multidimensional analysis of physiological signals
WO2009063463A2 (en) Pain monitoring using multidimensional analysis of physiological signals
Uçar et al. Automatic sleep staging in obstructive sleep apnea patients using photoplethysmography, heart rate variability signal and machine learning techniques
US11172835B2 (en) Method and system for monitoring sleep
Hosseini et al. Emotional stress recognition system using EEG and psychophysiological signals: Using new labelling process of EEG signals in emotional stress state
EP1562472B1 (en) Method, apparatus and system for characterizing sleep
US10213152B2 (en) System and method for real-time measurement of sleep quality
US11259708B2 (en) System and method for pain monitoring using a multidimensional analysis of physiological signals
Akbulut et al. Bimodal affect recognition based on autoregressive hidden Markov models from physiological signals
Anbalagan et al. Analysis of various techniques for ECG signal in healthcare, past, present, and future
Fernandez Rojas et al. A systematic review of neurophysiological sensing for the assessment of acute pain
Subasi Biomedical signal analysis and its usage in healthcare
Jegan et al. Mental Stress Detection and Classification using SVM Classifier: A Pilot Study
Moradhasel et al. Spectrogram classification of patient chin electromyography based on deep learning: A novel method for accurate diagnosis obstructive sleep apnea
Moradhasel et al. Chin electromyogram, an effectual and useful biosignal for the diagnosis of obstructive sleep apnea
Farooq et al. Classification of Sleep using Polysomnography
Mozafari et al. A Fusion Model for Cross-Subject Stress Level Detection Based on Transfer Learning
Nirabi et al. Eeg signal analysis for mental stress classification: A review
Zoubek Automatic classification of human sleep recordings combining artifact identification and relevant features selection
D'Alpaos AI for health: Using AI to identify stress from wearable devices data
Melek Comparison of EEG and EOG signals in classification of sleep stages
Yücelbaş et al. Identification of full-night sleep parameters using morphological features of ECG signals: A practical alternative to EEG and EOG signals
Piartli Personalized deep learning models for pain level classification from photoplethysmography signals
Hartmann Big data analysis of cyclic alternating pattern during sleep using deep learning
Correia Assessing Emotional States from Biosignals in Autistic and Neurotypically Developed Individuals

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 08849183

Country of ref document: EP

Kind code of ref document: A2

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 08849183

Country of ref document: EP

Kind code of ref document: A2