KR20190049430A - Pain depth classification apparatus and method using PPG and heart rate variability - Google Patents

Abstract

The present invention relates to a pain classification method using pulse wave and heart rate variability (HRV), which detects the time-domain characteristics of a pulse wave signal, detects the time- and frequency-domain characteristics of the HRV, normalizes the characteristics to have a value between 0 and 1, labels the normalized data with an output variable for each pain depth group, and performs machine learning pattern analysis through a pattern classifier to classify pain depths. The present invention comprises: a signal detecting step in which an operation processing unit detects a pulse wave time-domain characteristic including an average amplitude of a pulse wave from a pulse wave signal received from a pulse wave sensor, detects the HRV, RR intervals, detects, from the HRV, an HRV time-domain characteristic including an average RR interval, and detects an HRV frequency-domain characteristic including VLF intensity, a frequency intensity of 0.003 to 0.04 Hz of the HRV; a normalizing step in which an operation processing unit normalizes the pulse wave time-domain characteristic, the HRV time-domain characteristic, and the HRV frequency-domain characteristic, which are detected in the signal detecting step, so that the characteristics have the value between 0 and 1; a labeling step in which the operation processing unit labels the pulse wave time-domain characteristic, the HRV time-domain characteristic, and the HRV frequency-domain characteristic, which are normalized in the normalizing step, with an output variable for each pain depth group; and a pain depth classifying step in which the operation processing unit inputs, in a pattern classifier, a feature vector having the output variable for each pain depth group outputted in the labeling step to perform machine learning pattern analysis, thereby classifying the pain depths.

Description

[0002] Pain classification apparatus and method using PPG and heart rate variability [

The present invention detects time-domain characteristics of a pulse wave signal, detects time-domain and frequency-domain features of heartbeat-variability, normalizes these characteristics to have a value between 0 and 1, and outputs the normalized data to the intensity- To a method of classifying pain using pulse waves and heart beat variability, which classifies pain intensity by performing labeling with a variable and performing machine learning pattern analysis through a pattern classifier.

Generally, depending on the cause of the pain, the pain can be divided into nociceptive pain, neurogenic pain, non-tempered pain or psychogenic pain. It is divided into somatic pain, visceral pain, and central pain depending on the site of occurrence, and the somatic tub is divided into the surface penetration deep tub. It also classifies pain as either a fast pain or a slow pain.

 The depth of the pain is classified according to the site of occurrence. Generally, the pain varies depending on the subject of the patient, so it is very difficult to classify it.

 At present, there is no way to classify the depth of pain as an overview, numerical, or exact.

 Therefore, there is a need for an objective, numerical, simple, and highly accurate bio-signal-based pain intensity classification apparatus and method for determining the depth of pain.

Prior art is disclosed in Korean Patent Laid-Open No. 10-2006-0017510, entitled " Device and method for layering the risk of patients suffering from chest pain caused by suspected heart disease. In addition to the ECG of the patient, the present invention obtains a clinical result by at least one in vitro diagnostic assay, obtains a risk stratification result by comparing and analyzing the ECG result and the clinical result, does not analyze the depth of the pain, There is a hassle to get separately.

Another prior art is Korean Patent No. 10-1000761, entitled " Device and method for measuring the pain / consciousness level of a surgical patient. The apparatus for measuring the pain / consciousness level of a surgical patient according to the present invention is characterized in that a stimulation signal is given to a predetermined region of a surgical patient and a potential for the stimulation signal is detected at another region to obtain a somatic sensory evoked potential (SSEP) Generate pain / conscious level information of the surgical patient. That is, the present invention detects the pain response by detecting a physical response to the stimulus by directly stimulating the patient, but the pain felt by the patient is subjective and may be different from that of the patient, thus doubling the objectivity and accuracy. Also by stimulating, general patients can be avoided.

SUMMARY OF THE INVENTION The object of the present invention is to detect time-domain characteristics of a pulse wave signal, to detect time-domain and frequency-domain features of HRV, to normalize these characteristics to have values between 0 and 1, And classifying the pain intensity by performing a machine learning pattern analysis through a pattern classifier, and to provide a method of classifying pain using pulse waves and heart beat variability.

Another problem to be solved by the present invention is to provide a time domain characteristic of a pulse wave signal in which an average signal amplitude, an average rise time, an average fall time, and an average heart rate are obtained and an average RR interval, SDNN, rMSSD, The VLF component, the LF component, the HF component, and the LF / HF are detected by the frequency domain characteristics of heart beat variability, and these results are stored in a pattern classifier based on Multi-layer Perceptron Neural Network, Support A pattern classifier based on a Vector Machine (SVM), and a pattern classifier based on a Deep Belief Network, to classify pain depths, and a method for classifying pain using pulse waves and heart beat variability.

In order to solve the above problems, the present invention provides a blood pressure monitor for detecting pulse wave time-domain characteristics including an average amplitude of a pulse wave from a pulse wave signal received from a pulse wave sensor of an arithmetic processing unit and detecting a heart rate variation as an RR interval, A signal detection step of detecting a heart rate variability frequency domain feature including a mean RR interval and a VLF intensity that is a frequency intensity of 0.003 to 0.04 Hz of a heart beat variance; A normalizing step of normalizing the pulse-wave time-domain characteristic, the heartbeat-variability time-domain characteristic, and the heartbeat-variability frequency-domain characteristic detected in the signal detection step so as to have a value between 0 and 1; A labeling step of labeling the pulse wave time domain characteristic, the heartbeat mutation time domain characteristic, and the heartbeat mutation frequency domain characteristic normalized in the normalization step as an output variable for each pain intensity group in the normalization step; The arithmetic processing unit includes a pain intensity classifying step of classifying the pain intensity by performing a machine learning pattern analysis by inputting a feature vector having an output variable for each pain intensity group output in the labeling step into a pattern classifier.

In the signal detecting step, the arithmetic processing unit calculates a pulse wave signal average amplitude (Mean Amplitude), a pulse wave rise time (pulse wave rising time Rise time, a pulse time fall time, and an average heart rate of the pulse wave signal; After the detection of the time-domain characteristic of the pulse wave signal, the calculation processing section calculates the R-point of the next cycle from the R-point (maximum point) of each cycle in the pulse wave signal received from the pulse wave sensor or received from the pulse- An RR interval detecting step of obtaining an RR interval (heart beat variation) After the RR interval detection step, the arithmetic processing unit calculates the average square root of the mean of the values obtained by squaring the difference between the adjacent RR intervals, the mean RR interval (Mean RRI), the standard deviation (SDNN) of the total RR interval, the number of RR intervals exceeding 20 ms (NN20), the number of RR intervals exceeding 50 ms (NN50), the ratio of the total number of RR intervals exceeding 20 ms (pNN20), the total number of RR intervals And calculating a ratio (pNN50) of the number of RR intervals exceeding 50 ms in the heart beat variability time domain feature extraction step.

In the signal detection step, the arithmetic processing unit removes the DC value of the RR interval through the baseline correction at the RR interval values and the heartbeat deviation values obtained at the RR interval detection step, and performs a Cubic interpolation (cubic interpolation) A baseline correction and a cubic interpolation step, The arithmetic processing unit applies Fast Fourier Transform (FFT) by applying a 50% overlapped Hamming window to the RR interval value subjected to baseline correction and cubic interpolation in the baseline correction and cubic interpolation steps, The FFT step; A very low frequency band (VLF) intensity of 0.003 to 0.04 Hz, a low frequency band (LF) intensity of 0.04 to 0.15 Hz, a frequency of 0.15 (HRV) frequency domain feature extraction step for obtaining a high frequency band (HF) intensity of about 0.4 Hz and a band intensity ratio (LF / HF ratio) of a low frequency band and a high frequency band.

The arithmetic processing section obtains the amplitude between the maximum point Peak and the minimum point in one cycle of the pulse wave signal by the pulse wave signal amplitude and obtains the average pulse wave signal amplitude as the average pulse wave signal amplitude for five minutes, , The average time taken to rise from the minimum point (valley) to the maximum point (peak) in each cycle of the pulse wave signal for 5 minutes, and the pulse wave fall time is the minimum point (Peak) Valley) is the average of the time taken to descend.

The arithmetic processing unit calculates the time domain characteristics of the pulse wave signal, the heart rate variability (HRV) time domain features, and the HRV frequency domain features (D _{normalization} (i))

(Note that, D _raw (i) is the time domain of the pulse wave signal characteristics, heart rate variability (HRV) time domain features, and heart rate variability (HRV) is the one of the frequency-domain features, one feature value of, D _max (i ) _{Represents the} maximum value among the characteristics representing D _raw (i), and D _min (i) represents the minimum value among the characteristics representing D _raw (i)).

In the pain intensity classification step, the normalized feature vector, which is labeled in the operation unit labeling step, is divided into a pattern classifier based on a Multi-layer Perceptron Neural Network and a Radial Basis Function kernel The patterns are classified into Support Vector Machine (SVM) based pattern classifier and Deep Belief Network (DBN) based pattern classifier for machine learning pattern analysis.

The present invention is characterized by a recording medium storing a computer program source for a method of classifying a pain using a pulse wave and heart beat variability.

A signal preprocessing unit for amplifying a pulse wave received from the pulse wave sensor and removing noise and converting the pulse wave into a digital signal, And an arithmetic processing unit for receiving a pulse wave and including an average amplitude of a pulse wave and detecting an pulse wave time domain characteristic including an average amplitude of the pulse wave, wherein the arithmetic processing unit calculates, from the pulse wave signal received from the signal preprocessing unit, Detecting heart beat variability, detecting heart rate variability time domain features including the mean RR interval from heart beat variability, detecting heart rate variability frequency domain features including VLF intensity of 0.003-0.04 Hz of heart rate variability, and detecting Heart rate variability, time domain characteristics, heart beat variability, and frequency domain characteristics between 0 and 1 And labeling the normalized pulse wave time domain characteristic, the heartbeat mutation time domain characteristic, and the heartbeat mutation frequency domain characteristic as the output variable for each of the pain depth groups, and the characteristic vector having the output variable for each of the pain depth group is classified into the pattern classifier And the machine learning pattern analysis is performed.

The arithmetic processing section obtains a pulse wave signal mean amplitude, a pulse wave rise time, a pulse wave fall time and an average heart rate as time domain characteristics of a pulse wave signal, (Heart beat variation) that is an interval from the R point (maximum point) of each cycle to the R point of the next cycle in the pulse wave signal received from the pulse wave sensor or received from the pulse wave sensor and temporarily stored in the memory section, The processor calculates the mean square of the mean of the mean RR interval (SD), the standard deviation of the total RR interval (SDNN), the mean square of the difference between adjacent RR intervals (rMSSD), which is a characteristic of the HRV time domain, The number of RR intervals (NN20) exceeding 50 ms, the number of RR intervals exceeding 50 ms (NN50), the ratio of the number of RR intervals exceeding 20 ms among the total number of RR intervals (pNN20) (PNN50).

The arithmetic processing unit removes the DC value of the RR interval through the baseline correction at the RR interval values and the heartbeat deviation values, performs cubic interpolation (cubic interpolation) at 2 Hz, and performs the correction at the baseline correction and the cubic interpolation step Fast Fourier Transform (FFT) is performed by applying a 50% overlapped Hamming window to the RR interval values subjected to baseline correction and cubic interpolation, and the RR interval values converted into the frequency domain A very low frequency band (VLF) intensity of 0.003 to 0.04 Hz, a low frequency band (LF) intensity of 0.04 to 0.15 Hz, a high frequency band (HF) intensity of 0.15 to 0.4 Hz, Obtain the band intensity ratio (LF / HF ratio) of the frequency band high frequency.

The method for classifying pain using pulse waves and heart beat variability according to the present invention detects time-domain characteristics of a pulse wave signal, detects time-domain features and frequency-domain features of heart beat variability (HRV) And classifying the pain intensity by performing machine learning pattern analysis through a pattern classifier to classify the normalized data as an output variable for each of the pain intensity groups and to classify the pain intensity as an objective, numerical, simple, and highly accurate biological signal based pain A depth classification apparatus and method are provided.

Particularly, the present invention is characterized in that the average signal amplitude, the average rise time, the average fall time, and the average heart rate are obtained and the average RR interval, SDNN, rMSSD, NN20, NN50, pNN20 , and pNN50 were obtained. The VLF component, the LF component, the HF component, and the LF / HF were detected using the frequency domain characteristics of heart beat variability, and these results were analyzed using a Support Vector Machine (SVM) Radial Basis Function kernel based pattern classifier and Deep Belief Network based pattern classifier to classify pain depth to improve accuracy.

1 is a block diagram schematically illustrating an apparatus for classifying pain using pulse waves and heart beat variability according to the present invention.
2 is a schematic diagram for explaining time-domain characteristics of a pulse wave signal.
3 is a schematic diagram for explaining the HRV frequency domain characteristic.
4 is a flowchart schematically illustrating a method of performing pain classification using pulse waves and heart beat variability in the arithmetic processing unit 200 of FIG.

Hereinafter, a method for classifying pain intensity using pulse waves and heart beat variability according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram schematically illustrating an apparatus for classifying pain intensity using pulse waves and heartbeat variability according to the present invention. The pulse wave detector includes a pulse wave detecting unit 100, a signal preprocessing unit 120, an arithmetic processing unit 200, an output unit 250 ), And a memory unit 270.

The pulse wave detecting unit 100 includes a light emitting diode (not shown) and a light receiving unit (not shown) ). The electrocardiogram detecting unit may be provided instead of the pulse-wave detecting unit 100 according to circumstances.

The light emitting unit (not shown) may be composed of red light (wavelength: 650 to 750 nm) and light emitting diode (LED) having infrared light (wavelength: 850 to 1000 nm) and emits red light or infrared light to the blood vessel.

The light receiving unit (not shown) comprises a photodiode or a light receiving sensor (light sensor), and receives light reflected or transmitted through the blood vessel, that is, red light or infrared light, and outputs it as an electric signal.

The signal preprocessing unit 120 removes noise from the pulse wave detected by the pulse wave detecting unit 100, amplifies the signal, converts the signal into a digital signal, and transmits the digital signal to the arithmetic processing unit 200.

The arithmetic processing unit 200 includes seven arithmetic processing units including a time domain characteristic (parameter) of a pulse wave signal having four characteristics (parameters) including a pulse wave signal average amplitude and a mean RR interval (Mean RRI) The Heart Rate Variability (HRV), which consists of four characteristics, including heart rate variability (HRV) time domain characteristics and very low frequency (VLF) intensity of 0.003 to 0.04 Hz, , HRV) frequency domain characteristics. Since these extracted features (feature vectors) are all 15 features and 15 features vary from subject to subject, normalize them to eliminate such deviations and normalize each feature to have a value ranging from 0 to 1. The arithmetic processing unit 200 has the function of normalizing the normalized feature vector, that is, the normalized feature vector (input vector), to have an output variable for classification and to label the data, that is, the normalized feature vector, That is, the normalized input vectors are labeled as 0 for pain-free rest state and 1 for painful state), and the normalized feature vectors subjected to labeling are subjected to a machine learning pattern analysis To classify the pain intensity. The pattern classifier of the present invention includes a pattern classifier based on a Multi-layer Perceptron Neural Network, a Support Vector Machine (SVM) based pattern classifier using a Radial Basis Function kernel, , And Deep Belief Network (DBN) based pattern classifiers.

In other words, the normalized feature vector (input vector) must have an output variable for classification. (NRS), which expresses the degree of pain (degree), is used to label the output variable for each pain intensity group using the pain score based on the numerical rating scale (NRS) . ≪ / RTI > That is, a painless rest state is labeled 0, and a painful state is labeled 1.

The arithmetic processing unit (200) outputs the classified pain intensity to the output unit (270). The output unit 270 may be a monitor (not shown) or a printer (not shown).

The operation processing unit 200 may be a personal computer, a personal terminal, a microprocessor, a microcontroller, or the like.

A process of obtaining a time domain characteristic (parameter), a heart rate variability (HRV) time domain characteristic, and a heart rate variability (HRV) frequency domain characteristic of a pulse wave signal in the operation processing unit 200 will be described below .

FIG. 2 is a schematic diagram for explaining time-domain characteristics of a pulse wave signal. The Y-axis represents the size of the pulse wave and the X-axis represents time.

As the time domain characteristics of the pulse wave signal, the mean amplitude, pulse time, rise time, fall time, and average heart rate of the pulse wave signal as shown in FIG. 2 are obtained.

Peak and minimum points of each cycle were obtained for 5 minutes for the average amplitude of the pulse wave signal and the amplitude between the maximum point and the minimum point was determined as the pulse wave signal amplitude, The pulse wave signal amplitudes of the respective periods are averaged to obtain the mean amplitude of the pulse wave signal (Mean Amplitude).

For pulse wave rise time, calculate the average of the time taken to rise from the minimum point (peak) to the maximum point (peak) in each cycle for 5 minutes.

The pulse fall time is the average of the time taken to descend from the peak to the minimum point in each cycle for 5 minutes,

The average heart rate is determined by measuring the interval from one maximum peak (Peak) to the next maximum peak (Peak) (ie, the RR interval) as one beating interval, calculating the beats per minute, I ask.

The heart rate variability (HRV) time-domain feature extraction was performed using the mean RRI interval (SD), the standard deviation of the total RR interval (SDNN), the square root of the mean (rMSSD) of squared differences between adjacent RR intervals, The number of RR intervals exceeding 20 ms (NN20), the number of RR intervals exceeding 50 ms (NN50), the ratio of the total number of RR intervals exceeding 20 ms (pNN20), the total number of RR intervals exceeding 50 ms (PNN50) of the number of RR intervals.

The mean RR interval (Mean RRI) is the average of the RR intervals (RRI, RR Interval) in the pulse wave for 5 minutes. That is, the mean RR interval (mean RRI interval) obtains an average of the intervals from the R point (i.e., the maximum point) to the consecutive R points. The equation is expressed as follows.

은 RR 간격의 평균을 말하며, N은 RR 수열 샘플수, 즉, RR간격의 개수를 나타내며, Mean RRI는 평균 RR 간격을 말한다.In the above equation, RR denotes an RR interval,

Is the average of the RR intervals, N is the number of RR sequence samples, i.e., the number of RR intervals, and Mean RRI is the average RR interval.

The standard deviation (SDNN) of the total RR interval is the standard deviation of the RR interval (that is, the interval from the R point to the consecutive R points) in the pulse wave of 5 minutes, which can be expressed by the following equation.

The square root of the mean of the squared difference between adjacent RRs (rMSSD) is the square root of the mean square of the difference of adjacent RR intervals, ie, the difference between one RR interval and the subsequent RR interval And obtains the square root of the mean of the squared values. The equation is expressed as follows.

The number of RR intervals (NN20) exceeding 20 ms is obtained by counting the number of RR intervals (RRI) in which the RR interval exceeds 20 ms in a 5-minute pulse wave.

That is, the number of RR intervals (RRI) with the RR interval exceeding 20 ms is counted.

The number of RR intervals (NN50) in excess of 50 ms is calculated by the number of RR intervals (RRI) in which the RR interval exceeds 50 ms in the 5-minute pulse wave.

That is, the number of RR intervals (RRIs) whose RR intervals exceed 50 ms is counted.

The ratio of the number of RR intervals exceeding 20 ms (pNN20) of the total number of RR intervals is the ratio of the number of RR intervals (NN20) exceeding 20 ms of the total number of RR intervals in a pulse wave of 5 minutes, And is expressed by the following equation.

The ratio of the number of RR intervals exceeding 20 ms (pNN50) of the total number of RR intervals is a ratio of NN50 among the total number of RR intervals in a pulse wave of 5 minutes, in the form of percent (%) of NN50. The following is displayed.

In HRV frequency domain feature extraction, HRV time series data is transformed into frequency domain in order to analyze the HRV in the frequency domain. To do this, first, baseline detrend process is applied to RR 50% overlapped (50% overlapped) overlapping to reduce the direct current (DC) value of the interval and to perform frequency interpolation (Cubic interpolation) (Fast Fourier Transform) is performed for each window by applying a Hamming window to transform the frequency domain, and extract HRV features in the converted frequency domain.

3 is a schematic diagram for explaining the HRV frequency domain characteristic. In Fig. 3, the Y axis represents the frequency band intensity of HRV signal, and the X axis represents time.

The extracted HRV frequency domain characteristic is characterized by a very low frequency (VLF) power spectral density (PSD) of 0.003 to 0.04 Hz, a low frequency band of 0.04 to 0.15 Hz (LF) intensity, a high frequency band (HF) intensity of 0.15 to 0.4 Hz, and a low frequency band high frequency band intensity ratio (LF / HF ratio). HRV characteristics in these frequency regions indicate significant autonomic nervous system activity And is used as a pain classification index.

The intensity (PSD) of the very low frequency (VLF) is the intensity (PSD) in the frequency band of 0.003-0.04 Hz of the HRV signal, i.e. the VLF intensity is 0.003-0.04 Hz at all RR intervals Frequency intensity, which provides additional information on sympathetic nervous system, which can be expressed as follows.

PSD (x) is the power spectral density of the HRV signal, that is, PSD (x) is the frequency at which the frequency is x, and VLF represents the PSD at the VLF, i.e. the very low frequency band Power spectral density.

The intensity (PSD) of the LF (low frequency) is the intensity (PSD) of the low frequency band from 0.04 to 0.15 Hz of the HRV signal, i.e. the LF intensity is 0.04 to 0.15 Hz at all RR intervals Frequency intensity, and simultaneously reflects the activities of the sympathetic nervous system and parasympathetic nervous system.

In the above equation, LF represents the PSD in the LF, that is, the low frequency band (LF) intensity.

The intensity (PSD) of the HF (High frequency) is the PSD of the frequency band 0.15-0.4 Hz of the heart rate variation (HRV) signal, ie the HF intensity is the frequency intensity of 0.15-0.4 Hz at all RR intervals , Which is a respiratory band associated with respiration, that is, a relative high frequency component associated with respiratory activity. The PSD of HF is an index of the activity of the parasympathetic nervous system (vagus nerve), which can be expressed as follows.

In the above equation, HF represents the PSD in HF, that is, the high frequency band (HF) intensity.

The intensity of the low frequency band (LF) and the intensity of the high frequency band (HF) show the degree of control and balance of the two autonomic nervous systems (sympathetic and parasympathetic nerves).

The low frequency band high frequency band intensity ratio (LF / HF ratio) is a ratio between LF and HF, and the low frequency band high frequency band intensity ratio reflects the overall balance between sympathetic and parasympathetic nerves. The frequency ratio (LF / HF ratio) of the low frequency band and the high frequency band is proportional to the activity of the sympathetic nerve and inversely proportional to the activity of the parasympathetic nerve.

In the above equation, Ratio represents a low frequency band high frequency band intensity ratio.

Next, the operation processing unit 200 normalizes 15 characteristics of a pulse-wave signal's time domain characteristic (parameter), a heart rate variability (HRV) time domain characteristic, and a heart rate variability (HRV) frequency domain characteristic Describe the process.

That is, the time domain characteristics (parameter), the heart rate variability (HRV) time domain characteristics, and the heart rate variability (HRV) frequency domain characteristics of the pulse wave signal have different ranges for each feature, Each subject has a deviation. Therefore, a normalization technique is used to have different ranges for each feature, to correct deviations for each subject, and to avoid being influenced by specific feature values in pain classification. All extracted feature vectors, ie, data with 15 characteristics, are normalized to a range of 0 to 1 in the same manner as the following equation.

Here, D _raw (i) represents the value of the original data of each feature, i.e., the time domain characteristics of the pulse wave signal, the HRV time domain characteristics, the HRV frequency domain characteristics Value. (HRV) time domain features and HRV frequency domain features of the pulse wave signal are all 15, and when these 15 features are arranged in a predetermined order, the values of the i-th feature D _raw (i). D _min (i) represents the maximum value among the i-th feature values, that is, the maximum value for each feature, D _min (i) represents the minimum value among the i-th feature values, D _{normalization} (i) is normalized data, that is, a value obtained by normalizing D _raw (i) which is a value of the i-th feature, that is, a value obtained by normalizing each feature. In other words, if the i-th characteristic is a very low frequency (VLF) intensity, values obtained by normalizing D _raw (i), which are the values of the intensity of VLF (Very Low Frequency), are D _{normalization} value used D _max (i) means the maximum value among the values of the intensity of VLF, D _min (i) are the smallest values of values of the intensity of VLF.

In the machine learning pattern analysis through the pattern classifier, the operation processing unit 200 includes a pattern classifier based on a Multi-layer Perceptron Neural Network and a support vector machine using a Radial Basis Function kernel Support Vector Machine (SVM) based pattern classifier and Deep Belief Network (DBN) based pattern classifier. In other words, classification of pain depth classifies the pain depth by analyzing the machine learning pattern using multilayer perceptron neural network, support vector machine, and deep trust neural network using the features extracted from pulse wave and heart beat variability as inputs to the pain pattern classifier.

A pattern classifier based on a Multi-layer Perceptron Neural Network is a basic classification model composed of an input layer, a hidden layer, and an output layer for classifying data belonging to different classes. Classifier learning is performed while modifying the connection weight so that the error between the output value and the target value is minimized, and the pain depth classification is performed. In other words, the multi-layer perceptron neural network pattern classifier puts the features extracted from the pulse wave and heart beat variability into the input layer and puts the pain depth into the output layer to calculate the connection weight of the hidden layer in the direction of reducing the error between the output value and the target value through machine learning. .

The pattern classifier based on the multilayer perceptron neural network is well known and its detailed description is omitted. For example, Korean Patent Laid-Open Publication No. 10-2011-0027916 'Perceptron Artificial Neural Network Weighting Apparatus and Detection Apparatus, Detection System and Detection Method Using It', and Korean Patent Laid-Open No. 10-2001-0047163 ' Learning method of perceptron neural network circuit '.

The support vector machine (SVM) based pattern classifier using the Radial Basis Function kernel is an optimal hyperplane that best classifies data of two different classes, And is well known, and a detailed description thereof will be omitted. For example, the Wikipedia encyclopedia (https://en.wikipedia.org/wiki/Support_vector_machine, https://en.wikipedia.org/wiki/Radial_basis_function_kernel) is available. That is, according to the present invention, the support vector machine-based pattern classifier uses a Radial Basis Function kernel in consideration of a high dimensional input vector, sets the features extracted from the pulse wave and heartbeat variability as input vectors, The optimum value of the normal vector of the hyperplane that maximizes the margin and divides the pattern of pain intensity is obtained. Here, the two groups are two groups according to labeling, namely groups according to pain / absence (that is, painful group and painless group).

The pattern classifier based on Deep Belief Network (DBN) reduces the problem of local optimization through the Restricted Boltzmann Machine (RBM) of Unsupervised Learning, To model the complex nonlinear relationship between the features extracted from the PPG and the pain status based on the Numeric Rate Scale (NRS). Deep trust neural network (DBN), restricted Boltman machine (RBM), etc. are well known and will not be described in detail. For example, Wikipedia Encyclopedia, etc. (https://en.wikipedia.org/wiki/Deep_belief_network, https: //en.wikipedi a.org/wiki/Restricted_Boltzmann_machine) is available. In other words, the pattern classifier based on the deep trust neural network (DBN) uses the features extracted from the pulse wave and heartbeat variability as the input layer to initialize the connection weights and the biases based on the assumption that only the input values are known through the non- ) Learning to classify the patterns of pain intensity by adjusting the backward propagation algorithm to minimize errors of connection weights.

In the present invention, the classification of pain is performed by classifying the pain intensity by analyzing machine learning patterns of Multi-layer Perceptron Neural Network, Support Vector Machine (SVM), and Deep Belief Network from extracted features.

4 is a flowchart schematically illustrating a method of performing pain classification using pulse waves and heart beat variability in the arithmetic processing unit 200 of FIG.

In the time domain feature detection step S100 of the pulse wave signal, the calculation processing unit 200 extracts a pulse wave signal received from the signal preprocessing unit 120 or received by the signal preprocessing unit 120 and temporarily stored in the memory unit 270, A pulse wave signal average amplitude, a pulse wave rise time, a fall time, and an average heart rate, which are time-domain characteristics of the pulse wave signal, are obtained.

After the time domain feature detection step S100 of the pulse wave signal is detected in the RR interval detection step S150, the operation processing unit 200 receives the signal pre-processing unit 120 or the signal preprocessing unit 120, That is, the heartbeat variation, which is the interval from the R point (i.e., the maximum point) of each cycle to the R point of the next cycle in succession from the pulse wave signal temporarily stored in the memory 270.

In the heart rate variability time domain feature extraction step S200, the calculation processing unit 200 calculates an average RR interval (Mean RRI), a standard deviation (SDNN) of the total RR interval, a heart rate variability (HRV) (RMSSD), the number of RR intervals in excess of 20 ms (NN20), the number of RR intervals in excess of 50 ms (NN50), and the total number of RR intervals in excess of 20 ms (PNN20) of the number of RR intervals, and the ratio (pNN50) of the number of RR intervals exceeding 50 ms among the total number of RR intervals.

In the baseline correction and cubic interpolation step S250, the arithmetic processing unit 200 performs a baseline detrend process on the RR interval values and the heartbeat deviation values obtained in the RR interval detection step S150, And the cubic interpolation (cubic interpolation) is performed at 2 Hz to increase the frequency resolution.

In the FFT step S300, the arithmetic processing unit 200 performs a baseline correction and a 50% overlapped Hamming window 50% overlapping the RR interval value subjected to the baseline correction and cubic interpolation in the cubic interpolation step S250 And performs Fast Fourier Transform (FFT) on each of the wide-angle wands, thereby converting them into the frequency domain.

HRV frequency domain feature extraction step S350 is a step of extracting HRV frequency domain characteristics from 0.003 to 0.04 Hz in the RR interval (heart beat variation) values converted into the frequency domain in the FFT step S300 Very low frequency (VLF) power spectral density (PSD), low frequency band (LF) intensity of 0.04 to 0.15 Hz, high frequency band (HF) intensity of 0.15 to 0.4 Hz, band strength of low frequency band high frequency (LF / HF ratio).

Here, the time domain feature detection step S100 to the HRV frequency domain feature extraction step S350 of the pulse wave signal may be referred to as a signal detection step.

In the normalization step S400, the arithmetic processing unit 200 calculates the time-domain characteristic of the pulse wave signal obtained in the time-domain feature detection step S100 of the pulse wave signal, (HRV) time domain characteristic, HRV frequency domain characteristic (HRV) frequency domain characteristic extracted from the frequency domain feature extraction step (S350), and the values of the 15 characteristic values of the frequency domain characteristic .

 In the labeling step S450, the arithmetic processing unit 200 performs labeling of the features (feature vectors) normalized in the normalization step S400 by the pain intensity group-specific output variable.

In the pain intensity classification step S500, the operation processing unit 200 may classify the normalized feature vectors labeled in the labeling step S450 into a pattern classifier based on a Multi-layer Perceptron Neural Network, A pattern classifier based on a support vector machine (SVM) based on a radial basis function kernel and a pattern classifier (pattern classification algorithm) based on a deep trust network (DBN) Pattern analysis is performed to classify pain intensity. And outputs the classified pain intensity to the output unit 270.

It is to be understood that the present invention is not limited by what has been described above and illustrated in the drawings, and those skilled in the art will recognize that many modifications and variations are possible within the scope of the following claims.

100: a pulse-wave detecting unit 120: a signal preprocessing unit
200: operation processing unit 250:
270:

Claims (15)

The arithmetic processing section detects a pulse-wave time-domain characteristic including an average amplitude of the pulse wave from the pulse wave signal received from the pulse-wave sensor, detects a heartbeat variation in the RR interval, and calculates a heartbeat- A signal detection step of detecting a heart rate variability frequency domain characteristic including a VLF intensity which is a frequency intensity of 0.003 to 0.04 Hz of a heartbeat variation;
A normalizing step of normalizing the pulse-wave time-domain characteristic, the heartbeat-variability time-domain characteristic, and the heartbeat-variability frequency-domain characteristic detected in the signal detection step so as to have a value between 0 and 1;
A labeling step of labeling the pulse wave time domain characteristic, the heartbeat mutation time domain characteristic, and the heartbeat mutation frequency domain characteristic normalized in the normalization step as an output variable for each pain intensity group in the normalization step;
The operation processing unit classifies the pain depth by inputting a feature vector having an output variable for each pain depth group output in the labeling step into a pattern classifier and performing a machine learning pattern analysis;
Wherein the method comprises classifying the at least one of the at least one heartbeat and the heartbeat. 2. The method of claim 1,
The arithmetic processing unit is configured to calculate a pulse wave signal average amplitude, a pulse wave rise time, a pulse wave signal, and a pulse wave signal, which are time-domain characteristics of the pulse wave signal received from the pulse wave sensor or received from the pulse wave sensor and temporarily stored in the memory unit, A time domain feature detection step of obtaining a fall time and an average heart rate of a pulse wave signal;
After the detection of the time-domain characteristic of the pulse wave signal, the calculation processing section calculates the R-point of the next cycle from the R-point (maximum point) of each cycle in the pulse wave signal received from the pulse wave sensor or received from the pulse- An RR interval detecting step of obtaining an RR interval (heart beat variation)
After the RR interval detection step, the arithmetic processing unit calculates the average square root of the mean of the values obtained by squaring the difference between the adjacent RR intervals, the mean RR interval (Mean RRI), the standard deviation (SDNN) of the total RR interval, the number of RR intervals exceeding 20 ms (NN20), the number of RR intervals exceeding 50 ms (NN50), the ratio of the total number of RR intervals exceeding 20 ms (pNN20), the total number of RR intervals Calculating a ratio (pNN50) of the number of RR intervals exceeding 50 ms in the time domain;
Wherein the method comprises classifying the at least one of the at least one heartbeat and the heartbeat. 3. The method of claim 2,
The arithmetic processing unit performs a linear interpolation process for removing the DC value of the RR interval by baseline correction from the RR interval values and the heartbeat deviation values obtained in the RR interval detection step and performing a Cubic interpolation (cubic interpolation) at 2 Hz. Correction and cubic interpolation steps;
The arithmetic processing unit applies Fast Fourier Transform (FFT) by applying a 50% overlapped Hamming window to the RR interval value subjected to baseline correction and cubic interpolation in the baseline correction and cubic interpolation steps, The FFT step;
A very low frequency band (VLF) intensity of 0.003 to 0.04 Hz, a low frequency band (LF) intensity of 0.04 to 0.15 Hz, a frequency of 0.15 (HRV) frequency domain feature extraction step of obtaining a high frequency band (HF) intensity of ~0.4 Hz and a band intensity ratio (LF / HF ratio) of a low frequency band high frequency;
Further comprising the steps of: detecting the heartbeat and heartbeat. 3. The method of claim 2,
The arithmetic processing unit obtains the amplitude between the maximum point Peak and the minimum point in one cycle of the pulse wave signal by the amplitude of the pulse wave signal and obtains the average of the pulse wave signal amplitudes as the pulse wave signal average amplitude for five minutes,
The pulse wave rise time is an average of the time taken to rise from the minimum point (valley) to the maximum point (peak) in each cycle of the pulse wave signal for 5 minutes,
Wherein the pulse wave falling time is an average of a time taken to fall from a peak point to a minimum point in each cycle for 5 minutes. The method of claim 3,
The arithmetic processing unit calculates the time domain characteristics of the pulse wave signal, the heart rate variability (HRV) time domain features, and the HRV frequency domain features (D _{normalization} (i))

(Note that, D _raw (i) is the time domain of the pulse wave signal characteristics, heart rate variability (HRV) time domain features, and heart rate variability (HRV) is the one of the frequency-domain features, one feature value of, D _max (i ) _{Represents the} maximum value among the characteristics representing D _raw (i), and D _min (i) represents the minimum value among the characteristics representing D _raw (i)).
And the heartbeat and heartbeat variability. The method of claim 3,
In the pain intensity classification step, the normalized feature vector, which is labeled in the operation unit labeling step, is divided into a pattern classifier based on a Multi-layer Perceptron Neural Network and a Radial Basis Function kernel A pattern classifier based on a support vector machine (SVM) and a pattern classifier based on a Deep Belief Network (DBN) to classify the pain intensity by performing machine learning pattern analysis. Methods of pain classification using pulse and heart rate variability. A recording medium storing a computer program source for a method of classifying a pain using pulse waves and heart beat variability according to any one of claims 1 to 6. A pulse wave sensor that includes a light emitting diode and a photosensor and detects a pulse wave that is an optical pulse wave,
A signal preprocessing unit for amplifying the pulse wave received from the pulse wave sensor, removing noise, and converting it into a digital signal;
And an arithmetic processing unit for receiving a pulse wave from the signal preprocessing unit and detecting a pulse wave time-domain characteristic including an average amplitude of the pulse wave, the arithmetic processing unit comprising:
The heart rate variability time domain characteristic including the average RR interval is detected from the heart beat variation and the VLF of 0.003 to 0.04 Hz of the heart beat variation is detected from the pulse wave signal received from the signal preprocessing section, Detecting a heart rate variability frequency domain characteristic including intensity,
Normalizing the detected pulse-wave time-domain characteristics, heartbeat-mutation time-domain characteristics, heartbeat-mutation frequency-domain characteristics to have values between 0 and 1,
The normalized pulse wave time domain characteristic, the heartbeat mutation time domain characteristic, and the heartbeat mutation frequency domain characteristic are labeled with the output variable for each pain intensity group,
And a feature vector having an output variable for each of the pain intensity groups is input to a pattern classifier to perform a machine learning pattern analysis. 9. The method of claim 8,
The arithmetic processing unit obtains the mean amplitude of the pulse wave signal, the Rise time, the pulse wave fall time, and the average heart rate as the time domain characteristics of the pulse wave signal,
The arithmetic processing section calculates an RR interval (heartbeat variation), which is an interval from the R point (maximum point) of each cycle to the R point of the next cycle in the pulse wave signal received from the pulse wave sensor or received from the pulse wave sensor and temporarily stored in the memory section, ≪ / RTI >
The arithmetic processing unit calculates the average square root of the mean square value (rMSSD) of the value obtained by squaring the difference between the adjacent RR intervals, the standard deviation (SDNN) of the total RR interval, the mean RR interval (mean RRI) The number of RR intervals (NN20) exceeding 50 ms, the number of RR intervals (NN50) exceeding 50 ms, the ratio of the number of RR intervals exceeding 20 ms among the total number of RR intervals (pNN20), the number of RR And a ratio (pNN50) of the number of intervals is obtained. 10. The image processing apparatus according to claim 9,
(DC) value of the RR interval is removed from the RR interval values and the heartbeat deviation values by baseline correction, Cubic interpolation is performed at 2 Hz,
Fast Fourier transform (FFT) is performed by applying a 50% overlapped Hamming window to the RR interval value subjected to baseline correction and cubic interpolation in the baseline correction and cubic interpolation steps
A very low frequency band (VLF) intensity of 0.003 to 0.04 Hz, a low frequency band (LF) intensity of 0.04 to 0.15 Hz, a frequency of 0.15 to 0.4 Hz (LF / HF ratio) of the high frequency band (HF) intensity of the low frequency band and the high frequency band (HF) intensity of the low frequency band high frequency of the pulse wave. The method according to claim 6,
Wherein the output variable for each pain intensity group is represented by a pain score of a numeric grade scale (NRS) in a labeling step. 12. The method of claim 11,
Wherein the output variable for each of the pain intensity groups represents one of a '0' indicating a rest state without pain and a '1' indicating a state having a pain. 12. The method of claim 11,
The multi-layer perceptron neural network-based pattern classifier places the labeled normalized feature vector as an input layer, sets the pain depth as an output layer, and uses the machine learning to reduce the connection weight of the hidden layer in the direction of reducing the error between the output value and the target value And calculating a pattern of pain intensity according to the heart rate variability. 12. The method of claim 11,
A support vector machine based pattern classifier using a radial basis kernel is a set of two groups of painless and painful groups of labeled normalized feature vectors as input vectors and a margin And calculating an optimal value of a normal vector of a hyperplane that can be maximized and divided to classify a pattern of pain intensity. 12. The method of claim 11,
A deep class trust neural network (DBN)
Using the normalized feature vectors subjected to labeling, the connection weight and bias are initialized with the assumption that only the input value is known through the non-mapping learning using the input layer, and then the connection weighting and the bias are initialized by the supervised learning, Wherein the pattern of the pain intensity is classified by classifying the pattern of the pain intensity into the direction of minimizing the error of the heart.

Images