WO2024152392A1 - 混合机械通气模式下人机不同步波形识别方法及相关设备 - Google Patents

混合机械通气模式下人机不同步波形识别方法及相关设备 Download PDF

Info

Publication number
WO2024152392A1
WO2024152392A1 PCT/CN2023/075074 CN2023075074W WO2024152392A1 WO 2024152392 A1 WO2024152392 A1 WO 2024152392A1 CN 2023075074 W CN2023075074 W CN 2023075074W WO 2024152392 A1 WO2024152392 A1 WO 2024152392A1
Authority
WO
WIPO (PCT)
Prior art keywords
mechanical ventilation
phase space
waveform
ventilation mode
human
Prior art date
Application number
PCT/CN2023/075074
Other languages
English (en)
French (fr)
Inventor
颜延
马良
熊富海
刘语诗
闫旭东
黄意春
王磊
Original Assignee
中国科学院深圳先进技术研究院
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 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Publication of WO2024152392A1 publication Critical patent/WO2024152392A1/zh

Links

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/08—Detecting, measuring or recording devices for evaluating the respiratory organs
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61M—DEVICES FOR INTRODUCING MEDIA INTO, OR ONTO, THE BODY; DEVICES FOR TRANSDUCING BODY MEDIA OR FOR TAKING MEDIA FROM THE BODY; DEVICES FOR PRODUCING OR ENDING SLEEP OR STUPOR
    • A61M16/00—Devices for influencing the respiratory system of patients by gas treatment, e.g. mouth-to-mouth respiration; Tracheal tubes
    • 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
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2123/00—Data types
    • G06F2123/02—Data types in the time domain, e.g. time-series data
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02B—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO BUILDINGS, e.g. HOUSING, HOUSE APPLIANCES OR RELATED END-USER APPLICATIONS
    • Y02B30/00—Energy efficient heating, ventilation or air conditioning [HVAC]
    • Y02B30/70—Efficient control or regulation technologies, e.g. for control of refrigerant flow, motor or heating

Definitions

  • the present invention relates to the field of electrophysiological detection and monitoring technology, and in particular to a method, system, terminal and computer-readable storage medium for identifying human-machine asynchronous waveforms in a hybrid mechanical ventilation mode.
  • PVA Patient-Ventilator Asynchrony
  • Trigger type invalid trigger, repeated trigger, automatic trigger and reverse trigger.
  • VM ventilation mode
  • Traditional ventilation modes are mainly divided into two categories, (i) constant pressure ventilation mode (including pressure control mode (Pressure Controlled Ventilation, PCV) and pressure support mode (Pressure Support Ventilation, PSV)), and (ii) constant volume ventilation mode (including volume controlled ventilation (Volume Controlled Ventilation, VCV) and volume controlled assisted ventilation (Volume Control Assisted Ventilation, VC-AV)), as shown in Figure 1,
  • (a) and (d) in Figure 1 are normal ventilation cycle waveforms
  • (b) and (e) in Figure 1 are ineffective inspiratory effort cycle waveforms, as shown by the arrows
  • (c) and (f) in Figure 1 are early switching cycle waveforms, as shown by the arrows
  • (a)-(c) in Figure 1 are waveforms in constant volume ventilation mode
  • (d)-(f) in Figure 1 are waveforms in constant pressure ventilation mode.
  • constant pressure ventilation mode is that the peak pressure is lower and barotrauma occurs less frequently; however, the tidal volume is greatly affected by system compliance and viscous tissue; on the contrary, constant volume ventilation mode can ensure the supply of tidal volume, which is conducive to the rest of the respiratory muscles, but it is easy to cause barotrauma.
  • Both ventilation modes have a certain probability of causing asynchrony in human-machine interaction, and there is no conclusion on which mode is better so far.
  • PC-SIMV Pressure Controlled-Synchronized Intermittent Mandatory Ventilation
  • Variable Pressure Support Ventilation Variable Pressure Support Ventilation
  • Volume Controlled Synchronized Intermittent Mandatory Ventilation One feature of the waveforms generated by these modes is that they are more like a mixture of constant pressure and constant volume ventilation modes from a morphological point of view. Therefore, exploring the algorithm model for identifying and classifying PVA classification in mixed ventilation modes is of great significance to the intelligentization and physiological closed-loop control of ventilators.
  • the current PVA classification process can be summarized as follows: (1) obtaining respiratory mechanics waveform data derived from the ventilator; (2) having professional physicians annotate the waveforms of human-machine asynchrony; (3) data preprocessing and dividing the data set into training set, validation set, and test set; (4) inputting the preprocessed data into a defined model for model training; and (5) saving the trained model for application.
  • the prior art mentions the use of a recurrent neural network algorithm to detect human-machine asynchrony.
  • the GRU gated recurrent unit
  • the GRU extracts the pressure waveform features and the flow velocity time waveform features respectively, and then uses the BGRU (bidirectional gated recurrent neural unit) to extract higher-dimensional features after fusing the two features.
  • the softmax fully connected layer is used to obtain the classification results of the human-machine asynchrony type.
  • the data set is annotated by professional doctors, and four major types of human-machine asynchrony are detected: flow velocity, trigger, cycle and others.
  • a visual explanation of the classification decision of the model can be obtained by gradient weighted class activation mapping.
  • converting the collected original respiratory signal into a two-dimensional image first use the public image dataset ImageNet to train a two-dimensional image multi-classification model, and then use the transfer learning method to input the two-dimensional image composed of the respiratory waveform into the model and fine-tune the weights of the layers above the last fully connected layer to obtain a convolutional neural network that can be used for respiratory waveform classification.
  • the respiratory waveform data is read in real time to form a test sequence.
  • the DTW distance between the test sequence and all the sequences in the training set is calculated, and then the similarity distance is calculated using DTW.
  • the test sequence is classified; it is used to judge invalid inspiratory efforts in the phenomenon of human-machine asynchrony.
  • fuzzy entropy is used to analyze EEG signals, and then the fuzzy entropy under the corresponding electrode that reflects the characteristics of the EEG signals is extracted as input features through feature selection, and finally the features are used for classification.
  • the invention solves a binary classification problem.
  • the original respiratory waveform is first transformed once using wavelet scaling transform, and on this basis, a variety of entropy features are used to extract nonlinear features. After the best feature combination is selected using the preceding selection algorithm, it is used as the input of the support vector machine classification algorithm for classification. It still only classifies invalid inspiratory efforts as a type of human-machine asynchrony phenomenon, which belongs to a binary classification task.
  • the main purpose of the present invention is to provide a method, system, terminal and computer-readable storage medium for identifying human-machine asynchronous waveforms in a hybrid mechanical ventilation mode, aiming to solve the problem that various recognition algorithms in the prior art use ventilation data collected in a single ventilation mode when performing classification tasks and are unable to effectively identify human-machine asynchronous phenomena.
  • the present invention provides a method for identifying human-machine asynchronous waveforms in a hybrid mechanical ventilation mode, wherein the method for identifying human-machine asynchronous waveforms in a hybrid mechanical ventilation mode comprises the following steps:
  • the mechanical ventilation waveform is subjected to data segmentation, data labeling and data transformation to obtain a respiratory cycle waveform
  • a phase space reconstruction-convolutional neural network model is constructed, and based on the phase space reconstruction-convolutional neural network model, the human-machine asynchrony phenomenon under the mixed constant pressure and constant volume ventilation modes is identified according to the input respiratory cycle waveform.
  • the collected mechanical ventilation waveform is segmented according to the respiratory cycle, and features are extracted for each respiratory cycle, or the data of one respiratory cycle is regarded as a sample and input into the deep model.
  • ⁇ is the standard deviation of X
  • xi represents the i-th data.
  • phase space reconstruction-convolutional neural network model is specifically:
  • the delay vector after applying the delay embedding technology to X is obtained:
  • X i ⁇ x 1 ,x 2 ,...,x i+(m-1) ⁇ ⁇ ,i ⁇ [1,L];
  • x Lm refers to the value of the mth dimension of the Lth trajectory point.
  • the average mutual information I(X(t),X(t+ ⁇ )) of the one-dimensional time series X(t) and the delayed version X(t+ ⁇ ) is used as an indicator to quantify the independence of the two sequences, as shown in the following formula.
  • the optimal delay parameter ⁇ is determined based on the first minimum point of the mutual information:
  • p ij ( ⁇ ) refers to the joint distribution probability of the corresponding information in X(t) and X(t+ ⁇ ), and p i and p j refer to the probability of the current information;
  • the size after zero addition is 3*N X ;
  • the input data size after phase space reconstruction is 3*m*N psr , where N psr refers to the number of points after phase space reconstruction;
  • the phase space obtained after the respiratory cycle waveform is reconstructed is input into the convolution structure; the convolutional neural network has two convolutional layers for feature extraction.
  • the first convolutional layer includes a two-dimensional convolution, nonlinear transformation and maximum pooling; the second convolutional layer includes one-dimensional convolution and nonlinear transformation; the activation functions used in the two convolutional layers are both rectified linear units; after the convolutional layer extracts the features, the classification label is output through a fully connected layer.
  • the method for identifying human-machine asynchronous waveforms in the hybrid mechanical ventilation mode wherein after the original respiratory waveform is reconstructed in phase space, the waveform dimension obtained is 3*m*N psr , and when input into the convolutional neural network, it includes:
  • One-dimensional phase space convolution perform one-dimensional convolution on the trajectory points in the m-dimensional phase space in order according to the dimension, with a convolution kernel size of 1*K1 and a step size of S1, where K1 represents the size of the convolution kernel of the first convolution layer; merge the features obtained after convolution into one dimension, and perform another one-dimensional convolution with a convolution kernel size of 1*K2 and a step size of S2, where K2 represents the size of the convolution kernel of the second convolution layer; input the obtained features into the fully connected layer and then pass through the softmax function to obtain the classification probability;
  • Two-dimensional phase space convolution perform two-dimensional convolution on trajectory points, with a convolution kernel size of m*K1 and a step size of S2; merge the convolution features into one dimension, and perform another one-dimensional convolution with a convolution kernel size of 1*K2 and a step size of S2; input the obtained features into the fully connected layer and then pass the softmax function to obtain the classification probability.
  • the present invention also provides a human-machine asynchronous waveform recognition system in a hybrid mechanical ventilation mode, wherein the human-machine asynchronous waveform recognition system in a hybrid mechanical ventilation mode comprises:
  • a data acquisition module used to collect mechanical ventilation waveforms containing human-machine asynchrony in constant pressure ventilation mode and constant volume ventilation mode, wherein the mechanical ventilation waveforms select three channels: airway pressure, flow and tidal volume;
  • a data processing module used for performing data segmentation, data labeling and data transformation on the mechanical ventilation waveform to obtain a respiratory cycle waveform
  • a data recognition module is used to construct a phase space reconstruction-convolutional neural network model, and based on the phase space reconstruction-convolutional neural network model, identify the human-machine asynchrony phenomenon under the mixture of constant pressure and constant volume ventilation modes according to the input respiratory cycle waveform.
  • the present invention also provides a terminal, wherein the terminal comprises: a memory, a processor, and a hybrid mechanical ventilation mode program stored in the memory and executable on the processor.
  • a human-machine asynchronous waveform recognition program when executed by the processor, implements the steps of the human-machine asynchronous waveform recognition method under the hybrid mechanical ventilation mode as described above.
  • the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a human-machine asynchronous waveform recognition program under a hybrid mechanical ventilation mode, and when the human-machine asynchronous waveform recognition program under the hybrid mechanical ventilation mode is executed by a processor, the steps of the human-machine asynchronous waveform recognition method under the hybrid mechanical ventilation mode as described above are implemented.
  • mechanical ventilation waveforms containing human-machine asynchrony in constant pressure ventilation mode and constant volume ventilation mode are collected, and the mechanical ventilation waveform selects three channels of airway pressure, flow and tidal volume; the mechanical ventilation waveform is subjected to data segmentation, data labeling and data transformation to obtain a respiratory cycle waveform; a phase space reconstruction-convolutional neural network model is constructed, and based on the phase space reconstruction-convolutional neural network model, the human-machine asynchrony phenomenon under the mixed constant pressure and constant volume ventilation modes is identified according to the input respiratory cycle waveform.
  • phase space reconstruction-convolutional neural network model constructed by the present invention has good generalization performance, can accurately identify the human-machine asynchrony phenomenon under the mixed constant pressure and constant volume ventilation modes, and is of great significance to the intelligentization and physiological closed-loop control of the ventilator.
  • FIG1 is a schematic diagram of a mechanical ventilation waveform
  • FIG2 is a flow chart of a preferred embodiment of a method for identifying human-machine asynchronous waveforms in a hybrid mechanical ventilation mode of the present invention
  • FIG. 3 is a schematic diagram of time-delay embedding in phase space in a preferred embodiment of the method for identifying human-machine asynchronous waveforms in a hybrid mechanical ventilation mode of the present invention
  • phase space reconstruction-convolutional neural network structure in a preferred embodiment of the method for identifying human-machine asynchronous waveforms in a hybrid mechanical ventilation mode of the present invention
  • FIG5 is a schematic diagram showing the principle of a preferred embodiment of a human-machine asynchronous waveform recognition system under a hybrid mechanical ventilation mode of the present invention
  • FIG. 6 is a schematic diagram of an operating environment of a preferred embodiment of a terminal of the present invention.
  • PC-SIMV Pressure Controlled-Synchronized Intermittent Mandatory Ventilation
  • Variable Pressure Support Ventilation Variable Pressure Support Ventilation
  • Volume Controlled Synchronized Intermittent Mandatory Ventilation One feature of the waveforms generated by these modes is that they are more like a mixture of constant pressure and constant volume ventilation modes from a morphological point of view. Therefore, exploring the algorithm model for identifying and classifying PVA classification in mixed ventilation modes is of great significance to the intelligentization and physiological closed-loop control of ventilators.
  • the method for identifying human-machine asynchronous waveforms in a hybrid mechanical ventilation mode is shown in FIG1 , and the method for identifying human-machine asynchronous waveforms in a hybrid mechanical ventilation mode comprises the following steps:
  • Step S10 collecting mechanical ventilation waveforms containing human-machine asynchrony in constant pressure ventilation mode and constant volume ventilation mode, wherein the mechanical ventilation waveforms select three channels of airway pressure, flow and tidal volume.
  • Step S20 performing data segmentation, data labeling and data transformation on the mechanical ventilation waveform to obtain a respiratory cycle waveform.
  • data segmentation refers to segmenting the collected mechanical ventilation waveform (raw data) according to the respiratory cycle (the length of the respiratory cycle usually varies according to the patient's condition, generally 3-5 seconds, but in the waveform, due to different sampling frequencies, the number of points in a respiratory cycle is also different, for example, the sampling frequency is 50HZ, and a respiratory cycle is 3 seconds, then the number of points in this cycle is 150 points), and finally extracting features for each respiratory cycle or treating the data of a respiratory cycle as samples input into the deep model.
  • the respiratory cycle the length of the respiratory cycle usually varies according to the patient's condition, generally 3-5 seconds, but in the waveform, due to different sampling frequencies, the number of points in a respiratory cycle is also different, for example, the sampling frequency is 50HZ, and a respiratory cycle is 3 seconds, then the number of points in this cycle is 150 points
  • data labeling label the mechanical ventilation waveform of each respiratory cycle, where the labels include invalid triggering, repeated triggering, automatic triggering, reverse triggering, flow rate mismatch, early switching, delayed switching, intrinsic positive end-expiratory pressure and normal.
  • data transformation the mechanical ventilation waveforms of different channels (dimensions) are transformed (because the collected data come from different channels, the units of these channel data are different, for example, the unit of airway pressure is cm/H2O, and the unit of flow rate is L/min) to ensure the convergence and convergence speed of the deep model.
  • the data transformation uses z-transformation:
  • ⁇ is the standard deviation of X
  • xi represents the i-th data.
  • Step S30 constructing a phase space reconstruction-convolutional neural network model, and identifying human-machine asynchrony phenomenon under a mixture of constant pressure and constant volume ventilation modes based on the input respiratory cycle waveform based on the phase space reconstruction-convolutional neural network model.
  • the delay vector after applying the delay embedding technique to X can be obtained:
  • X i ⁇ x 1 ,x 2 ,...,x i+(m-1) ⁇ ⁇ ,i ⁇ [1,L];
  • Xi is a point in the m-dimensional phase space, which represents the delay vector
  • L N-(m-1) ⁇ is the number of trajectory points in the phase space.
  • phase space trajectory matrix (the result of phase space reconstruction, which is another representation of the original one-dimensional time series.
  • the convolutional neural network model can extract information from the reconstructed phase space of the original system, which cannot be extracted from the original one-dimensional time series and is conducive to improving the classification results):
  • x Lm refers to the value of the mth dimension of the Lth trajectory point.
  • Each row of the phase space trajectory matrix is a trajectory point in the phase space. This point has m dimensions and there are L rows in total, that is, L trajectory points.
  • FIG2 shows a schematic diagram of the time-delay embedded phase space, where the embedding dimension m is set to 3, the time-delay parameter ⁇ is set to 1, and finally a 3-dimensional phase space of three channels is obtained.
  • the delay parameter ⁇ is determined according to the average mutual information:
  • the average mutual information I(X(t),X(t+ ⁇ )) of the one-dimensional time series X(t) (i.e., the one-dimensional time series X) and the delayed version X(t+ ⁇ ) is used as an indicator to quantify the independence of the two sequences, as shown in the following formula.
  • the optimal delay parameter ⁇ is determined based on the first minimum point of the mutual information:
  • t is the time variable
  • p ij ( ⁇ ) refers to the joint distribution probability of the corresponding information in X(t) and X(t+ ⁇ )
  • p i and p j refer to the probability of the current information.
  • the embedding dimension m is determined based on the pseudo nearest neighbor: if the distance between points in a one-dimensional time series is less than the first preset distance, it is called a nearest neighbor (that is, if the distance between the points is very close, it is called a nearest neighbor). If the time series is embedded in m dimensions under a certain time delay, the distance between the points in the m-dimensional phase space is calculated. If the change in distance is greater than the second preset distance (that is, the distance changes greatly), it is called a pseudo nearest neighbor. Continue to change the embedding dimension m until the distance change is less than the third preset distance (that is, the distance change is no longer drastic), then m at this time is regarded as an estimate of the embedding dimension.
  • phase space reconstruction-convolutional neural network model is constructed:
  • the data is padded with 0, that is, a sufficient number of 0s are padded at the end of each channel to make the data size input to the convolutional neural network the same.
  • the size after 0-padded is 3*N X ;
  • the input data size after phase space reconstruction is 3*m*N psr ,
  • N psr refers to the number of points after phase space reconstruction, and 3 refers to 3 channels.
  • the convolutional neural network has two convolutional layers for feature extraction.
  • the first convolutional layer includes a two-dimensional convolution, nonlinear transformation and maximum pooling;
  • the second convolutional layer includes one-dimensional convolution and nonlinear transformation;
  • the activation functions used in the two convolutional layers are both rectified linear units (ReLU); after the convolutional layer extracts the features, it outputs the classification label through a fully connected layer.
  • ReLU rectified linear units
  • the waveform dimension is 3*m*N psr .
  • the convolutional neural network There are two different ways to input it into the convolutional neural network:
  • One-dimensional phase space convolution The trajectory points in the m-dimensional phase space are convolved one-dimensionally in sequence according to the dimension.
  • the convolution kernel size is 1*K1 and the step size is S1.
  • K1 represents the size of the convolution kernel of the first convolution layer.
  • the features obtained are merged into one dimension, and then a one-dimensional convolution is performed.
  • the convolution kernel size is 1*K2, the step size is S2, and K2 is the size of the convolution kernel of the second convolution layer.
  • the obtained features are input into the fully connected layer and then passed through the softmax function to obtain the classification probability.
  • Two-dimensional phase space convolution Perform two-dimensional convolution on the trajectory points, with a convolution kernel size of m*K1 and a step size of S2. Merge the convolution features into one dimension and perform another one-dimensional convolution with a convolution kernel size of 1*K2 and a step size of S2. Input the features into the fully connected layer and pass them through the softmax function to obtain the classification probability.
  • Model training and evaluation The training and evaluation process of the model is to find a model that can achieve the best process in various indicators with suitable parameters. In this process, it is necessary to set a leave-one-out cross-validation experiment to test the generalization performance of the model, that is, the data set of each case is used as a test set in turn, and the data of the remaining cases are used as a training set.
  • the training process usually inputs the data into the proposed model of the present invention in turn, sets the training [number of rounds] parameter, and then runs the computer algorithm. In each round, the program will output the current accuracy and F1-score indicators to judge the performance of the current model.
  • the leave-one-out cross-validation experiment mentioned above is a common model selection method.
  • Generalization performance refers to the recognition ability of a model when it encounters unseen data.
  • Accuracy and F1-score are used as evaluation indicators to test the generalization performance of the model. The larger the value of each indicator, the better the generalization of the model.
  • Model deployment The trained model can be deployed in ventilators or servers through various programming languages to provide services for various applications.
  • the data in this experiment came from 7 patients who were mechanically ventilated clinically, all of whom suffered from different respiratory diseases.
  • the mechanical ventilation modes used by the 7 patients also came from constant pressure control ventilation mode, volume control ventilation mode, and volume control-synchronous intermittent command ventilation mode. After data preprocessing, 16,760 respiratory cycles were finally obtained.
  • PSR-CNN refers to the model proposed in the present invention
  • CNN refers to a convolutional neural network model that has the same network structure as PSR-CNN but does not perform phase space reconstruction
  • time domain-RF refers to a random forest model based on time domain features
  • frequency domain-RF refers to a random forest model based on frequency domain features.
  • the three channels can be increased to multiple (more than three) channels, and the change of the data value simultaneously changes the phase space reconstruction process.
  • the structure of the convolution layer it is not limited to using only convolution, and the nonlinear transformation function can also be changed and the maximum pooling layer can be increased or decreased.
  • the present invention also provides a human-machine asynchronous waveform recognition system in the hybrid mechanical ventilation mode, wherein the human-machine asynchronous waveform recognition system in the hybrid mechanical ventilation mode includes:
  • a data acquisition module 51 is used to collect mechanical ventilation waveforms containing human-machine asynchrony in constant pressure ventilation mode and constant volume ventilation mode, wherein the mechanical ventilation waveforms select three channels of airway pressure, flow and tidal volume;
  • a data processing module 52 used for performing data segmentation, data labeling and data transformation on the mechanical ventilation waveform to obtain a respiratory cycle waveform
  • the data recognition module 53 is used to construct a phase space reconstruction-convolutional neural network model, and based on the phase space reconstruction-convolutional neural network model, identify the human-machine asynchrony phenomenon under the mixed constant pressure and constant volume ventilation modes according to the input respiratory cycle waveform.
  • the present invention also provides a terminal, which includes a processor 10, a memory 20, and a display 30.
  • FIG6 only shows some components of the terminal, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.
  • the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard disk, a smart media card (SMC), a secure digital (Secure Digital, SD) card, flash card (Flash Card), etc. Furthermore, the memory 20 may also include both the internal storage unit of the terminal and the external storage device.
  • the memory 20 is used to store the application software and various types of data installed on the terminal, such as the program code of the installation terminal, etc.
  • the memory 20 can also be used to temporarily store data that has been output or is to be output.
  • the memory 20 stores a human-machine asynchronous waveform recognition program 40 in a hybrid mechanical ventilation mode, and the human-machine asynchronous waveform recognition program 40 in the hybrid mechanical ventilation mode can be executed by the processor 10, thereby realizing the human-machine asynchronous waveform recognition method in the hybrid mechanical ventilation mode in the present application.
  • the processor 10 may be a central processing unit (CPU), a microprocessor or other data processing chip, which is used to run the program code or process data stored in the memory 20, such as executing the method for identifying human-machine asynchronous waveforms in the hybrid mechanical ventilation mode.
  • CPU central processing unit
  • microprocessor or other data processing chip, which is used to run the program code or process data stored in the memory 20, such as executing the method for identifying human-machine asynchronous waveforms in the hybrid mechanical ventilation mode.
  • the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc.
  • the display 30 is used to display information on the terminal and to display a visual user interface.
  • the components 10-30 of the terminal communicate with each other via a system bus.
  • the processor 10 executes the human-machine asynchronous waveform recognition program 40 in the memory 20 in the hybrid mechanical ventilation mode, the following steps are implemented:
  • the mechanical ventilation waveform is subjected to data segmentation, data labeling and data transformation to obtain a respiratory cycle waveform
  • a phase space reconstruction-convolutional neural network model is constructed, and based on the phase space reconstruction-convolutional neural network model, the human-machine asynchrony phenomenon under the mixed constant pressure and constant volume ventilation modes is identified according to the input respiratory cycle waveform.
  • the data segmentation is specifically as follows:
  • the collected mechanical ventilation waveform is segmented according to the respiratory cycle, and features are extracted for each respiratory cycle, or the data of one respiratory cycle is regarded as a sample and input into the deep model.
  • the data transformation is specifically as follows:
  • ⁇ is the standard deviation of X
  • xi represents the i-th data.
  • phase space reconstruction-convolutional neural network model is specifically constructed as follows:
  • the delay vector after applying the delay embedding technology to X is obtained:
  • X i ⁇ x 1 ,x 2 ,...,x i+(m-1) ⁇ ⁇ ,i ⁇ [1,L];
  • x Lm refers to the value of the mth dimension of the Lth trajectory point.
  • the delay parameter ⁇ is determined according to the average mutual information:
  • the average mutual information I(X(t),X(t+ ⁇ )) of the one-dimensional time series X(t) and the delayed version X(t+ ⁇ ) is used as an indicator to quantify the independence of the two sequences, as shown in the following formula.
  • the optimal delay parameter ⁇ is determined based on the first minimum point of the mutual information:
  • p ij ( ⁇ ) refers to the joint distribution probability of the corresponding information in X(t) and X(t+ ⁇ ), and p i and p j refer to the probability of the current information;
  • the size after zero addition is 3*N X ;
  • the input data size after phase space reconstruction is 3*m*N psr , where N psr refers to the number of points after phase space reconstruction;
  • the phase space obtained after the respiratory cycle waveform is reconstructed is input into the convolution structure; the convolutional neural network has two convolutional layers for feature extraction.
  • the first convolutional layer includes a two-dimensional convolution, nonlinear transformation and maximum pooling; the second convolutional layer includes one-dimensional convolution and nonlinear transformation; the activation functions used in the two convolutional layers are both rectified linear units; after the convolutional layer extracts the features, the classification label is output through a fully connected layer.
  • the waveform dimension obtained is 3*m*N psr , which includes:
  • One-dimensional phase space convolution perform one-dimensional convolution on the trajectory points in the m-dimensional phase space in order according to the dimension, with a convolution kernel size of 1*K1 and a step size of S1, where K1 represents the size of the convolution kernel of the first convolution layer; merge the features obtained after convolution into one dimension, and perform another one-dimensional convolution with a convolution kernel size of 1*K2 and a step size of S2, where K2 represents the size of the convolution kernel of the second convolution layer; input the obtained features into the fully connected layer and then pass through the softmax function to obtain the classification probability;
  • Two-dimensional phase space convolution perform two-dimensional convolution on trajectory points, with a convolution kernel size of m*K1 and a step size of S2; merge the convolution features into one dimension, and perform another one-dimensional convolution with a convolution kernel size of 1*K2 and a step size of S2; input the obtained features into the fully connected layer and then pass the softmax function to obtain the classification probability.
  • the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a human-machine asynchronous waveform recognition program under a hybrid mechanical ventilation mode, and when the human-machine asynchronous waveform recognition program under the hybrid mechanical ventilation mode is executed by a processor, the steps of the human-machine asynchronous waveform recognition method under the hybrid mechanical ventilation mode as described above are implemented.
  • the present invention provides a method for identifying human-machine asynchrony waveforms in a hybrid mechanical ventilation mode and related equipment, the method comprising: collecting mechanical ventilation waveforms containing human-machine asynchrony in a constant pressure ventilation mode and a constant volume ventilation mode, wherein the mechanical ventilation waveforms select three channels: airway pressure, flow and tidal volume; performing data segmentation, data labeling and data transformation on the mechanical ventilation waveforms to obtain respiratory cycle waveforms; constructing a corresponding The spatial reconstruction-convolutional neural network model, based on the phase space reconstruction-convolutional neural network model, identifies the human-machine asynchrony phenomenon under the mixed constant pressure type and constant volume type ventilation mode according to the input respiratory cycle waveform.
  • phase space reconstruction-convolutional neural network model constructed by the present invention has good generalization performance, can accurately identify the human-machine asynchrony phenomenon under the mixed constant pressure type and constant volume type ventilation mode, and is of great significance to the intelligentization and physiological closed-loop control of the ventilator.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Pulmonology (AREA)
  • Veterinary Medicine (AREA)
  • Public Health (AREA)
  • Animal Behavior & Ethology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Hematology (AREA)
  • Anesthesiology (AREA)
  • Emergency Medicine (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Physiology (AREA)
  • Pathology (AREA)
  • Medical Informatics (AREA)
  • Surgery (AREA)
  • Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)

Abstract

本发明公开了一种混合机械通气模式下人机不同步波形识别方法及相关设备，所述方法包括：采集定压型通气模式和定容型通气模式下包含人机不同步的机械通气波形，所述机械通气波形选取气道压力、流量和潮气量三个通道；将所述机械通气波形进行数据分割、数据标注和数据变换后得到呼吸周期波形；构建相空间重建-卷积神经网络模型，基于所述相空间重建-卷积神经网络模型根据输入的所述呼吸周期波形识别定压型与定容量型通气模式混合下的人机不同步现象。本发明构建的相空间重建-卷积神经网络模型具有良好的泛化性能，可以准确识别定压型与定容量型通气模式混合下的人机不同步现象，对呼吸机的智能化和生理闭环控制具有重要的意义。

Description

混合机械通气模式下人机不同步波形识别方法及相关设备 技术领域
本发明涉及电生理检测监护技术领域，尤其涉及一种混合机械通气模式下人机不同步波形识别方法、系统、终端及计算机可读存储介质。
背景技术
在呼吸机和患者的交互过程称为机械通气，即呼吸机给患者输送氧气，患者通过管路将二氧化碳排出的过程。人机不同步(Patient-Ventilator Asynchrony，PVA)现象是呼吸机与患者之间的交互过程中的一种不协调，过程的这种不协调可能会延长机械通气的时间甚至危害患者的生命。因此及时发现并采取有效措施防止PVA的发生对患者的临床治疗有着重要的意义。
目前人机不同步类型大致有以下四类八种：
(1)触发类型：无效触发、重复触发、自动触发和反向触发。
(2)吸气阶段：流速不匹配。
(3)切换阶段：提前切换和延迟切换。
(4)呼气阶段：内源性呼气末正压(PEEPi)。
在机械通气的临床实践中，医生往往根据患者病情及经验来确定通气模式(Ventilation Mode,VM)。传统的通气模式主要分为两大类，(一)定压型通气模式(包括压力控制模式(Pressure Controlled Ventilation，PCV)和压力支持模式(Pressure Support Ventilation，PSV)等)和(二)定容型通气模式(包括容量控制通气(Volume Controlled Ventilation，VCV)和容量控制辅助通气(Volume Control Assisted Ventilation，VC-AV)等)，如图1所示，图1中(a)和(d)为正常的通气周期波形，图1中(b)和(e)为无效吸气努力周期波形，如箭头所示，图1中(c)和(f)为提前切换周期波形，如箭头所示，图1中(a)-(c)为定容型通气模式下的波形；图1中(d)-(f)为定压型通气模式下的波形。
两种类型的通气模式各有优缺点。定压型通气模式的优势是峰压较低，较少出现气压伤；但潮气量受系统顺应性和粘性组织的影响较大；与之相反的是定容型通气模式能够保证潮气量的供给有利于呼吸肌的休息，但是易导致气压伤。两种通气模式均有一定概率造成人机交互的不同步，且迄今没有关于哪种模式更优的定论。
然而，现在市场上已经有许多呼吸机应用了具有肺保护通气策略的通气模式，如压力控制同步间歇指令通气模式(Pressure Controlled-Synchronized Intermittent Mandatory ventilation，PC-SIMV)、可变压力支持通气模式和容量控制同步间歇指令通气模式，这些模式产生的波形的一个特点是它们从形态学上考虑更像是定压型和定容型通气模式的混合。因此探索识别分类在混合通气模式下的PVA分类的算法模型对呼吸机的智能化和生理闭环控制具有重要的意义。
目前的PVA分类过程可以概括为：(1)获取从呼吸机导出的呼吸力学波形数据；(2)经过专业医师对人机不同步的波形进行标注；(3)数据预处理以及划分数据集为训练集、验证集和测试集；(4)将预处理的数据输入到已经定义好的模型中进行模型训练；(5)保存训练好的模型加以应用。
例如现有技术提到了使用循环神经网络的算法来检测人机不同步，两个通道的GRU(门控循环单元)分别提取压力波形特征和流速时间波形特征，然后将两种特征融合后使用BGRU(双向门控循环神经单元)提取更高维的特征，最后使用softmax全连接层得到对人机不同步类型的分类结果。数据集标注由专业医生标注，共检测四大类人机不同步：流速、触发、周期和其它。例如提出将由医生标注的数据集经过预处理后输入到一维卷积神经网络中进行学习训练，得到一个基于神经网络的预测模型。在预测过程中，通过梯度加权类激活映射的方式，可以获得该模型分类决策的可视化解释。例如通过把采集到的原始呼吸信号转换成二维图像，先使用公开的图像数据集ImageNet训练二维图像多分类的模型，之后以迁移学习的方式，将呼吸波形构成的二维图像输入到模型中并对最后一层的全连接层以上层的权重作微调，得到可用于呼吸波形分类的卷积神经网络。例如实时读取呼吸波形数据构成测试序列，经过标准化之后计算测试序列与训练集里面的所有序列的DTW距离，而后用DTW计算相似性距离，再结合KNN的聚类思想，对测试序列进行分类；用于判断人机不同步现象中无效吸气努力。例如使用模糊熵对脑电信号分析，接着通过特征选择提取出反应脑电信号特征的对应电极下模糊熵作为输入特征，最后将特征用于分类，该发明解决的属于二分类问题。例如首先采用小波尺度变换对原始呼吸波形作一次变换，在此基础上使用多种熵特征提取非线性特征，使用前项选择算法选择出最佳的特征组合后，将其作为支持向量机分类算法的输入进行分类，仍仅分类无效吸气努力这一种人机不同步现象，属于二分类任务。
因此，现有技术还有待于改进和发展。
发明内容
本发明的主要目的在于提供一种混合机械通气模式下人机不同步波形识别方法、系统、终端及计算机可读存储介质，旨在解决现有技术中各种识别算法在进行分类任务时，所利用到的通气数据均采集于单一的通气模式，且无法有效识别人机不同步现象的问题。
为实现上述目的，本发明提供一种混合机械通气模式下人机不同步波形识别方法，所述混合机械通气模式下人机不同步波形识别方法包括如下步骤：
采集定压型通气模式和定容型通气模式下包含人机不同步的机械通气波形，所述机械通气波形选取气道压力、流量和潮气量三个通道；
将所述机械通气波形进行数据分割、数据标注和数据变换后得到呼吸周期波形；
构建相空间重建-卷积神经网络模型，基于所述相空间重建-卷积神经网络模型根据输入的所述呼吸周期波形识别定压型与定容量型通气模式混合下的人机不同步现象。
所述的混合机械通气模式下人机不同步波形识别方法，其中，所述数据分割具体为：
将采集到的所述机械通气波形按呼吸周期分割，对每个呼吸周期提取特征或将一个呼吸周期的数据视作样本输入到深度模型。
所述的混合机械通气模式下人机不同步波形识别方法，其中，所述数据标注具体为：
将每个呼吸周期的机械通气波形进行标签标注。
所述的混合机械通气模式下人机不同步波形识别方法，其中，所述数据变换具体为：
将不同通道的所述机械通气波形进行数据变换，所述数据变换使用z变换：
其中，μ是时间序列X＝{x1,x2,…,xi,…,xn}的均值，σ是X的标准差，xi表示第i个数据。
所述的混合机械通气模式下人机不同步波形识别方法，其中，所述构建相空间重建-卷积神经网络模型具体为：
若一维时间序列X＝{x1,x2,…,xi,…,xn}的长度为N，根据确定的时延参量τ和嵌入维度m，得到对X应用延时嵌入技术后的延时向量：
Xi＝{x1,x2,…,xi+(m-1)τ},i∈[1,L]；
Xi为m维相空间中的一个点，表示延时向量，L＝N-(m-1)τ为相空间中轨迹点的个数；
最终得到相空间轨迹矩阵：
其中，xLm指的是第L个轨迹点的第m个维度的数值。
所述的混合机械通气模式下人机不同步波形识别方法，其中，根据平均互信息确定时延参量τ：
将一维时间序列X(t)和的时延版本X(t+τ)的平均互信息I(X(t),X(t+τ))作为量化两序列独立性的指标，如下公式所示，根据互信息的第一个最小点选取确定最优时延参量τ：
其中，t是时间变量，pij(τ)指X(t)和X(t+τ)中对应各信息的联合分布概率，pi和pj指当前信息的概率；
根据伪近邻确定嵌入维度m：若一维时间序列点之间的距离小于第一预设距离则称为近邻，若将时间序列在一定时延下进行m维嵌入，则计算m维相空间的点的距离，如果距离发生的变化大于第二预设距离，则称为伪近邻，继续改变嵌入维度m直到距离变化小于第三预设距离，则将此时m视为嵌入维度的估计；
构建相空间重建-卷积神经网络模型：
在每个通道的末尾补上足够数量的0使得输入到卷积神经网络的数据大小相同，补0之后的大小为3*NX；经过相空间重建后的输入数据大小为3*m*Npsr，Npsr指相空间重建后点的个数；
呼吸周期波形经过重构后得到的相空间输入卷积结构中；卷积神经网络共设置两个卷积层用于提取特征，第一个卷积层包括一个二维卷积、非线性变换和最大池化；第二个卷积层包括一维卷积和非线性变换；两个卷积层使用的激活函数均为整流线性单元；卷积层提取完特征后，经由一个全连接层输出分类标签。
所述的混合机械通气模式下人机不同步波形识别方法，其中，在原始呼吸波形经过相空间重建后，得到的波形维度为3*m*Npsr，在输入卷积神经网络时，包括：
一维相空间卷积：将m维的相空间中的轨迹点按照维度依次进行一维卷积，卷积核大小为1*K1，步长为S1，K1表示第一个卷积层卷积核的大小；将卷积后得到的特征合并为一个维度，再进行一次一维卷积，卷积核大小为1*K2，步长为S2，K2为第二个卷积层卷积核的大小；将得到的特征输入到全连接层再经过softmax函数得到分类概率；
二维相空间卷积：按轨迹点进行二维卷积，卷积核大小为m*K1，步长为S2；将卷积得到特征合并为一个维度，再进行一次一维卷积，卷积核大小为1*K2，步长为S2；将得到的特征输入到全连接层再经过softmax函数得到分类概率。
此外，为实现上述目的，本发明还提供一种混合机械通气模式下人机不同步波形识别系统，其中，所述混合机械通气模式下人机不同步波形识别系统包括：
数据采集模块，用于采集定压型通气模式和定容型通气模式下包含人机不同步的机械通气波形，所述机械通气波形选取气道压力、流量和潮气量三个通道；
数据处理模块，用于将所述机械通气波形进行数据分割、数据标注和数据变换后得到呼吸周期波形；
数据识别模块，用于构建相空间重建-卷积神经网络模型，基于所述相空间重建-卷积神经网络模型根据输入的所述呼吸周期波形识别定压型与定容量型通气模式混合下的人机不同步现象。
此外，为实现上述目的，本发明还提供一种终端，其中，所述终端包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的混合机械通气模式下 人机不同步波形识别程序，所述混合机械通气模式下人机不同步波形识别程序被所述处理器执行时实现如上所述的混合机械通气模式下人机不同步波形识别方法的步骤。
此外，为实现上述目的，本发明还提供一种计算机可读存储介质，其中，所述计算机可读存储介质存储有混合机械通气模式下人机不同步波形识别程序，所述混合机械通气模式下人机不同步波形识别程序被处理器执行时实现如上所述的混合机械通气模式下人机不同步波形识别方法的步骤。
本发明中，采集定压型通气模式和定容型通气模式下包含人机不同步的机械通气波形，所述机械通气波形选取气道压力、流量和潮气量三个通道；将所述机械通气波形进行数据分割、数据标注和数据变换后得到呼吸周期波形；构建相空间重建-卷积神经网络模型，基于所述相空间重建-卷积神经网络模型根据输入的所述呼吸周期波形识别定压型与定容量型通气模式混合下的人机不同步现象。本发明构建的相空间重建-卷积神经网络模型具有良好的泛化性能，可以准确识别定压型与定容量型通气模式混合下的人机不同步现象，对呼吸机的智能化和生理闭环控制具有重要的意义。
附图说明
图1是机械通气波形示意图；
图2是本发明混合机械通气模式下人机不同步波形识别方法的较佳实施例的流程图；
图3是本发明混合机械通气模式下人机不同步波形识别方法的较佳实施例中延时嵌入相空间的示意图；
图4是本发明混合机械通气模式下人机不同步波形识别方法的较佳实施例中相空间重建-卷积神经网络结构示意图；
图5是本发明混合机械通气模式下人机不同步波形识别系统的较佳实施例的原理示意图；
图6为本发明终端的较佳实施例的运行环境示意图。
具体实施方式
为使本发明的目的、技术方案及优点更加清楚、明确，以下参照附图并举实施 例对本发明进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。
现在市场上已经有许多呼吸机应用了具有肺保护通气策略的通气模式，如压力控制同步间歇指令通气模式(Pressure Controlled-Synchronized Intermittent Mandatory ventilation，PC-SIMV)、可变压力支持通气模式和容量控制同步间歇指令通气模式，这些模式产生的波形的一个特点是它们从形态学上考虑更像是定压型和定容型通气模式的混合。因此探索识别分类在混合通气模式下的PVA分类的算法模型对呼吸机的智能化和生理闭环控制具有重要的意义。
本发明较佳实施例所述的混合机械通气模式下人机不同步波形识别方法，如图1所示，所述混合机械通气模式下人机不同步波形识别方法包括以下步骤：
步骤S10、采集定压型通气模式和定容型通气模式下包含人机不同步的机械通气波形，所述机械通气波形选取气道压力、流量和潮气量三个通道。
具体地，收集不同通气模式(定压型通气模式和定容型通气模式)下的包含人机不同步的机械通气波形(例如这些数据从呼吸机自动导出)，波形选取气道压力、流量和潮气量三个通道。
步骤S20、将所述机械通气波形进行数据分割、数据标注和数据变换后得到呼吸周期波形。
具体地，数据分割：数据分割是指将采集到的所述机械通气波形(原始数据)按呼吸周期(呼吸周期的长短通常根据患者病情不同而不同，一般是3-5秒，但是在波形中，由于采样频率的不同，一个呼吸周期的点的个数也不尽相同，比如采样频率是50HZ，一个呼吸周期是3秒，那么这个周期的点的个数就是150个点)分割，最终将对每个呼吸周期提取特征或将一个呼吸周期的数据视作样本输入深度模型。
具体地，数据标注：将每个呼吸周期的机械通气波形进行标签标注，其中，所述标签包括无效触发、重复触发、自动触发、反向触发、流速不匹配、提前切换、延迟切换、内源性呼气末正压和正常。
具体地，数据变换：将不同通道(量纲)的所述机械通气波形进行数据变换(因为采集的数据来自不同通道，这些通道数据的单位是不一样的，比如，气道压力的单位是cm/H2O，流速的单位是L/min)，以保证深度模型的收敛性和收敛速度，本 发明中，所述数据变换使用z变换：
其中，μ是时间序列X＝{x1,x2,…,xi,…,xn}的均值，σ是X的标准差，xi表示第i个数据。
步骤S30、构建相空间重建-卷积神经网络模型，基于所述相空间重建-卷积神经网络模型根据输入的所述呼吸周期波形识别定压型与定容量型通气模式混合下的人机不同步现象。
具体地，若观察到的一维时间序列X＝{x1,x2,…,xi,…,xn}的长度为N，根据确定的时延参量τ和嵌入维度m，可以得到对X应用延时嵌入技术后的延时向量：
Xi＝{x1,x2,…,xi+(m-1)τ},i∈[1,L]；
Xi为m维相空间中的一个点，它表示延时向量，L＝N-(m-1)τ为相空间中轨迹点的个数。
最终可以得到如下相空间轨迹矩阵(相空间重建的结果，该结果是原始一维时间序列的另一种表现，通过该转化，可以使得卷积神经网络模型可以提取到原始系统的重建后的相空间中的信息，该信息从原来的一维时间序列中无法提取到，且该信息有利于提高分类结果)：
其中，xLm指的是第L个轨迹点的第m个维度的数值。该相空间轨迹矩阵的每一行，都是相空间中的一个轨迹点，这个点有m个维度，一共有L行，即L个轨迹点。
如图3所示，图2给出了延时嵌入相空间的示意图，其中令嵌入维度m为3，时延参量τ为1，最终得到三个通道的3维相空间。
具体地，根据平均互信息确定时延参量τ：
将一维时间序列X(t)(即一维时间序列X)和的时延版本X(t+τ)的平均互信息I(X(t),X(t+τ))作为量化两序列独立性的指标，如下公式所示，根据互信息的第一个最小点选取确定最优时延参量τ：
其中，t是时间变量，pij(τ)指X(t)和X(t+τ)中对应各信息的联合分布概率，pi和pj指当前信息的概率。
具体地，根据伪近邻确定嵌入维度m：若一维时间序列点之间的距离小于第一预设距离则称为近邻(即若点之间的距离很近则称为近邻)，若将时间序列在一定时延下进行m维嵌入，则计算m维相空间的点的距离，如果距离发生的变化大于第二预设距离(即距离发生很大变化)，则称为伪近邻，继续改变嵌入维度m直到距离变化小于第三预设距离(即距离变化不再剧烈)，则将此时m视为嵌入维度的估计。
具体地，构建相空间重建-卷积神经网络模型：
在输入卷积神经网络前，首先为了保证输入到深度模型的数据结构一致，对数据做了补0操作，即在每个通道的末尾补上足够数量的0使得输入到卷积神经网络的数据大小相同，补0之后的大小为3*NX；经过相空间重建后的输入数据大小为3*m*Npsr，Npsr指相空间重建后点的个数，3指的是3个通道。
呼吸周期波形经过重构后得到的相空间输入卷积结构中；如图4所示，卷积神经网络共设置两个卷积层用于提取特征，第一个卷积层包括一个二维卷积、非线性变换和最大池化；第二个卷积层包括一维卷积和非线性变换；两个卷积层使用的激活函数均为整流线性单元(Rectified Linear Units，ReLU)；卷积层提取完特征后，经由一个全连接层输出分类标签。
需要注意的是，在原始呼吸波形经过相空间重建后，得到的波形维度为3*m*Npsr，在输入卷积神经网络时，有两种不同的方式：
(1)一维相空间卷积：将m维的相空间中的轨迹点按照维度依次进行一维卷积，卷积核大小为1*K1，步长为S1，K1表示第一个卷积层卷积核的大小；将卷积 后得到的特征合并为一个维度，再进行一次一维卷积，卷积核大小为1*K2，步长为S2，K2为第二个卷积层卷积核的大小；将得到的特征输入到全连接层再经过softmax函数得到分类概率。
(2)二维相空间卷积：按轨迹点进行二维卷积，卷积核大小为m*K1，步长为S2；将卷积得到特征合并为一个维度，再进行一次一维卷积，卷积核大小为1*K2，步长为S2；将得到的特征输入到全连接层再经过softmax函数得到分类概率。
上述两种方式的差别为卷积是否提取相空间的空间特性信息。
模型训练与评估：模型的训练与评估过程是为了找到一个模型在合适的参数的可以在各项指标中达到最好的过程，此过程中需要设置留一法交叉验证实验检验模型泛化性能，即将每个病例的数据集依次作为测试集，其余病例的数据作为训练集。训练过程通常是将数据依次输入到本发明的所提出的模型中，并设置好训练的【回合次数】参数，然后运行计算机算法即可。每一个回合，程序都会输出当前的准确率和F1-score两个指标，用于判断当前模型的性能。前面提到留一法交叉验证实验，这是一种常见的模型选择方法，通过设置该实验可以得到公认的具有良好泛化性的模型。泛化性能指一个模型在遇到未见过的数据时它所表现出的识别能力。采取准确率和F1-score作为评价指标检验模型泛化性能。各项指标的数值越大，则说明模型泛化性越好。
模型的部署：训练好的模型可以通过各种编程语言部署在呼吸机中或服务器中为各项应用提供服务。
进一步地，设置了留一法交叉验证实验，并对比了基于特征的机器学习模型和未进行相空间重建的卷积神经网络模型在准确度(Accuracy)、灵敏度(Sensitivity)、特异性(Specificity)和F1-score四个指标上的差异。
本实验中的数据来自7名临床机械通气患者，他们均罹患不同的呼吸系统疾病。为7名患者所使用机械通气模式也来自于定压压力控制通气模式、容量控制通气模式和容量控制-同步间歇指令通气模式等。经过数据预处理，最终得到16760个呼吸周期。
留一法交叉验证实验过程中，个别病例的阳性样本过少，因此数据集仅作为训练集，而不用于测试集。最终得到的实验结果如下表所示：


表：实验结果(p<0.01)
上表结果均为重复试验50次后得到的，并进行了t检验，显著性水平α＝0.01。PSR-CNN指本发明提出的模型；CNN指未进行相空间重建但是与PSR-CNN具有相同网络结构的卷积神经网络模型；时域-RF指基于时域特征的随机森林模型；频域-RF指基于频域特征的随机森林模型。
上表的结果显示，本发明提出的模型在各种指标中均其它模型，显示了相空间重建-卷积神经网络模型良好的泛化性能。
进一步地，3个通道可以增加为多个(多于3个)通道，该数据值的变化同时改变相空间重建过程。关于卷积层的结构，则不限于仅使用卷积，还可改变非线性变换函数和增减最大池化层。
进一步地，如图5所示，基于上述混合机械通气模式下人机不同步波形识别方法，本发明还相应提供了一种混合机械通气模式下人机不同步波形识别系统，其中，所述混合机械通气模式下人机不同步波形识别系统包括：
数据采集模块51，用于采集定压型通气模式和定容型通气模式下包含人机不同步的机械通气波形，所述机械通气波形选取气道压力、流量和潮气量三个通道；
数据处理模块52，用于将所述机械通气波形进行数据分割、数据标注和数据变换后得到呼吸周期波形；
数据识别模块53，用于构建相空间重建-卷积神经网络模型，基于所述相空间重建-卷积神经网络模型根据输入的所述呼吸周期波形识别定压型与定容量型通气模式混合下的人机不同步现象。
进一步地，如图6所示，基于上述混合机械通气模式下人机不同步波形识别方法和系统，本发明还相应提供了一种终端，所述终端包括处理器10、存储器20及显示器30。图6仅示出了终端的部分组件，但是应理解的是，并不要求实施所有示出的组件，可以替代的实施更多或者更少的组件。
所述存储器20在一些实施例中可以是所述终端的内部存储单元，例如终端的硬盘或内存。所述存储器20在另一些实施例中也可以是所述终端的外部存储设备，例如所述终端上配备的插接式硬盘，智能存储卡(Smart Media Card,SMC)，安全数字 (Secure Digital,SD)卡，闪存卡(Flash Card)等。进一步地，所述存储器20还可以既包括所述终端的内部存储单元也包括外部存储设备。所述存储器20用于存储安装于所述终端的应用软件及各类数据，例如所述安装终端的程序代码等。所述存储器20还可以用于暂时地存储已经输出或者将要输出的数据。在一实施例中，存储器20上存储有混合机械通气模式下人机不同步波形识别程序40，该混合机械通气模式下人机不同步波形识别程序40可被处理器10所执行，从而实现本申请中混合机械通气模式下人机不同步波形识别方法。
所述处理器10在一些实施例中可以是一中央处理器(Central Processing Unit,CPU)，微处理器或其他数据处理芯片，用于运行所述存储器20中存储的程序代码或处理数据，例如执行所述混合机械通气模式下人机不同步波形识别方法等。
所述显示器30在一些实施例中可以是LED显示器、液晶显示器、触控式液晶显示器以及OLED(Organic Light-Emitting Diode，有机发光二极管)触摸器等。所述显示器30用于显示在所述终端的信息以及用于显示可视化的用户界面。所述终端的部件10-30通过系统总线相互通信。
在一实施例中，当处理器10执行所述存储器20中混合机械通气模式下人机不同步波形识别程序40时实现以下步骤：
采集定压型通气模式和定容型通气模式下包含人机不同步的机械通气波形，所述机械通气波形选取气道压力、流量和潮气量三个通道；
将所述机械通气波形进行数据分割、数据标注和数据变换后得到呼吸周期波形；
构建相空间重建-卷积神经网络模型，基于所述相空间重建-卷积神经网络模型根据输入的所述呼吸周期波形识别定压型与定容量型通气模式混合下的人机不同步现象。
其中，所述数据分割具体为：
将采集到的所述机械通气波形按呼吸周期分割，对每个呼吸周期提取特征或将一个呼吸周期的数据视作样本输入到深度模型。
其中，所述数据标注具体为：
将每个呼吸周期的机械通气波形进行标签标注。
其中，所述数据变换具体为：
将不同通道的所述机械通气波形进行数据变换，所述数据变换使用z变换：
其中，μ是时间序列X＝{x1,x2,…,xi,…,xn}的均值，σ是X的标准差，xi表示第i个数据。
其中，所述构建相空间重建-卷积神经网络模型具体为：
若一维时间序列X＝{x1,x2,…,xi,…,xn}的长度为N，根据确定的时延参量τ和嵌入维度m，得到对X应用延时嵌入技术后的延时向量：
Xi＝{x1,x2,…,xi+(m-1)τ},i∈[1,L]；
Xi为m维相空间中的一个点，表示延时向量，L＝N-(m-1)τ为相空间中轨迹点的个数；
最终得到相空间轨迹矩阵：
其中，xLm指的是第L个轨迹点的第m个维度的数值。
其中，根据平均互信息确定时延参量τ：
将一维时间序列X(t)和的时延版本X(t+τ)的平均互信息I(X(t),X(t+τ))作为量化两序列独立性的指标，如下公式所示，根据互信息的第一个最小点选取确定最优时延参量τ：
其中，t是时间变量，pij(τ)指X(t)和X(t+τ)中对应各信息的联合分布概率，pi和pj指当前信息的概率；
根据伪近邻确定嵌入维度m：若一维时间序列点之间的距离小于第一预设距离 则称为近邻，若将时间序列在一定时延下进行m维嵌入，则计算m维相空间的点的距离，如果距离发生的变化大于第二预设距离，则称为伪近邻，继续改变嵌入维度m直到距离变化小于第三预设距离，则将此时m视为嵌入维度的估计；
构建相空间重建-卷积神经网络模型：
在每个通道的末尾补上足够数量的0使得输入到卷积神经网络的数据大小相同，补0之后的大小为3*NX；经过相空间重建后的输入数据大小为3*m*Npsr，Npsr指相空间重建后点的个数；
呼吸周期波形经过重构后得到的相空间输入卷积结构中；卷积神经网络共设置两个卷积层用于提取特征，第一个卷积层包括一个二维卷积、非线性变换和最大池化；第二个卷积层包括一维卷积和非线性变换；两个卷积层使用的激活函数均为整流线性单元；卷积层提取完特征后，经由一个全连接层输出分类标签。
其中，在原始呼吸波形经过相空间重建后，得到的波形维度为3*m*Npsr，在输入卷积神经网络时，包括：
一维相空间卷积：将m维的相空间中的轨迹点按照维度依次进行一维卷积，卷积核大小为1*K1，步长为S1，K1表示第一个卷积层卷积核的大小；将卷积后得到的特征合并为一个维度，再进行一次一维卷积，卷积核大小为1*K2，步长为S2，K2为第二个卷积层卷积核的大小；将得到的特征输入到全连接层再经过softmax函数得到分类概率；
二维相空间卷积：按轨迹点进行二维卷积，卷积核大小为m*K1，步长为S2；将卷积得到特征合并为一个维度，再进行一次一维卷积，卷积核大小为1*K2，步长为S2；将得到的特征输入到全连接层再经过softmax函数得到分类概率。
本发明还提供一种计算机可读存储介质，其中，所述计算机可读存储介质存储有混合机械通气模式下人机不同步波形识别程序，所述混合机械通气模式下人机不同步波形识别程序被处理器执行时实现如上所述的混合机械通气模式下人机不同步波形识别方法的步骤。
综上所述，本发明提供一种混合机械通气模式下人机不同步波形识别方法及相关设备，所述方法包括：采集定压型通气模式和定容型通气模式下包含人机不同步的机械通气波形，所述机械通气波形选取气道压力、流量和潮气量三个通道；将所述机械通气波形进行数据分割、数据标注和数据变换后得到呼吸周期波形；构建相 空间重建-卷积神经网络模型，基于所述相空间重建-卷积神经网络模型根据输入的所述呼吸周期波形识别定压型与定容量型通气模式混合下的人机不同步现象。本发明构建的相空间重建-卷积神经网络模型具有良好的泛化性能，可以准确识别定压型与定容量型通气模式混合下的人机不同步现象，对呼吸机的智能化和生理闭环控制具有重要的意义。
需要说明的是，在本文中，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者终端不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者终端所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括该要素的过程、方法、物品或者终端中还存在另外的相同要素。
当然，本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关硬件(如处理器，控制器等)来完成，所述的程序可存储于一计算机可读取的计算机可读存储介质中，所述程序在执行时可包括如上述各方法实施例的流程。其中所述的计算机可读存储介质可为存储器、磁碟、光盘等。
应当理解的是，本发明的应用不限于上述的举例，对本领域普通技术人员来说，可以根据上述说明加以改进或变换，所有这些改进和变换都应属于本发明所附权利要求的保护范围。

Claims (10)

  1. 一种混合机械通气模式下人机不同步波形识别方法，其特征在于，所述混合机械通气模式下人机不同步波形识别方法包括：
    采集定压型通气模式和定容型通气模式下包含人机不同步的机械通气波形，所述机械通气波形选取气道压力、流量和潮气量三个通道；
    将所述机械通气波形进行数据分割、数据标注和数据变换后得到呼吸周期波形；
    构建相空间重建-卷积神经网络模型，基于所述相空间重建-卷积神经网络模型根据输入的所述呼吸周期波形识别定压型与定容量型通气模式混合下的人机不同步现象。
  2. 根据权利要求1所述的混合机械通气模式下人机不同步波形识别方法，其特征在于，所述数据分割具体为：
    将采集到的所述机械通气波形按呼吸周期分割，对每个呼吸周期提取特征或将一个呼吸周期的数据视作样本输入到深度模型。
  3. 根据权利要求2所述的混合机械通气模式下人机不同步波形识别方法，其特征在于，所述数据标注具体为：
    将每个呼吸周期的机械通气波形进行标签标注。
  4. 根据权利要求3所述的混合机械通气模式下人机不同步波形识别方法，其特征在于，所述数据变换具体为：
    将不同通道的所述机械通气波形进行数据变换，所述数据变换使用z变换：
    其中，μ是时间序列X＝{x1,x2,…,xi,…,xn}的均值，σ是X的标准差，xi表示第i个数据。
  5. 根据权利要求4所述的混合机械通气模式下人机不同步波形识别方法，其特征在于，所述构建相空间重建-卷积神经网络模型具体为：
    若一维时间序列X＝{x1,x2,…,xi,…,xn}的长度为N，根据确定的时延参量τ和嵌入维度m，得到对X应用延时嵌入技术后的延时向量：
    Xi＝{x1,x2,…,xi+(m-1)τ},i∈[1,L]；
    Xi为m维相空间中的一个点，表示延时向量，L＝N-(m-1)τ为相空间中轨迹点的个数；
    最终得到相空间轨迹矩阵：
    其中，xLm指的是第L个轨迹点的第m个维度的数值。
  6. 根据权利要求5所述的混合机械通气模式下人机不同步波形识别方法，其特征在于，根据平均互信息确定时延参量τ：
    将一维时间序列X(t)和的时延版本X(t+τ)的平均互信息I(X(t),X(t+τ))作为量化两序列独立性的指标，如下公式所示，根据互信息的第一个最小点选取确定最优时延参量τ：
    其中，t是时间变量，pij(τ)指X(t)和X(t+τ)中对应各信息的联合分布概率，pi和pj指当前信息的概率；
    根据伪近邻确定嵌入维度m：若一维时间序列点之间的距离小于第一预设距离则称为近邻，若将时间序列在一定时延下进行m维嵌入，则计算m维相空间的点的距离，如果距离发生的变化大于第二预设距离，则称为伪近邻，继续改变嵌入维度m直到距离变化小于第三预设距离，则将此时m视为嵌入维度的估计；
    构建相空间重建-卷积神经网络模型：
    在每个通道的末尾补上足够数量的0使得输入到卷积神经网络的数据大小相同，补0之后的大小为3*NX；经过相空间重建后的输入数据大小为3*m*Npsr，Npsr指相空间重建后点的个数；
    呼吸周期波形经过重构后得到的相空间输入卷积结构中；卷积神经网络共设置两个卷积层用于提取特征，第一个卷积层包括一个二维卷积、非线性变换和最大池化；第二个卷积层包括一维卷积和非线性变换；两个卷积层使用的激活函数均为整流线性单元；卷积层提取完特征后，经由一个全连接层输出分类标签。
  7. 根据权利要求6所述的混合机械通气模式下人机不同步波形识别方法，其特征在于，在原始呼吸波形经过相空间重建后，得到的波形维度为3*m*Npsr，在输入卷积神经网络时，包括：
    一维相空间卷积：将m维的相空间中的轨迹点按照维度依次进行一维卷积，卷积核大小为1*K1，步长为S1，K1表示第一个卷积层卷积核的大小；将卷积后得到的特征合并为一个维度，再进行一次一维卷积，卷积核大小为1*K2，步长为S2，K2为第二个卷积层卷积核的大小；将得到的特征输入到全连接层再经过softmax函数得到分类概率；
    二维相空间卷积：按轨迹点进行二维卷积，卷积核大小为m*K1，步长为S2；将卷积得到特征合并为一个维度，再进行一次一维卷积，卷积核大小为1*K2，步长为S2；将得到的特征输入到全连接层再经过softmax函数得到分类概率。
  8. 一种混合机械通气模式下人机不同步波形识别系统，其特征在于，所述混合机械通气模式下人机不同步波形识别系统包括：
    数据采集模块，用于采集定压型通气模式和定容型通气模式下包含人机不同步的机械通气波形，所述机械通气波形选取气道压力、流量和潮气量三个通道；
    数据处理模块，用于将所述机械通气波形进行数据分割、数据标注和数据变换后得到呼吸周期波形；
    数据识别模块，用于构建相空间重建-卷积神经网络模型，基于所述相空间重建-卷积神经网络模型根据输入的所述呼吸周期波形识别定压型与定容量型通气模式混合下的人机不同步现象。
  9. 一种终端，其特征在于，所述终端包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的混合机械通气模式下人机不同步波形识别程序，所述混合机械通气模式下人机不同步波形识别程序被所述处理器执行时实现如权利要求1-7任一项所述的混合机械通气模式下人机不同步波形识别方法的步骤。
  10. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质存储有 混合机械通气模式下人机不同步波形识别程序，所述混合机械通气模式下人机不同步波形识别程序被处理器执行时实现如权利要求1-7任一项所述的混合机械通气模式下人机不同步波形识别方法的步骤。
PCT/CN2023/075074 2023-01-17 2023-02-08 混合机械通气模式下人机不同步波形识别方法及相关设备 WO2024152392A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202310058842.1A CN115814222B (zh) 2023-01-17 2023-01-17 混合机械通气模式下人机不同步波形识别方法及相关设备
CN202310058842.1 2023-01-17

Publications (1)

Publication Number Publication Date
WO2024152392A1 true WO2024152392A1 (zh) 2024-07-25

Family

ID=85520758

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/075074 WO2024152392A1 (zh) 2023-01-17 2023-02-08 混合机械通气模式下人机不同步波形识别方法及相关设备

Country Status (2)

Country Link
CN (1) CN115814222B (zh)
WO (1) WO2024152392A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117045913B (zh) * 2023-07-14 2024-04-30 南通大学附属医院 一种基于呼吸变量监测的机械通气模式智能切换系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106453293A (zh) * 2016-09-30 2017-02-22 重庆邮电大学 一种基于改进bpnn的网络安全态势预测方法
CN110555468A (zh) * 2019-08-15 2019-12-10 武汉科技大学 一种联合递归图和cnn的脑电信号识别方法及系统
CN112560919A (zh) * 2020-12-07 2021-03-26 杭州智瑞思科技有限公司 基于一维可解释卷积神经网络的人机不同步识别方法
CN113521460A (zh) * 2021-05-20 2021-10-22 深圳先进技术研究院 机械通气人机异步检测方法、装置及计算机可读存储介质
CN114154551A (zh) * 2021-10-19 2022-03-08 中山大学 一种基于深度学习的多元时间序列分类方法
US20230000445A1 (en) * 2009-04-22 2023-01-05 Alton Reich Cardiovascular state determination apparatus and method of use thereof

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150090258A1 (en) * 2013-10-01 2015-04-02 Covidien Lp Ventilator-initiated prompt or setting regarding detection of asynchrony during ventilation
EP3706845B1 (en) * 2017-11-09 2023-10-18 Autonomous Healthcare, Inc. Clinical decision support system for patient-ventilator asynchrony detection and management
US12073945B2 (en) * 2020-06-30 2024-08-27 Cerner Innovation, Inc. Patient ventilator asynchrony detection
CN113539501A (zh) * 2021-06-25 2021-10-22 中国科学院深圳先进技术研究院 一种呼吸机人机异步分类方法、系统、终端以及存储介质
CN113951868B (zh) * 2021-10-29 2024-04-09 北京富通东方科技有限公司 机械通气患者人机不同步检测方法及装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20230000445A1 (en) * 2009-04-22 2023-01-05 Alton Reich Cardiovascular state determination apparatus and method of use thereof
CN106453293A (zh) * 2016-09-30 2017-02-22 重庆邮电大学 一种基于改进bpnn的网络安全态势预测方法
CN110555468A (zh) * 2019-08-15 2019-12-10 武汉科技大学 一种联合递归图和cnn的脑电信号识别方法及系统
CN112560919A (zh) * 2020-12-07 2021-03-26 杭州智瑞思科技有限公司 基于一维可解释卷积神经网络的人机不同步识别方法
CN113521460A (zh) * 2021-05-20 2021-10-22 深圳先进技术研究院 机械通气人机异步检测方法、装置及计算机可读存储介质
CN114154551A (zh) * 2021-10-19 2022-03-08 中山大学 一种基于深度学习的多元时间序列分类方法

Also Published As

Publication number Publication date
CN115814222A (zh) 2023-03-21
CN115814222B (zh) 2023-04-14

Similar Documents

Publication Publication Date Title
US9158382B2 (en) Medical information display apparatus, method, and program
CN109427060A (zh) 一种影像识别的方法、装置、终端设备和医疗系统
Ye et al. FENet: a frequency extraction network for obstructive sleep apnea detection
WO2024152392A1 (zh) 混合机械通气模式下人机不同步波形识别方法及相关设备
WO2022267381A1 (zh) 一种呼吸机人机异步分类方法、系统、终端以及存储介质
WO2020224433A1 (zh) 基于机器学习的目标对象属性预测方法及相关设备
Tobias et al. CNN-based deep learning model for chest X-ray health classification using tensorflow
WO2022267382A1 (zh) 一种呼吸机人机异步分类方法、系统、终端以及存储介质
WO2023097780A1 (zh) 机械通气过程中人机异步现象的分类方法和分类装置
Tahir et al. A CNN-ELM-based method for ballistocardiogram classification in a clinical environment
Rajasenbagam et al. Semantic content-based image retrieval system using deep learning model for lung cancer CT images
CN117258110A (zh) 生物反馈的深度放松导引系统及其方法
Chen et al. Attention-based convolutional long short-term memory neural network for detection of patient-ventilator asynchrony from mechanical ventilation
Hayashi et al. Patient deterioration detection using one-class classification via cluster period estimation subtask
CN113941061B (zh) 一种人机不同步识别方法、系统、终端以及存储介质
CN115565245A (zh) 一种基于rgb视频监控的icu患者自拔管行为预警方法
CN112487980B (zh) 基于微表情治疗方法、装置、系统与计算机可读存储介质
WO2023097785A1 (zh) 一种基于模糊熵特征提取的人机通气异步检测模型及装置
Alsalamah et al. Detection of obstructive sleep apnea using deep neural network
JP2021186675A (ja) 心臓マッピングにおける心構造の自動検出
Alsalamah et al. Clinical practice for diagnostic causes for obstructive sleep apnea using artificial intelligent neural networks
Zhou et al. Early Screening of ASD Based on Hand Gesture Analysis
Muammer et al. Employing X-ray images characteristics for early corona virus detection based on deep learning techniques
CN113730755B (zh) 一种基于注意力机制的机械通气人机异步检测识别方法
Dong et al. Experiences-based intelligence simulation in ECG recognition

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: 23916869

Country of ref document: EP

Kind code of ref document: A1