Anesthesia depth monitoring method, medium and electronic device

Abstract

The invention provides an anesthesia depth monitoring method, a medium and an electronic device, and belongs to the technical field of anesthesia depth monitoring, and the anesthesia depth monitoring method comprises the following steps: collecting an electroencephalogram signal, an electrocardiogram signal and a respiratory activity signal of a patient, and carrying out preprocessing, grouping and feature extraction to realize anesthesia depth monitoring. The method comprises the following steps: firstly, segmenting a signal, calculating the similarity, and classifying subsequences into stable and fluctuating subsequences; and then, determining a step length according to the alternative interval, processing stable and fluctuating subsequences, and meanwhile, filtering, reducing noise and recording an abnormal waveform segment. Then, the signals and the abnormal waveform segments are processed in parallel on the GPU and the CPU, and features are extracted; and finally, through clustering and dimension reduction optimization features and in combination with abnormal waveform segment features, obtaining final anesthesia depth features, and inputting the anesthesia depth features into a monitoring model to obtain a monitoring result. The technical problem that the anesthesia depth monitoring result is not accurate enough due to the fact that interference or small fluctuation is often difficult to process in an existing anesthesia depth monitoring technology is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of anesthesia depth monitoring, and in particular, relates to an anesthesia depth monitoring method, medium and electronic device. Background Art

[0002] Anesthesia depth monitoring is one of the key links in clinical anesthesia management. Accurately monitoring the patient's anesthesia depth can help anesthesiologists adjust the dosage of anesthetic drugs in time, avoid excessive or insufficient anesthesia, and ensure that the patient maintains the best anesthesia state during surgery. Existing anesthesia depth monitoring technologies mainly include monitoring methods based on physiological signals and monitoring methods based on pharmacodynamic models.

[0003] Physiological signal monitoring uses changes in the patient's physiological characteristics, such as electroencephalogram (EEG), electrocardiogram (ECG), respiratory activity, etc., and evaluates the depth of anesthesia by analyzing the characteristics of these signals. This method directly reflects the changes in the patient's physiological state and has strong objectivity and sensitivity. At present, the "bispectral index" (BIS) monitoring technology based on EEG is one of the most widely used physiological signal monitoring methods in clinical practice. BIS monitoring technology extracts and analyzes the time-frequency characteristics of EEG signals and gives an index value of 0-100 to reflect the depth of anesthesia. Although BIS technology has been widely used in clinical practice, it still has some defects: 1) BIS monitoring only relies on a single EEG signal and cannot fully utilize the information of other physiological signals; 2) BIS has certain individual differences in the response of different patients, and patients need to be calibrated; 3) In some surgical procedures, such as cardiac surgery, the accuracy of BIS monitoring will be disturbed.

[0004] The pharmacodynamic monitoring method is based on the establishment of a detailed pharmacodynamic model to predict the depth of anesthesia by monitoring the concentration changes of anesthetic drugs. This method establishes a mathematical relationship model between anesthetic drugs and physiological reactions, which can accurately describe the absorption, distribution, metabolism and action process of anesthetic drugs. However, the pharmacodynamic model requires a large amount of experimental data support, and the model parameters need to be individually calibrated for different patients, which makes the operation more complex. At the same time, this method cannot directly reflect the changes in the patient's physiological state, and may not be able to detect some abnormal conditions in time.

[0005] Currently, in some complex surgical scenarios, because the patient is under anesthesia, different physiological characteristic signals are often interfered with or produce small fluctuations themselves. The existing anesthesia depth monitoring technology often finds it difficult to deal with these interferences or small fluctuations, resulting in inaccurate anesthesia depth monitoring results. Summary of the invention

[0006] In view of this, the present invention provides an anesthesia depth monitoring method, medium and electronic device, which can solve the technical problem that the existing anesthesia depth monitoring technology is often difficult to deal with these interferences or small fluctuations, resulting in inaccurate anesthesia depth monitoring results.

[0007] The present invention is achieved in that:

[0008] A first aspect of the present invention provides a method for monitoring depth of anesthesia, comprising the following steps:

[0009] S10, continuously collecting the patient's electroencephalogram signal, electrocardiogram signal and respiratory activity signal as the original signal for anesthesia depth monitoring;

[0010] S20, dividing the signal into multiple sub-segments based on the average value of each preprocessed signal, marking the maximum value point and the minimum value point in each sub-segment, and determining the candidate interval of the grouping threshold according to the distance between the maximum value point and the minimum value point;

[0011] S30, dividing each preprocessed signal into a plurality of subsequences according to the determined grouping threshold, calculating the similarity between adjacent subsequences, and classifying the plurality of subsequences into stable subsequences and fluctuating subsequences based on the difference in similarity;

[0012] S40, based on the candidate interval of the grouping threshold determined in S20 and the classification results of the stable subsequence and the fluctuating subsequence in S30, performing differential processing on the anesthesia depth feature to determine a smaller step size used for the stable subsequence processing and a larger step size used for the fluctuating subsequence processing;

[0013] S50, filtering and denoising the collected original signal to obtain a preprocessed signal, and recording the abnormal mutation segments excluded in the filtering and denoising process to form an abnormal waveform segment set;

[0014] S60, according to the step size determined in S40, start the GPU parallel processing flow to process each pre-processed signal, and start the CPU to process the abnormal waveform segment set; perform short-time Fourier transform, Mel-frequency cepstrum coefficient analysis and wavelet packet decomposition on each pre-processed signal in the GPU, extract signal features, and obtain electroencephalogram features, electrocardiogram features, and respiratory activity features; use a sliding window method in the CPU to extract features from the abnormal waveform segment set using time-frequency analysis and statistical feature extraction methods, and record them as abnormal waveform segment features;

[0015] S70, performing cluster analysis on the electroencephalogram features, electrocardiogram features, and respiratory activity features respectively, obtaining a cluster center representing a typical anesthesia depth feature for each signal, and replacing the original features in the electroencephalogram features, electrocardiogram features, and respiratory activity features with the cluster center to form optimized electroencephalogram features, electrocardiogram features, and respiratory activity features;

[0016] S80, aligning and fusing the optimized electroencephalogram features, electrocardiogram features, respiratory activity features, and abnormal waveform segment features of each signal, using principal component analysis to reduce the dimension, obtaining the reduced dimension features, and performing standardization processing to finally obtain the anesthesia depth features;

[0017] S90: Use a pre-trained anesthesia depth monitoring model, input the anesthesia depth feature, obtain the patient's anesthesia depth monitoring result and output it.

[0018] The larger step size is 0.1 to 0.2 seconds, and the smaller step size is 0.01 to 0.02 seconds.

[0019] Wherein, in the step of calculating the similarity between adjacent subsequences, the similarity is calculated using correlation coefficient similarity.

[0020] Among them, the anesthesia depth monitoring model adopts a multi-branch parallel convolutional neural network structure, including an electroencephalogram branch network, an electrocardiogram branch network, a respiratory activity branch network, an abnormal waveform branch network and a fusion branch network.

[0021] Furthermore, the EEG branch network is used to process EEG features, the input is the optimized EEG features, the output is the EEG feature vector, and the structure is a multi-layer one-dimensional convolutional neural network;

[0022] The electrocardiogram branch network is used to process electrocardiogram features, the input is the optimized electrocardiogram features, the output is the electrocardiogram feature vector, and the structure is a multi-layer one-dimensional convolutional neural network;

[0023] The respiratory activity branch network is used to process respiratory activity features, the input is the optimized respiratory activity features, the output is the respiratory activity feature vector, and the structure is a multi-layer one-dimensional convolutional neural network;

[0024] The abnormal waveform branch network is used to process abnormal waveform segment features, the input is the abnormal waveform segment features, the output is the abnormal waveform feature vector, and the structure is a multi-layer fully connected neural network;

[0025] The fusion branch network is used to integrate the characteristics of each branch, the input is the output feature vector of each branch network, the output is the final anesthesia depth prediction value, and the structure is a multi-layer fully connected neural network.

[0026] Furthermore, the step of obtaining the training set for training the anesthesia depth monitoring model is specifically:

[0027] Step 11, collecting a large number of patients' anesthesia process data, including electroencephalogram, electrocardiogram, respiratory activity signal and corresponding anesthesia depth score;

[0028] Step 12: Process the original signal according to steps S10-S80 to extract features;

[0029] Step 13: Pair the extracted features with the corresponding anesthesia depth scores to form training samples;

[0030] Step 14: Perform data enhancement on the training samples, such as adding noise, time distortion, etc.

[0031] Step 15: Divide the processed training samples into training set, validation set and test set.

[0032] Further, the specific training steps of the anesthesia depth monitoring model are:

[0033] Step 21, initializing the parameters of each branch network;

[0034] Step 22: Use the mini-batch gradient descent method and the mean square error as the loss function to start training;

[0035] Step 23, for each batch of data: input the data into each branch network to obtain the prediction result; calculate the error between the prediction result and the actual anesthesia depth score; back propagate the error and update the network parameters;

[0036] Step 24: Evaluate the model performance on the validation set and save the model if the performance improves.

[0037] Step 25, repeat steps 3-4 until a preset number of iterations is reached or the model converges, to obtain the anesthesia depth monitoring model.

[0038] Wherein, in the step of performing cluster analysis on the electroencephalogram features, electrocardiogram features, and respiratory activity features respectively, the clustering method adopted is K-means clustering.

[0039] Specifically, the step S10 specifically includes: continuously collecting the patient's electroencephalogram signal, electrocardiogram signal and respiratory activity signal as the original signal for monitoring the depth of anesthesia. By wearing corresponding physiological signal acquisition equipment, such as electroencephalogram electrodes, electrocardiogram electrodes and respiratory sensors, the patient's electroencephalogram signal, electrocardiogram signal and respiratory activity signal during anesthesia are continuously captured. These three physiological signals all reflect the changes in the patient's physiological state under anesthesia and are important bases for evaluating the depth of anesthesia. By continuously collecting these original signals, basic data is provided for subsequent signal preprocessing and feature extraction.

[0040] Wherein, the step S20 specifically includes: firstly, preprocessing the collected original signal, such as denoising, filtering, etc., to obtain a preprocessed signal. Then, based on the average value of each preprocessed signal, it is divided into multiple sub-segments. For each sub-segment, the maximum value point and the minimum value point are found, and the distance between the two points is calculated. According to the distance between these maximum value points and minimum value points, the candidate interval of the threshold for subsequent grouping is determined. The purpose of this step is to find a suitable grouping threshold to lay the foundation for subsequent signal classification.

[0041] Wherein, the step S30 specifically includes: first, according to the grouping threshold determined in step S20, each pre-processed signal is divided into several groups of subsequences. Then, the similarity between adjacent subsequences is calculated, and the correlation coefficient or other similarity measurement methods can be used. Next, according to the difference in similarity between adjacent subsequences, these subsequences are classified into stable subsequences and fluctuating subsequences. Stable subsequences indicate that the signal is relatively stable during this period of time, and fluctuating subsequences indicate that the signal has large fluctuations during this period of time. The purpose of this step is to distinguish the stable and fluctuating features in the signal, and provide a basis for subsequent feature extraction and processing.

[0042] Wherein, the step S40 specifically includes: according to the alternative interval of the grouping threshold determined in step S20, combined with the classification results of the stable subsequence and the fluctuating subsequence obtained in step S30, differential processing is performed on the anesthesia depth features to be extracted subsequently. For the stable subsequence, a smaller step size is used for feature extraction; for the fluctuating subsequence, a larger step size is used for feature extraction. The purpose of this step is to use different feature extraction step sizes in stable and fluctuating states, so as to better capture the feature changes in the signal.

[0043] Wherein, the step S50 specifically includes: firstly, filtering and denoising the original signal collected in step S10 to remove some irrelevant interference components and obtain a preprocessed signal. In this process, some abnormal and sudden signal segments will be eliminated, and these abnormal waveform segments will be recorded to form an abnormal waveform segment set. Through this step, a relatively smooth preprocessed signal is obtained, and abnormal waveform segments that may contain valuable information are retained.

[0044] Wherein, the step S60 specifically includes: first, according to the different step sizes determined in step S40, respectively start the GPU and CPU for parallel processing. The GPU processing flow is used to process each preprocessed signal, and the features representing the electroencephalogram, electrocardiogram and respiratory activity are extracted by using methods such as short-time Fourier transform, Mel frequency cepstral coefficient analysis and wavelet packet decomposition. The CPU processing flow is used to process the abnormal waveform segment set recorded in step S50, and the sliding window is used. Wherein, the step S70 specifically includes: clustering analysis of the electroencephalogram features, electrocardiogram features and respiratory activity features extracted in step S60, for example, a clustering algorithm such as K-means or Gaussian mixture model can be used. Through clustering, the feature clustering center representing the typical anesthesia depth state in each signal is found. Then, these clustering centers are used to replace the original features to form an optimized feature set. The purpose of this step is to remove some redundant or noise features through clustering analysis and extract more representative anesthesia depth features.

[0045] Wherein, the step S80 specifically includes: firstly aligning and fusing the optimized EEG features, ECG features and respiratory activity features obtained in step S70, and the abnormal waveform segment features extracted in step S60, to form a comprehensive feature set. In order to reduce the data dimension, principal component analysis (PCA) is used to reduce the dimension and obtain the reduced dimension features. Finally, these reduced dimension features are standardized to obtain the final anesthesia depth features. The purpose of this step is to integrate the features of different signal sources, and further enhance the representativeness and robustness of the features through dimensionality reduction and standardization.

[0046] Wherein, the step S90 specifically includes: first, a pre-trained anesthesia depth monitoring model is required. The model adopts a multi-branch parallel convolutional neural network structure, including an electroencephalogram branch network, an electrocardiogram branch network, a respiratory activity branch network, an abnormal waveform branch network, and a fusion branch network. The input is the anesthesia depth feature obtained in step S80, and after processing and fusion of these branch networks, the patient's anesthesia depth monitoring result is finally output. The purpose of this step is to use the pre-trained deep learning model to quickly and accurately predict the patient's anesthesia depth state based on the extracted anesthesia depth features.

[0047] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned anesthesia depth monitoring method.

[0048] A third aspect of the present invention provides an electronic device, which comprises a processor and a memory, wherein the memory is used to store the step program of the above-mentioned method, and the processor reads the memory and executes the steps.

[0049] Compared with the prior art, the present invention provides an anesthesia depth monitoring method, medium and electronic device to solve the deficiencies in the prior art. The method, medium and electronic device continuously collect multiple physiological signals such as the patient's electroencephalogram, electrocardiogram and respiratory activity, use advanced signal processing and machine learning technology to extract key features representing the anesthesia depth, and train an intelligent prediction model to accurately monitor the patient's anesthesia depth.

[0050] Compared with the prior art, the main technical effects of the present invention are as follows:

[0051] 1. Multimodal signal fusion to improve monitoring accuracy: The present invention simultaneously collects and utilizes multiple physiological signals of the patient, such as the electroencephalogram, electrocardiogram and respiratory activity, and makes full use of the information contained in different signal sources. Compared with the monitoring method of a single signal, it can more comprehensively and accurately reflect the patient's anesthesia depth state.

[0052] 2. Differential processing and parallel computing to enhance robustness: This invention adopts differential processing methods and different feature extraction steps to better capture the feature changes in the signal, especially in some complex surgical scenarios, where different physiological feature signals are often interfered with or have small fluctuations themselves. At the same time, the parallel computing capabilities of GPU and CPU are fully utilized to greatly improve the efficiency and reliability of signal processing.

[0053] 3. Advanced machine learning model to achieve intelligent monitoring: The present invention adopts a multi-branch parallel convolutional neural network as an anesthesia depth prediction model, which can automatically learn the deep-level characteristics of various physiological signals without relying on complex artificial feature engineering, thereby achieving intelligent and high-precision monitoring of anesthesia depth.

[0054] 4. Identification and processing of abnormal situations to improve applicability: During the signal processing process, the present invention pays special attention to some abnormal signal waveforms, and uses methods such as sliding windows and time-frequency analysis to extract and classify their features, which can better cope with various complex surgical scenarios and improve the applicability of the method.

[0055] In summary, the anesthesia depth monitoring method based on multimodal physiological signal fusion proposed in the present invention makes full use of the information of multiple physiological signals and adopts advanced signal processing and machine learning technologies to solve the technical problem that existing anesthesia depth monitoring technologies are often difficult to deal with these interferences or small fluctuations, resulting in inaccurate anesthesia depth monitoring results. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flow chart of the method provided by the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0058] like Figure 1 FIG. 1 is a flow chart of a method for monitoring the depth of anesthesia provided by the present invention. The method comprises the following steps:

[0059] S10, continuously collecting the patient's electroencephalogram signal, electrocardiogram signal and respiratory activity signal as the original signal for anesthesia depth monitoring;

[0060] S20, dividing the signal into multiple sub-segments based on the average value of each preprocessed signal, marking the maximum value point and the minimum value point in each sub-segment, and determining the candidate interval of the grouping threshold according to the distance between the maximum value point and the minimum value point;

[0061] S30, dividing each preprocessed signal into a plurality of subsequences according to the determined grouping threshold, calculating the similarity between adjacent subsequences, and classifying the plurality of subsequences into stable subsequences and fluctuating subsequences based on the difference in similarity;

[0062] S40, based on the candidate interval of the grouping threshold determined in S20 and the classification results of the stable subsequence and the fluctuating subsequence in S30, performing differential processing on the anesthesia depth feature to determine a smaller step size used for the stable subsequence processing and a larger step size used for the fluctuating subsequence processing;

[0063] S50, filtering and denoising the collected original signal to obtain a preprocessed signal, and recording the abnormal mutation segments excluded in the filtering and denoising process to form an abnormal waveform segment set;

[0064] S60, according to the step size determined in S40, start the GPU parallel processing flow to process each pre-processed signal, and start the CPU to process the abnormal waveform segment set; perform short-time Fourier transform, Mel-frequency cepstrum coefficient analysis and wavelet packet decomposition on each pre-processed signal in the GPU, extract signal features, and obtain electroencephalogram features, electrocardiogram features, and respiratory activity features; use a sliding window method in the CPU to extract features from the abnormal waveform segment set using time-frequency analysis and statistical feature extraction methods, and record them as abnormal waveform segment features;

[0065] S70, performing cluster analysis on the electroencephalogram features, electrocardiogram features, and respiratory activity features respectively, obtaining the cluster center of each signal representing a typical anesthesia depth feature, and using the cluster center to replace the original features in the electroencephalogram features, electrocardiogram features, and respiratory activity features to form optimized electroencephalogram features, electrocardiogram features, and respiratory activity features;

[0066] S80, aligning and fusing the optimized electroencephalogram features, electrocardiogram features, respiratory activity features, and abnormal waveform segment features of each signal, using principal component analysis to reduce the dimension, obtaining the reduced dimension features, and performing standardization processing to finally obtain the anesthesia depth features;

[0067] S90: Use a pre-trained anesthesia depth monitoring model, input anesthesia depth characteristics, obtain the patient's anesthesia depth monitoring results and output them.

[0068] The specific implementation methods of the above steps are described in detail below:

[0069] Step S10: Continuously collect the patient's electroencephalogram (EEG) signal, electrocardiogram (ECG) signal and respiratory activity signal as the original signal for anesthesia depth monitoring.

[0070] Specifically, this step continuously captures the patient's EEG signals, ECG signals, and respiratory activity signals during anesthesia by wearing corresponding physiological signal acquisition equipment, such as EEG electrodes, ECG electrodes, and respiratory sensors. These three physiological signals reflect the changes in the patient's physiological state under anesthesia and are an important basis for assessing the depth of anesthesia. By continuously acquiring these original signals, the foundation is laid for subsequent signal preprocessing and feature extraction.

[0071] Step S20: Divide the signal into multiple sub-segments based on the average value of each preprocessed signal, mark the maximum value point and the minimum value point in each sub-segment, and determine the candidate interval of the grouping threshold according to the distance between the maximum value point and the minimum value point.

[0072] Specifically, the collected original signal is first preprocessed, such as denoising, filtering, etc., to obtain the preprocessed signal. Then, the average value of each preprocessed signal is used as a benchmark to divide it into multiple sub-segments. For each sub-segment, find the maximum and minimum points in it, and calculate the distance between the two points. According to the distance between these maximum and minimum points, the candidate interval of the threshold for subsequent grouping can be determined; usually, the interval between the maximum and minimum points is divided into multiple sub-intervals by equal division, and the segmentation point of each sub-interval is used as the grouping threshold, and each sub-interval is used as an alternative interval. The purpose of this step is to find a suitable grouping threshold to lay the foundation for subsequent signal classification.

[0073] Step S30: According to the determined grouping threshold, each preprocessed signal is divided into a plurality of groups of subsequences, similarities between adjacent subsequences are calculated, and the plurality of groups of subsequences are classified into stable subsequences and fluctuating subsequences based on the difference in similarities.

[0074] Specifically, first, each preprocessed signal is divided into several groups of subsequences according to the grouping threshold determined in step S20. Then, the similarity between adjacent subsequences is calculated, and the correlation coefficient or other similarity measurement methods can be used. Next, according to the difference in similarity between adjacent subsequences, these subsequences are classified into stable subsequences and fluctuating subsequences. A stable subsequence indicates that the signal is relatively stable during this period of time, and a fluctuating subsequence indicates that the signal has a large fluctuation during this period of time. The purpose of this step is to distinguish the stable and fluctuating features in the signal, and provide a basis for subsequent feature extraction and processing.

[0075] Step S40: Based on the alternative interval of the grouping threshold determined in S20 and the classification results of the stable subsequence and the fluctuating subsequence in S30, the anesthesia depth characteristics are differentially processed to determine the smaller step size used for the stable subsequence processing and the larger step size used for the fluctuating subsequence processing.

[0076] Specifically, according to the alternative interval of the grouping threshold determined in step S20, combined with the classification results of the stable subsequence and the fluctuating subsequence obtained in step S30, the anesthesia depth features to be extracted subsequently are differentially processed. For the stable subsequence, a smaller step size is used for feature extraction; for the fluctuating subsequence, a larger step size is used for feature extraction. For example, for the original signal with a sampling frequency of 100Hz, the step size of the stable subsequence can be set to 1 / 100 to 1 / 50 seconds, and the step size of the fluctuating subsequence can be set to 1 / 10 to 1 / 5 seconds. The purpose of this step is to use different feature extraction step sizes in stable and fluctuating states, so as to better capture the feature changes in the signal.

[0077] Step S50: Filter and denoise the collected original signal to obtain a preprocessed signal, and record the abnormal mutation segments excluded in the filtering and denoising process to form an abnormal waveform segment set.

[0078] Specifically, the original signal collected in step S10 is first filtered and denoised to remove some irrelevant interference components and obtain a preprocessed signal. In this process, some abnormal and sudden signal segments are eliminated, and these abnormal waveform segments are recorded to form an abnormal waveform segment set. These abnormal waveform segments may contain unique information related to the depth of anesthesia and need to be processed separately. Through this step, a relatively smooth preprocessed signal is obtained, and abnormal waveform segments that may contain valuable information are retained.

[0079] Step S60: according to the step size determined in S40, start the GPU parallel processing flow to process each preprocessed signal, and start the CPU to process the abnormal waveform segment set; in the GPU, perform short-time Fourier transform, Mel-frequency cepstrum coefficient analysis and wavelet packet decomposition on each preprocessed signal to extract signal features to obtain electroencephalogram features, electrocardiogram features, and respiratory activity features; in the CPU, use a sliding window method to extract features from the abnormal waveform segment set using time-frequency analysis and statistical feature extraction methods, and record them as abnormal waveform segment features.

[0080] Specifically, first, according to the different step sizes determined in step S40, the GPU and CPU are started for parallel processing. The GPU processing flow is used to process each preprocessed signal, and the features representing the electroencephalogram, electrocardiogram and respiratory activity are extracted by using methods such as short-time Fourier transform, Mel-frequency cepstral coefficient analysis and wavelet packet decomposition. The CPU processing flow is used to process the abnormal waveform segment set recorded in step S50, and the features of the abnormal waveform segment are extracted by using methods such as sliding window, time-frequency analysis and statistical feature extraction. The purpose of this step is to make full use of parallel computing resources and efficiently extract various representative features from the original signal.

[0081] Step S70: Perform cluster analysis on the EEG features, ECG features, and respiratory activity features respectively to obtain the cluster center of each signal representing a typical anesthesia depth feature, and use the cluster center to replace the original features in the EEG features, ECG features, and respiratory activity features to form optimized EEG features, ECG features, and respiratory activity features.

[0082] Specifically, the EEG features, ECG features, and respiratory activity features extracted in step S60 are clustered and analyzed respectively, for example, a clustering algorithm such as K-means or Gaussian mixture model can be used. Through clustering, the feature cluster center representing the typical anesthesia depth state in each signal can be found. Then, these cluster centers are used to replace the original features to form an optimized feature set. The purpose of this step is to remove some redundant or noise features through cluster analysis and extract more representative anesthesia depth features.

[0083] Step S80: Align and fuse the optimized EEG features, ECG features, respiratory activity features and abnormal waveform segment features of each signal, use principal component analysis to reduce the dimension, obtain the reduced dimension features, and perform standardization to finally obtain the anesthesia depth features.

[0084] Specifically, the optimized EEG features, ECG features, and respiratory activity features obtained in step S70, as well as the abnormal waveform segment features extracted in step S60, are first aligned and fused to form a comprehensive feature set. Since the feature dimension may be high, in order to reduce the data dimension, principal component analysis (PCA) is used to reduce the dimension and obtain the reduced-dimensional features. Finally, these reduced-dimensional features are standardized to obtain the final anesthesia depth features. The purpose of this step is to integrate the features of different signal sources and further enhance the representativeness and robustness of the features through dimensionality reduction and standardization.

[0085] Step S90: using a pre-trained anesthesia depth monitoring model, inputting the anesthesia depth feature, obtaining the patient's anesthesia depth monitoring result and outputting it.

[0086] In this step, a pre-trained anesthesia depth monitoring model is first required. The model adopts a multi-branch parallel convolutional neural network structure, including an electroencephalogram branch network, an electrocardiogram branch network, a respiratory activity branch network, an abnormal waveform branch network, and a fusion branch network. The input is the anesthesia depth feature obtained in step S80, and after processing and fusion of these branch networks, the patient's anesthesia depth monitoring result is finally output. The purpose of this step is to use the pre-trained deep learning model to quickly and accurately predict the patient's anesthesia depth state based on the extracted anesthesia depth features.

[0087] The training process of the entire method is as follows:

[0088] 1. Collect a large number of patients' anesthesia process data, including electroencephalogram, electrocardiogram, respiratory activity signals and corresponding anesthesia depth scores, as training samples.

[0089] 2. According to the process of steps S10-S80, feature extraction and preprocessing are performed on the training samples.

[0090] 3. Pair the extracted features with the corresponding anesthesia depth scores to form training samples.

[0091] 4. Perform data enhancement on training samples, such as adding noise, time distortion, etc., to enhance the generalization ability of the model.

[0092] 5. Divide the processed training samples into training set, validation set and test set.

[0093] 6. Initialize the parameters of the multi-branch network, use the small batch gradient descent method, and use the mean square error as the loss function to start training the model.

[0094] 7. Evaluate the model performance on the validation set and save the model if the performance improves. Repeat training and validation until the preset number of iterations is reached or the model converges.

[0095] 8. Finally, evaluate the performance of the final model on the test set.

[0096] In general, this method can effectively monitor the depth of anesthesia of patients through multi-step signal preprocessing, feature extraction and deep learning model training. The key is to make full use of the characteristics of different signal sources, adopt differential processing and parallel computing, and improve the representativeness of features and the prediction accuracy of the model.

[0097] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned anesthesia depth monitoring method.

[0098] A third aspect of the present invention provides an electronic device, which comprises a processor and a memory, wherein the memory is used to store the step program of the above-mentioned method, and the processor reads the memory and executes the steps.

[0099] Specifically, the principle of the present invention is to make full use of the information of various physiological signals, extract the key features representing the depth of anesthesia, and train an intelligent prediction model to achieve accurate monitoring of the patient's anesthesia depth. The specific technical principles are as follows:

[0100] 1. Multimodal signal acquisition and preprocessing: The present invention first continuously acquires the electroencephalogram (EEG) signal, electrocardiogram (ECG) signal and respiratory activity signal of the patient during anesthesia by wearing the corresponding physiological signal acquisition equipment. These three physiological signals reflect the changes in different physiological characteristics of the patient under anesthesia, and can provide a comprehensive data basis for subsequent feature extraction and model training. On the basis of acquiring the original signals, the present invention also filters and de-noises these signals to remove irrelevant interference components and obtain high-quality preprocessed signals.

[0101] 2. Differential feature extraction and abnormal waveform processing: In view of the stability and volatility of different signal features, the present invention adopts the differential processing method. For stable signal sub-segments, a smaller step size is used to extract features to capture subtle changes; for fluctuating signal sub-segments, a larger step size is used to extract features to cope with drastic changes. At the same time, the present invention also pays special attention to some abnormal signal waveforms, and extracts the unique features of these abnormal waveforms through methods such as sliding windows and time-frequency analysis to enhance the monitoring system's adaptability to complex situations.

[0102] 3. Multi-branch parallel neural network model: In order to make full use of the characteristic information contained in different signal sources, the present invention adopts a multi-branch parallel convolutional neural network as an anesthesia depth prediction model. The model includes an electroencephalogram branch network, an electrocardiogram branch network, a respiratory activity branch network, an abnormal waveform branch network, and a final fusion branch network. Each branch network is specifically responsible for extracting and learning the deep-level features of a certain physiological signal, and then these feature vectors are input into the fusion branch network for comprehensive prediction. This multi-branch parallel structure can make full use of the complementary information of different signal sources and improve the overall prediction performance of the model.

[0103] 4. Cluster optimization and feature fusion: In the process of feature extraction, the present invention uses the cluster analysis method to optimize various physiological signal features. Through clustering, the feature cluster centers representing the typical anesthesia depth state can be found, and these cluster centers can be used to replace the original features, remove some redundant or noise features, and improve the representativeness of the features. Finally, the optimized various physiological signal features and abnormal waveform features are aligned and fused, and then the dimension is reduced through principal component analysis to obtain the final anesthesia depth feature.

[0104] In summary, the anesthesia depth monitoring method of the present invention makes full use of the information of various physiological signals and adopts innovative technologies such as differential feature extraction, abnormal waveform processing, and multi-branch neural network modeling, which can monitor the patient's anesthesia depth state more accurately and robustly.

[0105] Since the present invention involves a large number of calculation processes, in order to better understand and implement the present invention, a specific embodiment of the present invention including a detailed calculation process is provided below. The steps of this embodiment are specifically described as follows:

[0106] Step S10: Continuously collect the patient's electroencephalogram (EEG) signal, electrocardiogram (ECG) signal and respiratory activity signal as the original signal for anesthesia depth monitoring.

[0107] In this step, the patient's electroencephalogram (EEG) signal during anesthesia is continuously captured by wearing the corresponding physiological signal acquisition equipment. EEG (t), ECG signal X ECG (t) and the respiratory activity signal X Resp (t). These three physiological signals all reflect the changes in the patient's physiological state under anesthesia and are important bases for evaluating the depth of anesthesia. Among them, t represents the time variable.

[0108] Step S20: Divide the signal into multiple sub-segments based on the average value of each preprocessed signal, mark the maximum value point and the minimum value point in each sub-segment, and determine the candidate interval of the grouping threshold according to the distance between the maximum value point and the minimum value point.

[0109] First, the collected original signal is preprocessed, such as denoising, filtering, etc., to obtain the preprocessed signal and in, and Respectively represent the preprocessed EEG signal, ECG signal and respiratory activity signal. Then, the average value μ of each preprocessed signal is EEG , μ ECG and μ Resp As a benchmark, divide the signal into multiple sub-segments. For each sub-segment, find the maximum value point x max and the minimum point x min , calculate the distance between these two points d = |x max -x min According to the distance d between these maximum and minimum points, the candidate interval of the threshold for subsequent grouping can be determined [d min ,d max ].

[0110] Step S30: According to the determined grouping threshold, each preprocessed signal is divided into a plurality of groups of subsequences, similarities between adjacent subsequences are calculated, and the plurality of groups of subsequences are classified into stable subsequences and fluctuating subsequences based on the difference in similarities.

[0111] First, according to the grouping threshold [d min ,d max ], each preprocessed signal and Divide into several groups of subsequences and Where i represents the index of the subsequence. Then, the similarity S between adjacent subsequences is calculated. i,i+1 , you can use correlation coefficients or other similarity measures, such as Next, according to the similarity S between adjacent subsequences i,i+1 , these subsequences are classified as stable subsequences And the wave subsequence and where j and k represent the indices of the stable and volatile subsequences, respectively.

[0112] Step S40: Based on the alternative interval of the grouping threshold determined in S20 and the classification results of the stable subsequence and the fluctuating subsequence in S30, the anesthesia depth characteristics are differentially processed to determine the smaller step size used for the stable subsequence processing and the larger step size used for the fluctuating subsequence processing.

[0113] According to the grouping threshold candidate interval [d min ,dmax ], combined with the classification results of the stable subsequence and the fluctuating subsequence obtained in step S30, the anesthesia depth feature to be extracted later is subjected to differential processing. and Use a smaller step size Seconds to extract features; for the fluctuation subsequence and Use a larger step size The purpose of this is to use different feature extraction steps in stable and fluctuating states to better capture the feature changes in the signal.

[0114] Step S50: Filter and denoise the collected original signal to obtain a preprocessed signal, and record the abnormal mutation segments excluded in the filtering and denoising process to form an abnormal waveform segment set.

[0115] First, the original signal X collected in step S10 EEG (t), X ECG (t) and X Resp (t) Perform filtering and noise reduction to remove some irrelevant interference components and obtain the preprocessed signal and In this process, some abnormal signal segments will be eliminated and recorded to form an abnormal waveform segment set. Where l represents the index of the abnormal waveform segment. Through this step, a relatively smooth preprocessed signal is obtained, and the abnormal waveform segments that may contain valuable information are retained.

[0116] Step S60: according to the step size determined in S40, start the GPU parallel processing flow to process each preprocessed signal, and start the CPU to process the abnormal waveform segment set; in the GPU, perform short-time Fourier transform, Mel-frequency cepstrum coefficient analysis and wavelet packet decomposition on each preprocessed signal to extract signal features to obtain electroencephalogram features, electrocardiogram features, and respiratory activity features; in the CPU, use a sliding window method to extract features from the abnormal waveform segment set using time-frequency analysis and statistical feature extraction methods, and record them as abnormal waveform segment features.

[0117] First, according to the different step lengths Δt determined in step S40 s and Δt v , respectively start the GPU and CPU for parallel processing. The GPU processing flow is used to process each preprocessed signal and The short-time Fourier transform (STFT), Mel-frequency cepstral coefficient (MFCC) analysis and wavelet packet decomposition (WPD) were used to extract the representative EEG features F.EEG 、ECG characteristics F ECG and respiratory activity characteristics F Resp The CPU processing flow is used to process the abnormal waveform segment set recorded in step S50. The sliding window, time-frequency analysis and statistical feature extraction methods are used to extract the abnormal waveform segment features F. a .

[0118] Step S70: Perform cluster analysis on the EEG features, ECG features, and respiratory activity features respectively to obtain the cluster center of each signal representing a typical anesthesia depth feature, and use the cluster center to replace the original features in the EEG features, ECG features, and respiratory activity features to form optimized EEG features, ECG features, and respiratory activity features.

[0119] The EEG feature F extracted in step S60 EEG 、ECG characteristics F ECG and respiratory activity characteristics F Resp Cluster analysis can be performed separately, for example, K-means or Gaussian mixture model (GMM) and other clustering algorithms can be used. Through clustering, the characteristic cluster center representing the typical anesthesia depth state in each signal can be found. and Where p, q and r represent the index of the cluster center respectively. Then, these cluster centers are used to replace the original features to form the optimized EEG features. ECG characteristics and respiratory activity characteristics The purpose of this step is to remove some redundant or noisy features through cluster analysis and extract more representative anesthesia depth features.

[0120] Step S80: Align and fuse the optimized EEG features, ECG features, respiratory activity features and abnormal waveform segment features of each signal, use principal component analysis to reduce the dimension, obtain the reduced dimension features, and perform standardization to finally obtain the anesthesia depth features.

[0121] First, the optimized EEG feature obtained in step S70 is ECG characteristics and respiratory activity characteristics And the abnormal waveform segment feature F extracted in step S60 a Align and fuse to form a comprehensive feature set Since the feature dimension may be high, in order to reduce the data dimension, principal component analysis (PCA) is used to reduce the dimension and obtain the feature after dimension reduction. Finally, for these reduced features Standardization was performed to obtain the final anesthesia depth characteristics

[0122] Step S90: using a pre-trained anesthesia depth monitoring model, inputting the anesthesia depth feature, obtaining the patient's anesthesia depth monitoring result and outputting it.

[0123] In this step, a pre-trained anesthesia depth monitoring model is first required. The model adopts a multi-branch parallel convolutional neural network structure, including an electroencephalogram branch network, an electrocardiogram branch network, a respiratory activity branch network, an abnormal waveform branch network, and a fusion branch network. The input is the anesthesia depth feature obtained in step S80. After processing and fusion of these branch networks, the patient's anesthesia depth monitoring results are finally output. The purpose of this step is to use the pre-trained deep learning model to quickly and accurately predict the patient's anesthesia depth status based on the extracted anesthesia depth features.

[0124] The training process of the entire method is as follows:

[0125] 1. Collect a large number of patients' anesthesia process data, including EEG X EEG (t), electrocardiogram X ECG (t), respiratory activity signal X Resp (t) and the corresponding anesthesia depth score y, as training samples.

[0126] 2. According to the process of steps S10-S80, feature extraction and preprocessing are performed on the training samples to obtain the final anesthesia depth feature

[0127] 3. Extract the features Paired with the corresponding anesthesia depth score to form a training sample.

[0128] 4. Perform data enhancement on the training samples, such as adding noise, time distortion, etc., to enhance the generalization ability of the model. After data enhancement, the training samples can be expressed as Where N is the total number of training samples.

[0129] 5. Divide the processed training samples into training set, validation set and test set.

[0130] Next are the specific training steps of the model:

[0131] 6. Initialize the parameters of the multi-branch network, use the small batch gradient descent method, and use the mean square error is the loss function and starts training the model. Where B is the size of the mini-batch.

[0132] 7. Evaluate the model performance on the validation set and save the model parameters if the performance improves. Repeat training and validation until the preset number of iterations is reached or the model converges.

[0133] Specifically, the operations for each training batch are as follows:

[0134] (1) Randomly sample B samples from the training set

[0135] (2) Input these samples into each branch network to obtain the prediction results

[0136] (3) Calculate the mean square error between the predicted result and the actual anesthesia depth score

[0137] (4) Update the network parameters through back propagation according to the gradient descent method.

[0138] 8. Evaluate the performance of the final model on the test set.

[0139] The specific structure of the model is as follows:

[0140] EEG branch network: input is optimized EEG features A multi-layer one-dimensional convolutional neural network structure is used, and the output is the EEG feature vector h EEG .

[0141] ECG branch network: input is the optimized ECG features A multi-layer one-dimensional convolutional neural network structure is used, and the output is the electrocardiogram feature vector h ECG .

[0142] Respiratory activity branch network: input is the optimized respiratory activity feature A multi-layer one-dimensional convolutional neural network structure is used, and the output is the respiratory activity feature vector h Resp .

[0143] Abnormal waveform branch network: the input is the abnormal waveform segment feature F a , a multi-layer fully connected neural network structure is used, and the output is the abnormal waveform feature vector h a .

[0144] Fusion branch network: The input is the output feature vector {h EEG ,h ECG ,h Resp ,h a}, using a multi-layer fully connected neural network structure, the output is the final anesthesia depth prediction value

[0145] The entire training process aims to learn a deep learning model that can accurately predict the depth of anesthesia of patients. By making full use of the characteristics of different signal sources, using differential processing and parallel computing, the representativeness of the features and the prediction accuracy of the model are improved. Ultimately, the model can effectively monitor the depth of anesthesia of patients.

[0146] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for monitoring depth of anesthesia, characterized in that: The following steps are involved: S10, continuously collecting the patient's electroencephalogram signal, electrocardiogram signal and respiratory activity signal as the original signal for anesthesia depth monitoring; S20, dividing the signal into multiple sub-segments based on the average value of each preprocessed signal, marking the maximum value point and the minimum value point in each sub-segment, and determining the candidate interval of the grouping threshold according to the distance between the maximum value point and the minimum value point; S30, dividing each preprocessed signal into a plurality of subsequences according to the determined grouping threshold, calculating the similarity between adjacent subsequences, and classifying the plurality of subsequences into stable subsequences and fluctuating subsequences based on the difference in similarity; S40, based on the candidate interval of the grouping threshold determined in S20 and the classification results of the stable subsequence and the fluctuating subsequence in S30, performing differential processing on the anesthesia depth feature to determine a smaller step size used for the stable subsequence processing and a larger step size used for the fluctuating subsequence processing; S50, filtering and denoising the collected original signal to obtain a preprocessed signal, and recording the abnormal mutation segments excluded in the filtering and denoising process to form an abnormal waveform segment set; S60, according to the step size determined in S40, start the GPU parallel processing flow to process each pre-processed signal, and start the CPU to process the abnormal waveform segment set; perform short-time Fourier transform, Mel-frequency cepstrum coefficient analysis and wavelet packet decomposition on each pre-processed signal in the GPU, extract signal features, and obtain electroencephalogram features, electrocardiogram features, and respiratory activity features; use a sliding window method in the CPU to extract features from the abnormal waveform segment set using time-frequency analysis and statistical feature extraction methods, and record them as abnormal waveform segment features; S70, performing cluster analysis on the electroencephalogram features, electrocardiogram features, and respiratory activity features respectively, obtaining a cluster center representing a typical anesthesia depth feature for each signal, and replacing the original features in the electroencephalogram features, electrocardiogram features, and respiratory activity features with the cluster center to form optimized electroencephalogram features, electrocardiogram features, and respiratory activity features; S80, aligning and fusing the optimized electroencephalogram features, electrocardiogram features, respiratory activity features, and abnormal waveform segment features of each signal, using principal component analysis to reduce the dimension, obtaining the reduced dimension features, and performing standardization processing to finally obtain the anesthesia depth features; S90: Use a pre-trained anesthesia depth monitoring model, input the anesthesia depth feature, obtain the patient's anesthesia depth monitoring result and output it.

2. A method for monitoring depth of anesthesia according to claim 1, characterized in that: The larger step size is 0.1 to 0.2 seconds, and the smaller step size is 0.01 to 0.02 seconds.

3. The method for monitoring depth of anesthesia according to claim 1, characterized in that: In the step of calculating the similarity between adjacent subsequences, the similarity is calculated using correlation coefficient similarity.

4. The method for monitoring depth of anesthesia according to claim 1, characterized in that: The anesthesia depth monitoring model adopts a multi-branch parallel convolutional neural network structure, including an electroencephalogram branch network, an electrocardiogram branch network, a respiratory activity branch network, an abnormal waveform branch network, and a fusion branch network.

5. A method for monitoring depth of anesthesia according to claim 4, characterized in that: The EEG branch network is used to process EEG features, the input is the optimized EEG features, the output is the EEG feature vector, and the structure is a multi-layer one-dimensional convolutional neural network; The electrocardiogram branch network is used to process electrocardiogram features, the input is the optimized electrocardiogram features, the output is the electrocardiogram feature vector, and the structure is a multi-layer one-dimensional convolutional neural network; The respiratory activity branch network is used to process respiratory activity features, the input is the optimized respiratory activity features, the output is the respiratory activity feature vector, and the structure is a multi-layer one-dimensional convolutional neural network; The abnormal waveform branch network is used to process abnormal waveform segment features, the input is the abnormal waveform segment features, the output is the abnormal waveform feature vector, and the structure is a multi-layer fully connected neural network; The fusion branch network is used to integrate the characteristics of each branch, the input is the output feature vector of each branch network, the output is the final anesthesia depth prediction value, and the structure is a multi-layer fully connected neural network.

6. A method for monitoring depth of anesthesia according to claim 5, characterized in that: The steps of obtaining the training set for training the anesthesia depth monitoring model are specifically: Step 11, collecting a large number of patients' anesthesia process data, including electroencephalogram, electrocardiogram, respiratory activity signal and corresponding anesthesia depth score; Step 12: Process the original signal according to steps S10-S80 to extract features; Step 13: Pair the extracted features with the corresponding anesthesia depth scores to form training samples; Step 14: Perform data enhancement on the training samples, such as adding noise, time distortion, etc. Step 15: Divide the processed training samples into training set, validation set and test set.

7. A method for monitoring depth of anesthesia according to claim 6, characterized in that: The specific training steps of the anesthesia depth monitoring model are: Step 21, initializing the parameters of each branch network; Step 22: Use the mini-batch gradient descent method and the mean square error as the loss function to start training; Step 23, for each batch of data: input the data into each branch network to obtain the prediction result; calculate the error between the prediction result and the actual anesthesia depth score; back propagate the error and update the network parameters; Step 24: Evaluate the model performance on the validation set and save the model if the performance improves. Step 25, repeat steps 3-4 until a preset number of iterations is reached or the model converges, to obtain the anesthesia depth monitoring model.

8. The method for monitoring depth of anesthesia according to claim 1, characterized in that: In the steps of respectively performing cluster analysis on the electroencephalogram features, electrocardiogram features, and respiratory activity features, the clustering method adopted is K-means clustering.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the anesthesia depth monitoring method according to any one of claims 1 to 8.

10. An electronic device, characterized in that: A processor and a memory are provided, wherein the memory is used to store the step program of the method described in any one of claims 1 to 8, and the processor reads the memory and executes the steps.

Images