Anesthesia depth monitoring method and system

Abstract

The invention relates to the technical field of medical data analysis, in particular to an anesthesia depth monitoring method and system. The anesthesia depth monitoring system comprises a reference data acquisition module, a real-time data acquisition module and an anesthesia depth analysis module. According to the method, the acquired reference brain wave data, the acquired reference physiological signal data, the acquired real-time brain wave data and the acquired real-time physiological signal data are processed through the anesthesia depth analysis model to output the anesthesia depth value, and then an anesthetist is assisted to control the dosage of anesthetic through the anesthesia depth value; the brain wave differential feature vector and the physiological signal differential data are used for reflecting personalized features of the patient, higher accuracy of anesthesia depth analysis results can be obtained for different patients, a self-attention mechanism is performed through brain wave data changes of the patient, and personalized features of the patient are further enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data analysis, and in particular to an anesthesia depth monitoring method and system. Background Art

[0002] Accurate monitoring of anesthesia depth is an important part of ensuring patient safety during surgery. At present, the anesthesia depth monitoring method widely used in clinical practice is mainly based on the analysis of electroencephalogram (EEG) signals, such as calculating the bispectral index (BIS) of brain waves to assess the patient's anesthesia depth. The bispectral index is a single numerical indicator generated by comprehensive processing of multiple brain wave characteristics. It can reflect the patient's central nervous system activity level and provide anesthesiologists with a quantitative reference for the depth of anesthesia. However, this anesthesia depth assessment method based on a fixed algorithm has certain limitations.

[0003] Specifically, the calculation of the bispectral index usually relies on standardized EEG analysis models, which do not fully consider individual differences among patients, such as age, gender, weight, metabolic level, drug sensitivity, and pathophysiological status. This "one-size-fits-all" assessment method may lead to inaccurate assessment of the depth of anesthesia for some patients. For example, the same bispectral index value may correspond to different depths of anesthesia between different patients, and may even lead to the risk of too deep or too shallow anesthesia. Summary of the invention

[0004] The present invention processes the acquired baseline brain wave data, baseline physiological signal data, real-time brain wave data and real-time physiological signal data through an anesthesia depth analysis model to output an anesthesia depth value, and then assists the anesthesiologist in controlling the dosage of anesthetic drugs through the anesthesia depth value. In the anesthesia depth analysis model, brain wave differential feature vectors and physiological signal differential data are used to reflect the personalized characteristics of the patient, so that a higher accuracy of the anesthesia depth analysis result can be obtained when facing different patients, and a self-attention mechanism is performed through changes in the patient's brain wave data to further enhance the patient's personalized characteristics. The second feature extraction layer is also updated through changes in the patient's brain wave data and physiological signal data, so that the extracted anesthesia depth analysis vector is more in line with the patient's brain wave data changes and physiological signal data changes.

[0005] The present invention provides a method for monitoring the depth of anesthesia, comprising: Obtaining brain wave data of the patient in a waking state and recording it as baseline brain wave data, obtaining physiological signal data of the patient in a waking state and recording it as baseline physiological signal data; At the monitoring time point, the patient's brain wave data and physiological signal data during the anesthesia process are obtained, and recorded as real-time brain wave data and real-time physiological signal data, respectively. The real-time brain wave data and real-time physiological signal data obtained at the current monitoring time point are stored in a monitoring database, and the baseline brain wave data, baseline physiological signal data, real-time brain wave data and real-time physiological signal data are sent to the anesthesia depth analysis model for processing, and the anesthesia depth value is output; The anesthesia depth analysis model includes a first feature extraction layer, a feature enhancement layer, a feature splicing layer, a second feature extraction layer and an anesthesia depth value output layer, wherein the first feature extraction layer is used to perform feature extraction on the baseline EEG data and the real-time EEG data to construct the corresponding baseline EEG feature vector and the real-time EEG feature vector, and the feature extraction specifically includes power spectral density extraction and entropy feature extraction; the feature enhancement layer is used to perform differential processing on the baseline EEG feature vector and the real-time EEG feature vector to construct the EEG differential feature vector, and perform self-attention mechanism operation on the EEG differential feature vector based on the monitoring database to construct the enhanced brain The electric wave differential feature vector; the feature splicing layer is used to perform differential operations on the baseline physiological signal data and the real-time physiological signal data to construct the physiological signal differential data, and then splice the enhanced brain wave differential feature vector with the physiological signal differential data to construct the anesthesia depth analysis data; the second feature extraction layer is established based on the restricted Boltzmann machine, and is used to process the anesthesia depth analysis data to construct the anesthesia depth analysis vector, and the second feature extraction layer is updated in real time based on the monitoring database at the model update time point; the anesthesia depth value output layer is used to process the anesthesia depth analysis vector, which can be specifically a fully connected operation to output the anesthesia depth value.

[0006] Preferably, a self-attention mechanism operation is performed on the brain wave differential feature vector based on the monitoring database to construct an enhanced brain wave differential feature vector, which specifically includes the following steps: Acquire the real-time brain wave data corresponding to the N-1 monitoring time points before the current monitoring time point from the monitoring database, and process the real-time brain wave data corresponding to the N-1 monitoring time points before the current monitoring time point through the first feature extraction layer to obtain the corresponding real-time brain wave feature vector, and form a brain wave timing analysis data set with the real-time brain wave feature vector corresponding to the N-1 monitoring time points before the current monitoring time point and the real-time brain wave feature vector corresponding to the current monitoring time point, and then send the brain wave timing analysis data set to the timing analysis network for processing to construct a brain wave change feature vector; The brain wave difference feature vector is matrix multiplied with the value weight matrix and the key weight matrix to obtain the brain wave value vector V and the brain wave key vector K. Then the brain wave change feature vector is matrix multiplied with the query weight matrix to obtain the brain wave query vector Q. The self-attention mechanism operation is implemented on the brain wave difference feature vector through the following formula: , where H is the enhanced EEG differential feature vector, softmax() is the activation function, and D is the dimension of the EEG key vector K.

[0007] Preferably, at the model update time point, the second feature extraction layer is updated in real time based on the monitoring database, and the specific steps are as follows: At the model update time point, the real-time brain wave data corresponding to all monitoring time points in the monitoring database are processed by the first feature extraction layer to obtain the corresponding real-time brain wave feature vectors. For any two adjacent monitoring time points, the real-time brain wave feature vectors corresponding to the two adjacent monitoring time points are subjected to a differential operation to obtain a training brain wave differential feature vector. The real-time physiological signal data corresponding to the two adjacent monitoring time points are subjected to a differential operation to obtain training physiological signal differential data. The training brain wave differential feature vectors and the training physiological signal differential data corresponding to the two adjacent monitoring time points are then subjected to a feature extraction real-time training sample. All feature extraction real-time training samples are combined into a feature extraction real-time training set. A number of monitoring time points are included between two adjacent model update time points. Performing unsupervised training on the second feature extraction layer through the real-time training set to obtain a trained second feature extraction layer; The second feature extraction layer is updated in real time using the following formula: i =1+βε i , where θ i is the parameter in the second feature extraction layer, i=1,2,3,…,I, I is the total number of parameters in the second feature extraction layer, β is the learning rate of the second feature extraction layer, ε i is the parameter θ in the trained second feature extraction layer i The corresponding parameters.

[0008] Preferably, the second feature extraction layer is trained unsupervisedly using a real-time training set, specifically in the following manner: The second feature extraction layer includes a visible layer and a hidden layer. The feature extraction real-time training sample is sent to the visible layer. The activation probability of the hidden layer is calculated through the feature extraction weight matrix and the feature extraction bias value between the visible layer and the hidden layer. Then, the first hidden layer state data in the hidden layer is obtained based on the activation probability sampling of the hidden layer. The first hidden layer state data in the hidden layer is back-propagated to the visible layer through the feature extraction weight matrix and the feature extraction bias value between the visible layer and the hidden layer, and the visible layer state data of the visible layer is reconstructed. Then, the visible layer state data of the visible layer is transmitted through the visible layer and the hidden layer. The feature extraction weight matrix and feature extraction bias value between the layers are forward propagated to the hidden layer to obtain the second hidden layer state data in the hidden layer. Finally, the contrast divergence processing is performed through the feature extraction real-time training samples, the first hidden layer state data, the visible layer state data and the second hidden layer state data to maximize the logarithmic likelihood value of the feature extraction real-time training samples to adjust the parameters of the feature extraction weight matrix and feature extraction bias value between the visible layer and the hidden layer; and the training goal is that the number of training times reaches the maximum number of training times. When the training goal is achieved, the trained second feature extraction layer is output.

[0009] Preferably, the timing analysis network is established based on the LSTM model, including an LSTM layer and a fully connected layer, wherein the LSTM layer is used to perform timing analysis on the EEG timing analysis data set, and the fully connected layer is used to perform a fully connected operation on the output of the LSTM layer to obtain a brain wave change feature vector.

[0010] Preferably, training the anesthesia depth analysis model specifically includes the following steps: A number of anesthesia depth analysis training samples are obtained, and the anesthesia depth analysis training samples include real-time EEG data and real-time physiological signal data obtained from the same monitoring time point and the same patient, and a one-to-one correspondence is established between the patient's corresponding baseline EEG data, baseline physiological signal data and monitoring database and the anesthesia depth analysis training samples, and the anesthesia depth analysis training samples are labeled by anesthesia depth values, and all labeled anesthesia depth analysis training samples are formed into an anesthesia depth analysis training set, and an anesthesia depth analysis model is trained by the anesthesia depth analysis training set to determine whether the training conditions are met. If the training conditions are met, the trained anesthesia depth analysis model is output; otherwise, the anesthesia depth analysis model continues to be trained by the anesthesia depth analysis training set.

[0011] The present invention also provides an anesthesia depth monitoring system, comprising: A reference data acquisition module is used to obtain the brain wave data of the patient in the awake state and record it as the reference brain wave data, and obtain the physiological signal data of the patient in the awake state and record it as the reference physiological signal data; The real-time data acquisition module is used to obtain the brain wave data and physiological signal data of the patient during the anesthesia process at the monitoring time point, and record them as real-time brain wave data and real-time physiological signal data respectively, and store the real-time brain wave data and real-time physiological signal data obtained at the current monitoring time point in the monitoring database; An anesthesia depth analysis module is used to send the baseline brain wave data, the baseline physiological signal data, the real-time brain wave data and the real-time physiological signal data to the anesthesia depth analysis model for processing and output the anesthesia depth value; The anesthesia depth analysis model includes a first feature extraction layer, a feature enhancement layer, a feature splicing layer, a second feature extraction layer and an anesthesia depth value output layer, wherein the first feature extraction layer is used to perform feature extraction on the baseline EEG data and the real-time EEG data to construct the corresponding baseline EEG feature vector and the real-time EEG feature vector, and the feature extraction specifically includes power spectral density extraction and entropy feature extraction; the feature enhancement layer is used to perform differential processing on the baseline EEG feature vector and the real-time EEG feature vector to construct the EEG differential feature vector, and perform self-attention mechanism operation on the EEG differential feature vector based on the monitoring database to construct the enhanced brain The electric wave differential feature vector; the feature splicing layer is used to perform differential operations on the baseline physiological signal data and the real-time physiological signal data to construct the physiological signal differential data, and then splice the enhanced brain wave differential feature vector with the physiological signal differential data to construct the anesthesia depth analysis data; the second feature extraction layer is established based on the restricted Boltzmann machine, and is used to process the anesthesia depth analysis data to construct the anesthesia depth analysis vector, and the second feature extraction layer is updated in real time based on the monitoring database at the model update time point; the anesthesia depth value output layer is used to process the anesthesia depth analysis vector, which can be specifically a fully connected operation to output the anesthesia depth value.

[0012] The present invention has the following advantages: The present invention processes the acquired baseline brain wave data, baseline physiological signal data, real-time brain wave data and real-time physiological signal data through an anesthesia depth analysis model to output an anesthesia depth value, and then assists the anesthesiologist in controlling the dosage of anesthetic drugs through the anesthesia depth value. In the anesthesia depth analysis model, brain wave differential feature vectors and physiological signal differential data are used to reflect the personalized characteristics of the patient, so that a higher accuracy of the anesthesia depth analysis result can be obtained when facing different patients, and a self-attention mechanism is performed through changes in the patient's brain wave data to further enhance the patient's personalized characteristics. The second feature extraction layer is also updated through changes in the patient's brain wave data and physiological signal data, so that the extracted anesthesia depth analysis vector is more in line with the patient's brain wave data changes and physiological signal data changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1It is a schematic diagram of the structure of the anesthesia depth analysis model adopted in an embodiment of the present invention.

[0014] Figure 2 It is a schematic diagram of the structure of the anesthesia depth monitoring system adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions 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.

[0016] Embodiment 1, a method for monitoring depth of anesthesia, comprising: Obtain the patient's brain wave data when awake and record it as the baseline brain wave data. Obtain the patient's physiological signal data when awake and record it as the baseline physiological signal data. It should be noted that the awake state here refers to the patient before surgery and when the patient's cognition is clear. In the process of obtaining brain wave data, operations such as denoising, data segmentation and standardization are generally included to improve the signal quality of brain wave data. With the baseline brain wave data as a reference, the changes in the patient's brain waves during anesthesia can be analyzed. The physiological signal data may include heart rate, blood pressure, blood oxygen saturation and respiratory rate, etc.; At the monitoring time point, the brain wave data and physiological signal data of the patient during the anesthesia process are obtained, and recorded as real-time brain wave data and real-time physiological signal data respectively. The real-time brain wave data here can reflect the changes in the electrical signals of the cerebral cortex of the patient during the anesthesia process, and the real-time physiological signal data can reflect the changes in the physiological signal data of the patient during the anesthesia process. The real-time brain wave data and real-time physiological signal data obtained at the current monitoring time point are stored in the monitoring database, and the baseline brain wave data, baseline physiological signal data, real-time brain wave data and real-time physiological signal data are sent to the anesthesia depth analysis model for processing, and the anesthesia depth value is output. The anesthesia depth value is generally a value between 0-100, and 80-100 is a wakeful state, 60-80 is mild sedation, 40-60 is an appropriate anesthesia depth, and 0-40 is deep anesthesia or excessive anesthesia. The anesthesia depth value can reflect the patient's anesthetic state during the anesthesia process to assist the anesthesiologist in controlling the dosage of anesthetic drugs; the time interval between adjacent monitoring time points is generally 30s; See also Figure 1The anesthesia depth analysis model includes a first feature extraction layer, a feature enhancement layer, a feature splicing layer, a second feature extraction layer and an anesthesia depth value output layer, wherein the first feature extraction layer is used to extract features from the baseline EEG data and the real-time EEG data to construct the corresponding baseline EEG feature vector and the real-time EEG feature vector. The feature extraction specifically includes power spectrum density extraction and entropy feature extraction, which can be regarded as a preprocessing operation for the baseline EEG data and the real-time EEG data. The power spectrum density extraction is specifically as follows: after the EEG data is subjected to wavelet transformation, the power extraction of δ, θ, α, β, and γ waves is performed. The entropy features include approximate entropy, sample entropy, sorting entropy and wavelet entropy. The specific calculation of approximate entropy, sample entropy, sorting entropy and wavelet entropy can be directly adopted. The yEntropy library is used to implement it, and because the baseline EEG feature vectors are consistent, the baseline EEG feature vectors can be temporarily stored without repeated feature extraction; the feature enhancement layer is used to perform differential processing on the baseline EEG feature vectors and the real-time EEG feature vectors to construct EEG differential feature vectors, and perform self-attention mechanism operations on the EEG differential feature vectors based on the monitoring database to construct enhanced EEG differential feature vectors. It should be noted that the EEG differential feature vectors can reflect the degree of deviation of the currently acquired EEG data from the baseline EEG data. Compared with direct EEG data, they can better reflect the patient's personalized factors (such as age, gender, weight, metabolic level, drug sensitivity, and pathophysiological status). The change characteristics of EEG data caused by factors such as anxiety, anxiety, and stress (such as anxiety, anxiety, and stress) are used to perform subsequent anesthesia depth analysis through EEG differential feature vectors, which can obtain more accurate anesthesia depth analysis results based on the patient's own personalized factors, thereby making the control during anesthesia more accurate. By performing self-attention mechanism operations on EEG differential feature vectors through the monitoring database, it is possible to analyze the real-time changes in the patient's EEG data. For example, elderly patients may need to pay more attention to the changes in low-frequency delta waves, while young patients may need to pay more attention to the changes in high-frequency beta waves. Based on the real-time changes in the patient's EEG data, the self-attention mechanism operation on the EEG differential feature vector can make the anesthesia depth analysis model pay more attention to the personalization of the patient's EEG data. changes, further improving the accuracy of anesthesia depth analysis results; the feature splicing layer is used to perform differential operations on the baseline physiological signal data and the real-time physiological signal data to construct physiological signal differential data. Similarly, the physiological signal differential data can better reflect the patient's personalized physiological signal data, and then the enhanced EEG differential feature vector is spliced ​​with the physiological signal differential data to construct anesthesia depth analysis data; the second feature extraction layer is established based on the restricted Boltzmann machine, which is used to process the anesthesia depth analysis data to construct anesthesia depth analysis vector, and the second feature extraction layer is updated in real time based on the monitoring database at the model update time point, so that the extracted anesthesia depth analysis vector is more in line with the patient's EEG data changes and physiological signal data changes;The anesthesia depth value output layer is used to process the anesthesia depth analysis vector, which can be a fully connected operation to output the anesthesia depth value; The present application processes the acquired baseline EEG data, baseline physiological signal data, real-time EEG data and real-time physiological signal data through an anesthesia depth analysis model to output an anesthesia depth value, and then assists the anesthesiologist in controlling the dosage of anesthetic drugs through the anesthesia depth value. In the anesthesia depth analysis model, the EEG differential feature vector and the physiological signal differential data are used to reflect the personalized characteristics of the patient, so that a higher accuracy of the anesthesia depth analysis results can be obtained when facing different patients, and a self-attention mechanism is performed through the changes in the patient's brain wave data to further enhance the patient's personalized characteristics. The second feature extraction layer is also updated through the changes in the patient's brain wave data and the physiological signal, so that the extracted anesthesia depth analysis vector is more in line with the patient's brain wave data changes and the physiological signal data changes.

[0017] Based on the monitoring database, a self-attention mechanism operation is performed on the EEG differential feature vector to construct an enhanced EEG differential feature vector, which specifically includes the following steps: The real-time EEG data corresponding to the N-1 monitoring time points before the current monitoring time point are obtained from the monitoring database, and the real-time EEG data corresponding to the N-1 monitoring time points before the current monitoring time point are respectively processed by the first feature extraction layer to obtain the corresponding real-time EEG feature vector. It should be noted that in the actual use process, the real-time EEG data in the monitoring database gradually increases and does not reach N at the beginning. Therefore, the feature enhancement layer is disabled at the beginning. Until the real-time EEG data in the monitoring database reaches N, the feature enhancement layer starts to be enabled, and the real-time EEG feature vector corresponding to the N-1 monitoring time points before the current monitoring time point is compared with the current The real-time EEG feature vectors corresponding to the monitoring time points constitute the EEG timing analysis data set, and then the EEG timing analysis data set is sent to the timing analysis network for processing to construct the EEG change feature vector. The timing analysis network is established based on the LSTM model, including the LSTM layer and the fully connected layer, wherein the LSTM layer is used to perform timing analysis on the EEG timing analysis data set, and the fully connected layer is used to perform a fully connected operation on the output of the LSTM layer to obtain the EEG change feature vector. The timing analysis network established based on the LSTM model can analyze the changes in the patient's EEG data, and the parameters of the timing analysis network are adjusted following the training of the anesthesia depth analysis model; The brain wave difference feature vector is matrix multiplied with the value weight matrix and the key weight matrix to obtain the brain wave value vector V and the brain wave key vector K. Then the brain wave change feature vector is matrix multiplied with the query weight matrix to obtain the brain wave query vector Q. The self-attention mechanism operation is implemented on the brain wave difference feature vector through the following formula: , where H is the enhanced EEG differential feature vector, softmax() is the activation function, and D is the dimension of the EEG key vector K; it should be noted that the value weight matrix, key weight matrix, and query weight matrix are set with reference to the self-attention mechanism layer in the Transformer model, and the parameters of the value weight matrix, key weight matrix, and query weight matrix are adjusted following the training of the anesthesia depth analysis model; At the model update time point, the second feature extraction layer is updated in real time based on the monitoring database. The specific steps are as follows: At the model update time point, the real-time brain wave data corresponding to all monitoring time points in the monitoring database are processed by the first feature extraction layer to obtain the corresponding real-time brain wave feature vectors. For any two adjacent monitoring time points, the real-time brain wave feature vectors corresponding to the two adjacent monitoring time points are subjected to a differential operation to obtain a training brain wave differential feature vector. The real-time physiological signal data corresponding to the two adjacent monitoring time points are subjected to a differential operation to obtain training physiological signal differential data. Then, the training brain wave differential feature vectors and the training physiological signal differential data corresponding to the two adjacent monitoring time points are subjected to a feature extraction real-time training sample. All feature extraction real-time training samples are combined into a feature extraction real-time training set. A number of monitoring time points are included between two adjacent model update time points, generally including 10 monitoring time points. The second feature extraction layer is trained unsupervisedly through the real-time training set. Specifically, the second feature extraction layer includes a visible layer and a hidden layer, and the feature extraction real-time training sample is sent to the visible layer. The activation probability of the hidden layer is calculated through the feature extraction weight matrix and the feature extraction bias value between the visible layer and the hidden layer, and then the first hidden layer state data in the hidden layer is obtained based on the activation probability sampling of the hidden layer. The first hidden layer state data in the hidden layer is back-propagated to the visible layer through the feature extraction weight matrix and the feature extraction bias value between the visible layer and the hidden layer, and the visible layer state data of the visible layer is reconstructed. Then, the visible layer state data of the visible layer is forward-propagated to the hidden layer through the feature extraction weight matrix and the feature extraction bias value between the visible layer and the hidden layer. The state data of the second hidden layer in the hidden layer is obtained, and finally, contrast divergence processing is performed on the feature extraction real-time training samples, the first hidden layer state data, the visible layer state data and the second hidden layer state data, so as to maximize the logarithmic likelihood value of the feature extraction real-time training samples to adjust the parameters of the feature extraction weight matrix and the feature extraction bias value between the visible layer and the hidden layer; and the training goal is that the number of training times reaches the maximum number of training times, and in the process of unsupervised training of the second feature extraction layer, it can adapt to the distribution of changes in brain wave data and physiological signal data, so that the anesthesia depth analysis vector extracted by the second feature extraction layer pays more attention to the feature information of the patient's own adaptation that causes the change of anesthesia conditions, and when the training goal is completed, the trained second feature extraction layer is output; The second feature extraction layer is updated in real time using the following formula: i =1+βε i , where θ i is the parameter in the second feature extraction layer, i=1,2,3,…,I, I is the total number of parameters in the second feature extraction layer, β is the learning rate of the second feature extraction layer, generally 0.01, ε i is the parameter θ in the trained second feature extraction layer i The corresponding parameters here refer to the feature extraction weight matrix and feature extraction bias value of the second feature extraction layer.

[0018] Training the anesthesia depth analysis model includes the following steps: A number of anesthesia depth analysis training samples are obtained, and the anesthesia depth analysis training samples include real-time EEG data and real-time physiological signal data obtained from the same monitoring time point and the same patient, and a one-to-one corresponding link is established with the anesthesia depth analysis training samples through the patient's corresponding baseline EEG data, baseline physiological signal data and monitoring database. It should be noted that the real-time EEG data and real-time physiological signal data here are obtained during the patient's actual anesthesia process, and the data source is guaranteed to be consistent. The monitoring database is a number of real-time EEG data and real-time physiological signal data obtained by the patient corresponding to the anesthesia depth analysis training samples. The anesthesia depth analysis training samples are annotated with anesthesia depth values, and the anesthesia depth analysis training samples are manually annotated by experienced anesthesiologists based on the patient's EEG data, the patient's physiological signal data and the patient's performance. All annotated anesthesia depth analysis training samples form an anesthesia depth analysis training set, and the anesthesia depth analysis model is trained with the anesthesia depth analysis training set to determine whether the training conditions are met. The training condition may be that the accuracy of the anesthesia depth analysis model reaches a threshold set by the developer. If the training conditions are met, the trained anesthesia depth analysis model is output; otherwise, the anesthesia depth analysis model is continuously trained with the anesthesia depth analysis training set.

[0019] Embodiment 2, an anesthesia depth monitoring system, such as Figure 2 As shown, including: The reference data acquisition module is used to obtain the brain wave data of the patient in the awake state and record it as the reference brain wave data, and obtain the physiological signal data of the patient in the awake state and record it as the reference physiological signal data. It should be noted that the awake state here refers to the patient before the operation and when the cognition is clear. In the process of obtaining the brain wave data, it generally includes denoising, data segmentation and standardization operations to improve the signal quality of the brain wave data. With the reference brain wave data as a reference, the changes in the patient's brain waves during anesthesia can be analyzed, and the physiological signal data can include heart rate, blood pressure, blood oxygen saturation and respiratory rate, etc.; The real-time data acquisition module is used to obtain the brain wave data and physiological signal data of the patient during the anesthesia process at the monitoring time point, and record them as real-time brain wave data and real-time physiological signal data respectively. The real-time brain wave data here can reflect the changes in the electrical signals of the cerebral cortex of the patient during the anesthesia process, and the real-time physiological signal data can reflect the changes in the physiological signal data of the patient during the anesthesia process. The real-time brain wave data and real-time physiological signal data obtained at the current monitoring time point are stored in the monitoring database; The anesthesia depth analysis module is used to send the baseline EEG data, baseline physiological signal data, real-time EEG data and real-time physiological signal data to the anesthesia depth analysis model for processing and output the anesthesia depth value. The anesthesia depth value is generally a value between 0 and 100, and 80-100 is a wakeful state, 60-80 is mild sedation, 40-60 is a suitable anesthesia depth, and 0-40 is deep anesthesia or excessive anesthesia. The anesthesia depth value can reflect the patient's anesthetic state during anesthesia to assist the anesthesiologist in controlling the dosage of anesthetic drugs; the anesthesia depth analysis model includes a first feature extraction layer, a feature enhancement layer, a feature splicing layer, a second feature extraction layer and anesthesia depth value input layer. The first feature extraction layer is used to extract features from the baseline EEG data and the real-time EEG data to construct the corresponding baseline EEG feature vector and the real-time EEG feature vector. The feature extraction specifically includes power spectrum density extraction and entropy feature extraction, which can be regarded as the preprocessing operation of the baseline EEG data and the real-time EEG data. The power spectrum density extraction is specifically as follows: after the EEG data is subjected to wavelet transformation, the power extraction of δ, θ, α, β, and γ waves is performed. The entropy features include approximate entropy, sample entropy, sorting entropy, and wavelet entropy. The specific calculation of approximate entropy, sample entropy, sorting entropy, and wavelet entropy can be directly implemented using the pyEntropy library. Since the baseline EEG feature vector are consistent, the baseline EEG feature vector can be temporarily stored without repeated feature extraction; the feature enhancement layer is used to perform differential processing on the baseline EEG feature vector and the real-time EEG feature vector to construct an EEG differential feature vector, and perform a self-attention mechanism operation on the EEG differential feature vector based on the monitoring database to construct an enhanced EEG differential feature vector. It should be noted that the EEG differential feature vector can reflect the degree of deviation of the currently acquired EEG data from the baseline EEG data. Compared with direct EEG data, it can better reflect the patient's personalized EEG data characteristics, and perform subsequent anesthesia depth analysis through the EEG differential feature vector. A more accurate anesthesia depth analysis result can be obtained based on the individual characteristics of the patient, thereby making the control during anesthesia more accurate. By performing a self-attention mechanism operation on the EEG differential feature vector through the monitoring database, the real-time changes of the patient's EEG data can be analyzed. For example, elderly patients may need to pay more attention to the changes in low-frequency delta waves, while young patients may need to pay more attention to the changes in high-frequency beta waves. Based on the real-time changes in the patient's EEG data, the self-attention mechanism operation on the EEG differential feature vector can make the anesthesia depth analysis model pay more attention to the individual changes in the patient's EEG data, further improving the accuracy of the anesthesia depth analysis results.The feature splicing layer is used to perform differential operations on the baseline physiological signal data and the real-time physiological signal data to construct physiological signal differential data. Similarly, the physiological signal differential data can better reflect the patient's personalized physiological signal data, and then the enhanced brain wave differential feature vector is spliced ​​with the physiological signal differential data to construct anesthesia depth analysis data; the second feature extraction layer is established based on the restricted Boltzmann machine and is used to process the anesthesia depth analysis data to construct anesthesia depth analysis vector, and the second feature extraction layer is updated in real time based on the monitoring database at the model update time point, so that the extracted anesthesia depth analysis vector is more in line with the patient's brain wave data changes and physiological signal data changes; the anesthesia depth value output layer is used to process the anesthesia depth analysis vector, which can be a full connection operation to output the anesthesia depth value. ;

[0020] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A method for monitoring depth of anesthesia, characterized in that: include: Obtaining brain wave data of the patient in a waking state and recording it as baseline brain wave data, obtaining physiological signal data of the patient in a waking state and recording it as baseline physiological signal data; At the monitoring time point, the patient's brain wave data and physiological signal data during the anesthesia process are obtained, and recorded as real-time brain wave data and real-time physiological signal data, respectively. The real-time brain wave data and real-time physiological signal data obtained at the current monitoring time point are stored in a monitoring database, and the baseline brain wave data, baseline physiological signal data, real-time brain wave data and real-time physiological signal data are sent to the anesthesia depth analysis model for processing, and the anesthesia depth value is output; The anesthesia depth analysis model includes a first feature extraction layer, a feature enhancement layer, a feature splicing layer, a second feature extraction layer and an anesthesia depth value output layer, wherein the first feature extraction layer is used to perform feature extraction on the baseline EEG data and the real-time EEG data to construct the corresponding baseline EEG feature vector and the real-time EEG feature vector, and the feature extraction specifically includes power spectrum density extraction and entropy feature extraction; the feature enhancement layer is used to perform differential processing on the baseline EEG feature vector and the real-time EEG feature vector to construct an EEG differential feature vector, and perform a self-attention mechanism operation on the EEG differential feature vector based on the monitoring database to construct an enhanced EEG differential feature vector; The feature concatenation layer is used to perform a differential operation on the baseline physiological signal data and the real-time physiological signal data to construct the physiological signal differential data, and then concatenate the enhanced EEG differential feature vector with the physiological signal differential data to construct the anesthesia depth analysis data; The second feature extraction layer is established based on the restricted Boltzmann machine, and is used to process the anesthesia depth analysis data to construct an anesthesia depth analysis vector, and the second feature extraction layer is updated in real time based on the monitoring database at the model update time point; the anesthesia depth value output layer is used to process the anesthesia depth analysis vector to output the anesthesia depth value.

2. A method for monitoring depth of anesthesia according to claim 1, characterized in that: Based on the monitoring database, a self-attention mechanism operation is performed on the EEG differential feature vector to construct an enhanced EEG differential feature vector, which specifically includes the following steps: Acquire the real-time brain wave data corresponding to the N-1 monitoring time points before the current monitoring time point from the monitoring database, and process the real-time brain wave data corresponding to the N-1 monitoring time points before the current monitoring time point through the first feature extraction layer to obtain the corresponding real-time brain wave feature vector, and form a brain wave timing analysis data set with the real-time brain wave feature vector corresponding to the N-1 monitoring time points before the current monitoring time point and the real-time brain wave feature vector corresponding to the current monitoring time point, and then send the brain wave timing analysis data set to the timing analysis network for processing to construct a brain wave change feature vector; The brain wave difference feature vector is matrix multiplied with the value weight matrix and the key weight matrix to obtain the brain wave value vector V and the brain wave key vector K. Then the brain wave change feature vector is matrix multiplied with the query weight matrix to obtain the brain wave query vector Q. The self-attention mechanism operation is implemented on the brain wave difference feature vector through the following formula: , where H is the enhanced EEG differential feature vector, softmax() is the activation function, and D is the dimension of the EEG key vector K.

3. A method for monitoring depth of anesthesia according to claim 2, characterized in that: At the model update time point, the second feature extraction layer is updated in real time based on the monitoring database. The specific steps are as follows: At the model update time point, the real-time brain wave data corresponding to all monitoring time points in the monitoring database are processed by the first feature extraction layer respectively to obtain the corresponding real-time brain wave feature vectors, and for any two adjacent monitoring time points, the real-time brain wave feature vectors corresponding to the two adjacent monitoring time points are subjected to a differential operation to obtain a training brain wave differential feature vector, and the real-time physiological signal data corresponding to the two adjacent monitoring time points are subjected to a differential operation to obtain training physiological signal differential data, and then the training brain wave differential feature vectors and the training physiological signal differential data corresponding to the two adjacent monitoring time points are subjected to a feature extraction real-time training sample; all the feature extraction real-time training samples are combined into a feature extraction real-time training set; and a number of monitoring time points are included between two adjacent model update time points; Performing unsupervised training on the second feature extraction layer through the real-time training set to obtain a trained second feature extraction layer; The second feature extraction layer is updated in real time using the following formula: i =1+βε i , where θ i is the parameter in the second feature extraction layer, i=1,2,3,…,I, I is the total number of parameters in the second feature extraction layer, β is the learning rate of the second feature extraction layer, ε i is the parameter θ in the trained second feature extraction layer i The corresponding parameters.

4. A method for monitoring depth of anesthesia according to claim 3, characterized in that: The second feature extraction layer is trained unsupervised using the real-time training set as follows: The second feature extraction layer includes a visible layer and a hidden layer. The feature extraction real-time training sample is sent to the visible layer. The activation probability of the hidden layer is calculated through the feature extraction weight matrix and the feature extraction bias value between the visible layer and the hidden layer. Then, the first hidden layer state data in the hidden layer is obtained based on the activation probability sampling of the hidden layer. The first hidden layer state data in the hidden layer is back-propagated to the visible layer through the feature extraction weight matrix and the feature extraction bias value between the visible layer and the hidden layer, and the visible layer state data of the visible layer is reconstructed. Then, the visible layer state data of the visible layer is transmitted through the visible layer and the hidden layer. The feature extraction weight matrix and feature extraction bias value between the layers are forward propagated to the hidden layer to obtain the second hidden layer state data in the hidden layer. Finally, the contrast divergence processing is performed through the feature extraction real-time training samples, the first hidden layer state data, the visible layer state data and the second hidden layer state data to maximize the logarithmic likelihood value of the feature extraction real-time training samples to adjust the parameters of the feature extraction weight matrix and feature extraction bias value between the visible layer and the hidden layer; and the training goal is that the number of training times reaches the maximum number of training times. When the training goal is achieved, the trained second feature extraction layer is output.

5. A method for monitoring depth of anesthesia according to claim 4, characterized in that: The timing analysis network is established based on the LSTM model, including an LSTM layer and a fully connected layer. The LSTM layer is used to perform timing analysis on the EEG timing analysis dataset, and the fully connected layer is used to perform a fully connected operation on the output of the LSTM layer to obtain the EEG change feature vector.

6. A method for monitoring depth of anesthesia according to claim 5, characterized in that: Training the anesthesia depth analysis model includes the following steps: Acquire several anesthesia depth analysis training samples, the anesthesia depth analysis training samples include real-time EEG data and real-time physiological signal data acquired from the same monitoring time point and the same patient, and establish a one-to-one correspondence link with the anesthesia depth analysis training samples through the patient's corresponding baseline EEG data, baseline physiological signal data and monitoring database, anesthesia depth analysis training samples are labeled by anesthesia depth values, all labeled anesthesia depth analysis training samples form an anesthesia depth analysis training set, anesthesia depth analysis model is trained by the anesthesia depth analysis training set, and it is determined whether the training conditions are met. If the training conditions are met, the trained anesthesia depth analysis model is output; Otherwise, continue to train the anesthesia depth analysis model using the anesthesia depth analysis training set.

7. An anesthesia depth monitoring system, characterized in that: The system applies a method for monitoring depth of anesthesia according to any one of claims 1 to 6, comprising: A reference data acquisition module is used to obtain the brain wave data of the patient in the awake state and record it as the reference brain wave data, and obtain the physiological signal data of the patient in the awake state and record it as the reference physiological signal data; The real-time data acquisition module is used to obtain the brain wave data and physiological signal data of the patient during the anesthesia process at the monitoring time point, and record them as real-time brain wave data and real-time physiological signal data respectively, and store the real-time brain wave data and real-time physiological signal data obtained at the current monitoring time point in the monitoring database; An anesthesia depth analysis module is used to send the baseline brain wave data, the baseline physiological signal data, the real-time brain wave data and the real-time physiological signal data to the anesthesia depth analysis model for processing and output the anesthesia depth value; The anesthesia depth analysis model includes a first feature extraction layer, a feature enhancement layer, a feature splicing layer, a second feature extraction layer and an anesthesia depth value output layer, wherein the first feature extraction layer is used to perform feature extraction on the baseline EEG data and the real-time EEG data to construct the corresponding baseline EEG feature vector and the real-time EEG feature vector, and the feature extraction specifically includes power spectral density extraction and entropy feature extraction; the feature enhancement layer is used to perform differential processing on the baseline EEG feature vector and the real-time EEG feature vector to construct the EEG differential feature vector, and perform a self-attention mechanism operation on the EEG differential feature vector based on the monitoring database to obtain the corresponding baseline EEG feature vector and the real-time EEG feature vector. Construct an enhanced brain wave differential feature vector; the feature splicing layer is used to perform a differential operation on the baseline physiological signal data and the real-time physiological signal data to construct the physiological signal differential data, and then splice the enhanced brain wave differential feature vector with the physiological signal differential data to construct the anesthesia depth analysis data; the second feature extraction layer is established based on the restricted Boltzmann machine, and is used to process the anesthesia depth analysis data to construct the anesthesia depth analysis vector, and the second feature extraction layer is updated in real time based on the monitoring database at the model update time point; the anesthesia depth value output layer is used to process the anesthesia depth analysis vector to output the anesthesia depth value.

Images