Intraoperative patient anesthesia state evaluation system based on electroencephalogram signal and ear temperature monitoring

Abstract

The invention relates to the technical field of anesthesia state evaluation, in particular to an intra-operative patient anesthesia state evaluation system based on EEG (electroencephalogram) signals and ear temperature monitoring, which comprises a data acquisition unit, a data processing unit and an evaluation unit, performing data synchronization on the electroencephalogram signal and the external auditory canal temperature signal of the target patient according to a synchronous clock model; the correlation modeling unit is used for conducting dynamic correlation modeling on the target patient according to the electroencephalogram signals and the external auditory canal temperature signals so as to obtain a PAC feature matrix between the electroencephalogram signals and the external auditory canal temperature signals. By combining the electroencephalogram signals and the ear temperature monitoring data, the anesthesia depth of a patient can be evaluated more comprehensively and accurately, the electroencephalogram signals can reflect the electrical activity of the brain, the ear temperature reflects the overall physiological state of the body, especially the influence of anesthetic on the body temperature, the two signals are comprehensively analyzed, and the anesthesia depth of the patient can be evaluated more comprehensively and accurately. And the accuracy of anesthesia state evaluation can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of anesthesia state assessment, and particularly to an intraoperative patient anesthesia state assessment system based on electroencephalogram signal and ear temperature monitoring. Background Art

[0002] Traditional systems usually rely on a single physiological parameter, such as electroencephalogram signal or body temperature, etc. This single-signal monitoring method may not comprehensively reflect the patient's anesthesia state; and because traditional systems rely on a single signal, and these signals are often affected by external interferences (such as individual differences in anesthetic drugs, external environment, etc.), this may lead to inaccurate assessment of the patient's anesthesia state, thereby affecting anesthesia regulation during the operation. Especially in a changing surgical environment, traditional methods may not be able to timely and accurately capture the changes in the patient's body; and traditional systems usually adopt a static model and cannot capture the dynamic changes of the patient's anesthesia state in real time. The assessment of anesthesia depth may lag behind the actual physiological changes of the patient, resulting in a slow response speed for adjusting the anesthesia depth, thereby affecting the safety of the anesthesia process; moreover, traditional systems often do not have an intelligent feedback mechanism. The infusion rate of anesthetic drugs usually relies on manual monitoring and adjustment, and there is a risk of human error. Especially in complex surgeries, it is difficult for anesthesiologists to monitor all physiological parameters simultaneously, and manual adjustment is easily affected by factors such as fatigue and misjudgment. Without a real-time and automated feedback mechanism, it may lead to over-anesthesia or under-anesthesia, increasing the risk of anesthesia accidents. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the prior art and provide an intraoperative patient anesthesia state assessment system based on electroencephalogram signal and ear temperature monitoring.

[0004] The technical solution adopted to solve the above technical problem is: an intraoperative patient anesthesia state assessment system based on electroencephalogram signal and ear temperature monitoring, including:

[0005] A data acquisition unit, which is used to obtain the electroencephalogram signal and the external auditory canal temperature of a target patient according to EEG leads and an ear temperature sensor, and synchronize the electroencephalogram signal and the external auditory canal temperature signal of the target patient according to a synchronous clock model;

[0006] An association modeling unit, which is used to perform dynamic association modeling on the target patient according to the electroencephalogram signal and the external auditory canal temperature signal to obtain a PAC feature matrix between the electroencephalogram signal and the external auditory canal temperature signal;

[0007] A feature extraction unit, which is used to construct a dynamic graph based on each channel of the EEG leads, the ear temperature sensor, and the PAC feature matrix, and perform spatio-temporal feature extraction on the dynamic graph according to a graph neural network to obtain the spatio-temporal features of each node of the dynamic graph;

[0008] A feature fusion unit, which is used to fuse the features of each node of the dynamic graph through the spatio-temporal features according to the attention mechanism to obtain the fusion features of the dynamic graph;

[0009] An anesthesia evaluation unit, which is used to perform anesthesia depth regression and anesthesia state classification on the target patient according to the fusion features of the dynamic graph to obtain the CSI index and anesthesia state level of the target patient;

[0010] An anesthesia control unit, which is used to real-time feedback the CSI index and anesthesia state level of the target patient to an anesthesia infusion pump, and dynamically adjust the anesthesia drug infusion rate according to a fuzzy PID controller.

[0011] Preferably, the expression of the synchronous clock model is as follows:

[0012] t sync =t EEG +Δt delay ;

[0013] Wherein, t sync represents the synchronized signal time, t EEG represents the acquisition time of the EEG signal, Δt delay represents the acquisition time delay of the external auditory canal temperature signal, and Δt delay ~N(0,z 2 ), N(0,z 2 ) represents a Gaussian distribution, and z represents the standard deviation of the delay.

[0014] Preferably, performing dynamic association modeling on the target patient according to the EEG signal and the external auditory canal temperature signal to obtain the PAC feature matrix between the EEG signal and the external auditory canal temperature signal, including:

[0015] Obtaining each frequency band signal of the EEG signal according to a band-pass filter, wherein each frequency band signal includes a δ wave, a θ wave, an α wave, and a β wave, and performing a Hilbert transform on each frequency band signal to obtain the phase of each frequency band signal;

[0016] Calculating the low-frequency amplitude of the external auditory canal temperature signal according to the short-time Fourier transform;

[0017] Calculating the coupling index between the phase of each frequency band signal and the low-frequency amplitude of the external auditory canal temperature signal according to PAC;

[0018] Obtain the PAC feature matrix between the EEG signal and the external auditory canal temperature signal according to the coupling index between the phases of the respective band signals and the low-frequency amplitude of the external auditory canal temperature signal.

[0019] Preferably, the calculation formula of the coupling index is as follows:

[0020]

[0021] Among them, PAC (f,TT) represents the phase-amplitude coupling index, (f, TT) represents the frequency band combination of phase-amplitude coupling, and φ f (t) represents the phase of each frequency band of the EEG signal, and A TT (t) represents the amplitude of the external auditory canal temperature, and N represents the total number of samples.

[0022] Preferably, construct a dynamic graph according to each channel of the EEG lead, the ear temperature sensor, and the PAC feature matrix, including:

[0023] Take each channel of the EEG lead and the ear temperature sensor as the nodes of the dynamic graph, where the nodes of the dynamic graph include EEG lead nodes and ear temperature sensor nodes;

[0024] Take the low-frequency amplitude of the external auditory canal temperature signal and the PAC feature matrix as the feature vectors of each node of the dynamic graph;

[0025] Calculate the weights of each dynamic edge in the dynamic graph according to the feature vectors of each node of the dynamic graph, and obtain the adjacency matrix of the dynamic graph according to the weights of each dynamic edge in the dynamic graph;

[0026] Update the weights of each dynamic edge in the dynamic graph according to the preset graph structure update frequency.

[0027] Preferably, the calculation formula of the weight of the dynamic edge is as follows:

[0028]

[0029] Among them, ω ij represents the weight of the dynamic edge between the i-th node and the j-th node in the dynamic graph, σ represents the bandwidth parameter, and f i and f j represent the feature vectors of the i-th node and the j-th node in the dynamic graph.

[0030] Preferably, perform spatio-temporal feature extraction on the dynamic graph according to the graph neural network to obtain the spatio-temporal features of the dynamic graph, including:

[0031] Perform spatial graph convolution on the dynamic graph according to the adjacency matrix to obtain the spatial features of each node in the dynamic graph. The calculation formula for the spatial features is as follows:

[0032]

[0033] where represents the spatial feature of the (l + 1)-th layer, which is the aggregation result of brain region information, represents the normalized adjacency matrix, represents the degree matrix of the normalized adjacency matrix, represents the trainable spatial convolution kernel of the (l + 1)-th layer, H (l) represents the input feature of the l-th layer, and the input feature represents the feature vector of each node in the dynamic graph;

[0034] Perform temporal feature extraction on the feature vectors of each node in the dynamic graph according to the LSTM model to obtain the temporal features of each node in the dynamic graph. The calculation formula for the spatial features is as follows:

[0035]

[0036] where represents the temporal feature of the (l + 1)-th layer LSTM, C (l+1) represents the memory state of the (l + 1)-th layer LSTM, represents the trainable weight matrix of the l-th layer LSTM;

[0037] Perform residual connection on the spatial features, temporal features and feature vectors of each node in the dynamic graph to obtain the spatio-temporal features of each node in the dynamic graph.

[0038] Preferably, perform feature fusion on each node of the dynamic graph through the spatio-temporal features according to the attention mechanism to obtain the fusion features of the dynamic graph, including:

[0039] Calculate the attention weights of each EEG lead node in the dynamic graph according to the fusion features of each EEG lead node in the dynamic graph. The calculation formula for the attention weights of the EEG lead nodes is as follows:

[0040]

[0041] where represents the attention weight of the i-th EEG lead node, σ represents the Sigmoid activation function, represents the trainable projection matrix of the EEG lead node, represents the bias term of the EEG lead node, represents the spatio-temporal feature of the EEG lead node;

[0042] Calculate the attention weight of the ear temperature sensor nodes in the dynamic graph according to the fusion features of the ear temperature sensor nodes in the dynamic graph, where the calculation formula of the attention weight of the ear temperature sensor nodes is as follows:

[0043]

[0044] Where represents the attention weight of the ear temperature sensor node, represents the trainable projection matrix of the ear temperature sensor node, represents the bias term of the ear temperature sensor node, represents the spatio-temporal feature of the ear temperature sensor node;

[0045] Fuse the spatio-temporal features of each node in the dynamic graph according to the attention weight of each EEG lead node in the dynamic graph and the attention weight of the ear temperature sensor node in the dynamic graph to obtain the fusion feature of the dynamic graph, where the calculation formula of the fusion feature is as follows:

[0046]

[0047] Where H fused represents the fusion feature of the dynamic graph, V EEG represents each EEG lead node in the dynamic graph, V TT represents the ear temperature sensor node in the dynamic graph.

[0048] Preferably, perform anesthesia depth regression and anesthesia state classification on the target patient according to the fusion feature of the dynamic graph to obtain the CSI index and anesthesia state level of the target patient, including:

[0049] Perform anesthesia depth regression on the fusion feature of the dynamic graph to obtain the CSI index of the target patient, where the calculation formula of the CSI index is as follows:

[0050]

[0051] Where represents the CSI index, W r represents the weight matrix of anesthesia depth regression, b r represents the bias term of anesthesia depth regression;

[0052] Perform anesthesia state classification on the fusion feature of the dynamic graph to obtain the anesthesia state level of the target patient, where the calculation formula of the probability of the anesthesia state level is as follows:

[0053] p = Softmax(W c Hfused +b c );

[0054] Among them, p represents the probability of the anesthesia state level, and W c represents the weight matrix of the anesthesia state classification, and b c represents the bias term of the anesthesia state classification.

[0055] Preferably, the CSI index and the anesthesia state level of the target patient are fed back to the anesthesia infusion pump in real time, and the infusion rate of the anesthetic is dynamically adjusted according to the fuzzy PID controller, including:

[0056] Define the state space according to the CSI index and the anesthesia state level of the target patient;

[0057] Define the action space according to the adjustment proportional gain, the adjustment integral gain, the adjustment differential gain of the fuzzy PID controller, and the weight factor of the adjustment fuzzy rule;

[0058] Determine the reward function according to the CSI index and the anesthesia state level of the target patient, where the calculation formula of the reward function is as follows:

[0059] r t = -(a1CSI + a2ST);

[0060] Among them, r t represents the reward function, a1 and a2 represent preset balance weights, CSI represents the CSI index of the target patient, and ST represents the anesthesia state level of the target patient;

[0061] Construct an agent according to the state space, the action space, and the reward function, and train the agent according to deep reinforcement learning.

[0062] The beneficial effects of the present invention are as follows: (1) By combining electroencephalogram (EEG) signals and ear temperature monitoring data, the present invention can more comprehensively and accurately evaluate the depth of anesthesia of patients. EEG signals can reflect the electrical activity of the brain, while ear temperature reflects the overall physiological state of the body, especially the impact of anesthetic drugs on body temperature. By comprehensively analyzing these two signals, the accuracy of anesthesia state assessment can be effectively improved, avoiding misjudgment caused by single-signal monitoring in traditional methods. Moreover, through the correlation modeling unit, dynamic correlation modeling is performed on EEG and ear temperature signals to extract the PAC feature matrix between the two. This modeling method can reveal the dynamic interaction between EEG signals and ear temperature, enabling anesthesia assessment to reflect the changes in the physiological state of patients in real time, and improving the sensitivity and response speed of assessment; (2) The present invention extracts spatio-temporal features through a graph neural network, extracts the spatio-temporal features of each node from a dynamic graph, and can effectively capture the spatial structure information and temporal variation law of EEG signals and ear temperature signals. By making full use of this information for in-depth anesthesia state analysis, the accuracy of anesthesia assessment is improved. Moreover, through the attention mechanism for feature fusion of spatio-temporal features, the fusion strategy can be dynamically adjusted according to the spatio-temporal features of each node, enabling different types of features to be fully utilized. The attention mechanism can automatically select the most relevant features, thereby improving the accuracy of anesthesia assessment and ensuring accurate judgment of the current anesthesia state of patients during anesthesia depth regression and state classification; (3) Through the anesthesia assessment unit, the present invention generates the CSI and anesthesia state level of patients in real time, and feeds this information back to the anesthesia control unit. Using a fuzzy PID controller, the system can dynamically adjust the infusion rate of anesthetic drugs according to the anesthesia state of patients. Such an intelligent feedback mechanism can optimize the anesthesia process in real time, ensure that patients maintain the best anesthesia state during surgery, avoid the situation of over-anesthesia or under-anesthesia, and can evaluate and adjust the anesthesia state in real time and accurately, reducing errors in manual operations, enhancing the safety of the anesthesia process. Especially in complex surgeries, anesthesiologists may have difficulty continuously monitoring all physiological parameters, while the system can provide real-time feedback in multiple dimensions, thereby reducing the occurrence of anesthesia accidents and ensuring the safety of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 FIG. is a schematic diagram of the system architecture of the overall system in an embodiment proposed by the present invention.

[0064] Reference numerals: 1, data acquisition unit; 2, correlation modeling unit; 3, feature extraction unit; 4, feature fusion unit; 5, anesthesia assessment unit; 6, anesthesia control unit. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] Example 1, as Figure 1 shown, the intraoperative patient anesthesia state assessment system based on EEG signals and ear temperature monitoring proposed by the present invention includes:

[0066] Data acquisition unit 1 is used to obtain the electroencephalogram signal and external auditory canal temperature of the target patient according to the EEG leads and ear temperature sensor, and synchronize the electroencephalogram signal and external auditory canal temperature signal of the target patient according to the synchronous clock model;

[0067] Association modeling unit 2 is used to perform dynamic association modeling on the target patient according to the electroencephalogram signal and external auditory canal temperature signal to obtain the PAC feature matrix between the electroencephalogram signal and external auditory canal temperature signal;

[0068] Feature extraction unit 3 is used to construct a dynamic graph according to each channel of the EEG leads, the ear temperature sensor and the PAC feature matrix, and perform spatio-temporal feature extraction on the dynamic graph according to the graph neural network to obtain the spatio-temporal features of each node of the dynamic graph;

[0069] Feature fusion unit 4 is used to fuse the features of each node of the dynamic graph through spatio-temporal features according to the attention mechanism to obtain the fusion features of the dynamic graph;

[0070] Anesthesia evaluation unit 5 is used to perform anesthesia depth regression and anesthesia state classification on the target patient according to the fusion features of the dynamic graph to obtain the CSI index and anesthesia state level of the target patient;

[0071] Anesthesia control unit 6 is used to real-time feedback the CSI index and anesthesia state level of the target patient to the anesthesia infusion pump, and dynamically adjust the anesthesia drug infusion rate according to the fuzzy PID controller.

[0072] In the present invention, EEG is a technology that measures and records the electrical activity of the cerebral cortex through electrodes, and EEG leads refer to electrodes placed in different positions for collecting brain electrical signals. Through these electrodes, the patient's brain wave activity can be monitored, which is very important for evaluating the depth of anesthesia; the ear temperature sensor is used to measure the temperature of the external auditory canal, which is considered to be an effective indicator of human body temperature and has a certain correlation with the anesthesia state in the body. During anesthesia, changes in body temperature may reflect the patient's anesthesia depth, so ear temperature monitoring can be used as an important data source for anesthesia state evaluation; the synchronous clock model is used to ensure that data collected from different sensors can be aligned in time; PAC refers to the coupling relationship between the phase and amplitude of the EEG signal in the frequency domain, and the PAC feature matrix is a feature matrix extracted between the EEG signal and the ear temperature signal, representing the phase-amplitude coupling relationship between them. This feature helps to find the correlation between the depth of anesthesia and the EEG signal and body temperature, and provides a basis for anesthesia evaluation; a dynamic graph is a graph structure used to represent the time-space relationship that changes over time. In this system, dynamic graphs are used to express how the interactions between different sensors (EEG, ear temperature) change over time. Each node represents a sensor or a specific signal channel, and the connections between nodes represent the correlation between signals. Graph neural networks are a type of neural network based on graph structures that can process graph data with nodes and edges. In this system, graph neural networks are used to extract spatiotemporal features from dynamic graphs and analyze the changing characteristics of each node (i.e., EEG leads or ear temperature sensors) in time and space dimensions, which helps to mine the complex dependencies and interactions between signals; spatiotemporal features refer to the characteristics of signals in time and space. Temporal features reflect the law of signal changes over time, while spatial features represent the relationships between different sensors; the attention mechanism is a technology that mimics the human attention concentration process. It emphasizes important information and ignores irrelevant information by assigning different weights to different information (or nodes); CSI is an index used to represent the depth of anesthesia, usually calculated based on EEG data. It evaluates the patient's state of consciousness by quantifying the characteristics of EEG signals (such as fluctuation frequency, amplitude, etc.). A lower CSI index usually indicates a deeper state of anesthesia; PID controller is a feedback control algorithm commonly used in automatic control systems. The fuzzy PID controller combines fuzzy logic with PID control and can effectively adjust system parameters under uncertain or ambiguous conditions. It is used to dynamically adjust the infusion rate of anesthetic drugs according to the patient's anesthesia state to ensure that the depth of anesthesia remains in an appropriate range.

[0073] Embodiment 2, the present invention proposes an intraoperative patient anesthesia state assessment system based on EEG signal and ear temperature monitoring. Compared with embodiment 1, this embodiment further includes: the expression of the synchronous clock model is as follows:

[0074] t sync = t EEG + Δt delay ;

[0075] Wherein, t sync represents the time of the synchronization signal, t EEG represents the acquisition time of the electroencephalogram signal, Δt delay represents the acquisition time delay of the external auditory canal temperature signal, and Δt delay ~ N(0, z 2 ), N(0, z 2 ) represents the Gaussian distribution, and z represents the standard deviation of the delay.

[0076] In an alternative embodiment, a dynamic association model is built for a target patient based on the electroencephalogram signal and the external auditory canal temperature signal to obtain a PAC feature matrix between the electroencephalogram signal and the external auditory canal temperature signal, including:

[0077] Obtain the band signals of the electroencephalogram signal according to a band-pass filter, wherein the band signals include delta waves, theta waves, alpha waves and beta waves, and perform Hilbert transform on the band signals to obtain the phases of the band signals;

[0078] Calculate the low-frequency amplitude of the external auditory canal temperature signal according to the short-time Fourier transform;

[0079] Calculate the coupling index between the phases of the band signals and the low-frequency amplitude of the external auditory canal temperature signal according to PAC;

[0080] Obtain the PAC feature matrix between the electroencephalogram signal and the external auditory canal temperature signal according to the coupling index between the phases of the band signals and the low-frequency amplitude of the external auditory canal temperature signal.

[0081] It should be noted that a band-pass filter is an electronic filter that allows signals within a certain frequency range to pass through while blocking signals below and above that range. In electroencephalogram (EEG) signal processing, band-pass filters are usually used to extract signals within a specific frequency range; Hilbert transform is a mathematical transform commonly used in signal processing. Through Hilbert transform, a complex signal can be generated from a real signal, where the real part is the original signal and the imaginary part is its phase information; short-time Fourier transform is a method for time-frequency analysis of signals. It divides the signal into small segments and then performs Fourier transform on each segment to obtain the time-frequency spectrum of the signal; low-frequency amplitude refers to the amplitude of the signal in the low-frequency range (usually 0.5 - 4 Hz). For the external auditory canal temperature signal, the low-frequency amplitude represents the intensity of the low-frequency components of the temperature signal, and these components are usually related to physiological processes and the anesthetic state.

[0082] In an alternative embodiment, the calculation formula of the coupling index is as follows:

[0083]

[0084] Among them, PAC (f,TT) represents the phase-amplitude coupling index, (f, TT) represents the frequency band combination of phase-amplitude coupling, and φ f (t) represents the phase of each frequency band of the EEG signal, and A TT (t) represents the amplitude of the external auditory canal temperature, and N represents the total number of samples.

[0085] In an optional embodiment, a dynamic graph is constructed based on each channel of the EEG lead, the ear temperature sensor, and the PAC feature matrix, including:

[0086] Regarding each channel of the EEG lead and the ear temperature sensor as the nodes of the dynamic graph, where the nodes of the dynamic graph include EEG lead nodes and ear temperature sensor nodes;

[0087] Regarding the low-frequency amplitude of the external auditory canal temperature signal and the PAC feature matrix as the feature vectors of each node of the dynamic graph;

[0088] Calculating the weights of each dynamic edge in the dynamic graph according to the feature vectors of each node of the dynamic graph, and obtaining the adjacency matrix of the dynamic graph according to the weights of each dynamic edge in the dynamic graph;

[0089] Updating the weights of each dynamic edge in the dynamic graph according to the preset graph structure update frequency.

[0090] It should be noted that in the dynamic graph, an edge is a connection representing the relationship between nodes, and a dynamic edge means that the strength or existence of these connections will change over time, which can reflect how the relationship between the EEG lead and the ear temperature sensor changes over time.

[0091] In an optional embodiment, the calculation formula for the weight of the dynamic edge is as follows:

[0092]

[0093] Among them, ω ij represents the weight of the dynamic edge between the i-th node and the j-th node in the dynamic graph, σ represents the bandwidth parameter, and f i and f j represent the feature vectors of the i-th node and the j-th node in the dynamic graph.

[0094] In an optional embodiment, spatio-temporal feature extraction is performed on the dynamic graph according to the graph neural network to obtain the spatio-temporal features of the dynamic graph, including:

[0095] Performing spatial graph convolution on the dynamic graph according to the adjacency matrix to obtain the spatial features of each node of the dynamic graph, where the calculation formula for the spatial features is as follows:

[0096]

[0097] Among them, represents the spatial feature of the (l + 1)-th layer, which is the aggregation result of the brain region information. represents the normalized adjacency matrix. represents the degree matrix of the normalized adjacency matrix. represents the trainable spatial convolution kernel of the (l + 1)-th layer, H (l) represents the input feature of the l-th layer, and the input feature represents the feature vectors of each node in the dynamic graph.

[0098] According to the LSTM model, time features are extracted from the feature vectors of each node in the dynamic graph to obtain the time features of each node in the dynamic graph. Among them, the calculation formula of the spatial feature is as follows:

[0099]

[0100] Among them, represents the time feature of the (l + 1)-th layer LSTM, C (l+1) represents the memory state of the (l + 1)-th layer LSTM. represents the trainable weight matrix of the l-th layer LSTM.

[0101] The spatial features, time features, and feature vectors of each node in the dynamic graph are subjected to residual connection to obtain the spatio-temporal features of each node in the dynamic graph.

[0102] It should be noted that the residual connection is a structure in a neural network. By means of skip connection, the input information is directly transmitted to the deep layer of the network, thereby avoiding the problem of gradient disappearance and improving the training efficiency.

[0103] In an optional embodiment, according to the attention mechanism, the features of each node in the dynamic graph are fused through spatio-temporal features to obtain the fused features of the dynamic graph, including:

[0104] Calculate the attention weights of each EEG lead node in the dynamic graph according to the fused features of each EEG lead node in the dynamic graph. Among them, the calculation formula of the attention weight of the EEG lead node is as follows:

[0105]

[0106] Among them, represents the attention weight of the i-th EEG lead node, and σ represents the Sigmoid activation function. represents the trainable projection matrix of the EEG lead node. represents the bias term of the EEG lead node. represents the spatio-temporal feature of the EEG lead node.

[0107] Calculate the attention weight of the ear temperature sensor nodes in the dynamic graph according to the fusion features of the ear temperature sensor nodes in the dynamic graph. The calculation formula of the attention weight of the ear temperature sensor nodes is as follows:

[0108]

[0109] Wherein, represents the attention weight of the ear temperature sensor node, represents the trainable projection matrix of the ear temperature sensor node, represents the bias term of the ear temperature sensor node, represents the spatio-temporal feature of the ear temperature sensor node;

[0110] Fuse the spatio-temporal features of each node in the dynamic graph according to the attention weight of each EEG lead node in the dynamic graph and the attention weight of the ear temperature sensor node in the dynamic graph to obtain the fusion feature of the dynamic graph. The calculation formula of the fusion feature is as follows:

[0111]

[0112] Wherein, H fused represents the fusion feature of the dynamic graph, V EEG represents each EEG lead node in the dynamic graph, V TT represents the ear temperature sensor node in the dynamic graph.

[0113] In an optional embodiment, perform anesthesia depth regression and anesthesia state classification on the target patient according to the fusion feature of the dynamic graph to obtain the CSI index and anesthesia state level of the target patient, including:

[0114] Perform anesthesia depth regression on the fusion feature of the dynamic graph to obtain the CSI index of the target patient. The calculation formula of the CSI index is as follows:

[0115]

[0116] Wherein, represents the CSI index, W r represents the weight matrix of anesthesia depth regression, b r represents the bias term of anesthesia depth regression;

[0117] Perform anesthesia state classification on the fusion feature of the dynamic graph to obtain the anesthesia state level of the target patient. The calculation formula of the probability of the anesthesia state level is as follows:

[0118] p = Softmax(W c H fused + b c );

[0119] Among them, p represents the probability of the anesthesia state level, and W c represents the weight matrix of the anesthesia state classification, and b c represents the bias term of the anesthesia state classification.

[0120] In an optional embodiment, the CSI index and the anesthesia state level of the target patient are fed back to the anesthesia infusion pump in real time, and the infusion rate of the anesthetic is dynamically adjusted according to the fuzzy PID controller, including:

[0121] Defining a state space according to the CSI index and the anesthesia state level of the target patient;

[0122] Defining an action space according to the adjustment proportional gain, adjustment integral gain, adjustment differential gain of the fuzzy PID controller, and the weight factor of the adjustment fuzzy rule;

[0123] Determining a reward function according to the CSI index and the anesthesia state level of the target patient, where the calculation formula of the reward function is as follows:

[0124] r t = -(a1CSI + a2ST);

[0125] Among them, r t represents the reward function, a1 and a2 represent preset balance weights, CSI represents the CSI index of the target patient, and ST represents the anesthesia state level of the target patient;

[0126] Constructing an agent according to the state space, action space and reward function, and training the agent according to deep reinforcement learning.

[0127] It should be noted that the state space refers to the set of all possible system states in reinforcement learning or control systems; the action space refers to the set of all possible actions that the agent can take in each state; the reward function is the criterion used to evaluate the behavior of the agent in reinforcement learning; the agent is a system that can perceive the environment and take actions to achieve the goal. In reinforcement learning, the agent learns the optimal policy through interaction with the environment. Here, the agent perceives the environment through the state space (the CSI index and the anesthesia state level of the target patient), and performs control through the action space (adjusting the parameters of the PID controller), and finally evaluates the effect of the action through the reward function; deep reinforcement learning is a method that combines deep learning and reinforcement learning, used to train the agent to learn the optimal policy in a complex environment. Through the deep neural network, the agent can process and learn high-dimensional and complex state spaces and action spaces. During the training process, the agent tries and makes mistakes according to the reward function and gradually learns to optimize its behavior.

[0128] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto, and various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art.

Claims

1. An intraoperative patient anesthesia state evaluation system based on electroencephalogram signal and ear temperature monitoring, characterized in that, Including: A data acquisition unit (1) which is used to acquire the electroencephalogram signal and the external auditory canal temperature of a target patient according to EEG leads and an ear temperature sensor, and synchronize the electroencephalogram signal and the external auditory canal temperature signal of the target patient according to a synchronous clock model; A correlation modeling unit (2) which is used to perform dynamic correlation modeling on the target patient according to the electroencephalogram signal and the external auditory canal temperature signal to obtain a PAC feature matrix between the electroencephalogram signal and the external auditory canal temperature signal; A feature extraction unit (3) which is used to construct a dynamic graph according to each channel of the EEG leads, the ear temperature sensor and the PAC feature matrix, and perform spatio-temporal feature extraction on the dynamic graph according to a graph neural network to obtain the spatio-temporal features of each node of the dynamic graph; A feature fusion unit (4) which is used to fuse the features of each node of the dynamic graph through the spatio-temporal features according to an attention mechanism to obtain the fusion features of the dynamic graph; An anesthesia evaluation unit (5) which is used to perform anesthesia depth regression and anesthesia state classification on the target patient according to the fusion features of the dynamic graph to obtain the CSI index and the anesthesia state level of the target patient; An anesthesia control unit (6) which is used to feedback the CSI index and the anesthesia state level of the target patient to an anesthesia infusion pump in real time, and dynamically adjust the anesthesia drug infusion rate according to a fuzzy PID controller.

2. The intraoperative patient anesthesia state evaluation system based on electroencephalogram signal and ear temperature monitoring according to claim 1, characterized in that, The expression of the synchronous clock model is as follows: t sync = t EEG + Δt delay ; Among them, t sync represents the signal time of synchronization, t EEG represents the acquisition time of the electroencephalogram signal, Δt delay represents the acquisition time of the delay of the external auditory canal temperature signal, and Δt delay ~N(0, z 2 ), N(0, z 2 ) represents the Gaussian distribution, and z represents the standard deviation of the delay.

3. The intraoperative patient anesthesia state evaluation system based on EEG signal and ear temperature monitoring according to claim 2, wherein Performing dynamic correlation modeling on the target patient according to the electroencephalogram signal and the external auditory canal temperature signal to obtain a PAC feature matrix between the electroencephalogram signal and the external auditory canal temperature signal, including: Obtaining each frequency band signal of the electroencephalogram signal according to a band-pass filter, where each frequency band signal includes a δ wave, a θ wave, an α wave and a β wave, and performing Hilbert transform on each frequency band signal to obtain the phase of each frequency band signal; Calculating the low-frequency amplitude of the external auditory canal temperature signal according to short-time Fourier transform; Calculating the coupling index between the phase of each frequency band signal and the low-frequency amplitude of the external auditory canal temperature signal according to PAC; Obtaining the PAC feature matrix between the electroencephalogram signal and the external auditory canal temperature signal according to the coupling index between the phase of each frequency band signal and the low-frequency amplitude of the external auditory canal temperature signal.

4. The intraoperative patient anesthesia state evaluation system based on EEG signal and ear temperature monitoring according to claim 3, characterized in that, The calculation formula of the coupling index is as follows: Among them, PAC (f,TT) represents the phase-amplitude coupling index, (f, TT) represents the frequency band combination of phase-amplitude coupling, and φ f (t) represents the phase of each frequency band of the electroencephalogram signal, and A TT (t) represents the amplitude of the external auditory canal temperature, and N represents the total number of samples.

5. The intraoperative patient anesthesia state evaluation system based on electroencephalogram signal and ear temperature monitoring according to claim 4, characterized in that, Constructing a dynamic graph according to each channel of the EEG leads, the ear temperature sensor and the PAC feature matrix, including: Regarding each channel of the EEG leads and the ear temperature sensor as the nodes of the dynamic graph, where the nodes of the dynamic graph include EEG lead nodes and ear temperature sensor nodes; Regarding the low-frequency amplitude of the external auditory canal temperature signal and the PAC feature matrix as the feature vectors of each node of the dynamic graph; Calculate the weights of each dynamic edge in the dynamic graph according to the eigenvectors of each node in the dynamic graph, and obtain the adjacency matrix of the dynamic graph according to the weights of each dynamic edge in the dynamic graph; Update the weights of each dynamic edge in the dynamic graph according to the preset graph structure update frequency.

6. The intraoperative patient anesthesia state evaluation system based on electroencephalogram signal and ear temperature monitoring according to claim 5, characterized in that, The calculation formula of the weight of the dynamic edge is as follows: where ω ij represents the weight of the dynamic edge between the i-th node and the j-th node in the dynamic graph, σ represents the bandwidth parameter, f i and f j represent the feature vectors of the i-th node and the j-th node in the dynamic graph.

7. The intraoperative patient anesthesia state evaluation system based on EEG signal and ear temperature monitoring according to claim 6, characterized in that, Perform spatio-temporal feature extraction on the dynamic graph according to the graph neural network to obtain the spatio-temporal features of the dynamic graph, including: Perform spatial graph convolution on the dynamic graph according to the adjacency matrix to obtain the spatial features of each node in the dynamic graph. The calculation formula of the spatial features is as follows: Among them, represents the spatial features of the (l + 1)-th layer, which represents the aggregation result of brain region information, represents the normalized adjacency matrix, represents the degree matrix of the normalized adjacency matrix, represents the trainable spatial convolution kernel of the (l + 1)-th layer, H (l) represents the input features of the l-th layer, and the input features represent the feature vectors of each node in the dynamic graph; Perform temporal feature extraction on the eigenvectors of each node in the dynamic graph according to the LSTM model to obtain the temporal features of each node in the dynamic graph. The calculation formula of the spatial features is as follows: Among them, represents the time feature of the (l + 1)-th layer LSTM, C (l+1) represents the memory state of the (l + 1)-th layer LSTM, represents the trainable weight matrix of the l-th layer LSTM; Perform residual connection on the spatial features, temporal features and eigenvectors of each node in the dynamic graph to obtain the spatio-temporal features of each node in the dynamic graph.

8. The intraoperative patient anesthesia state evaluation system based on electroencephalogram signal and ear temperature monitoring according to claim 7, characterized in that, Perform feature fusion on each node of the dynamic graph through the spatio-temporal features according to the attention mechanism to obtain the fusion features of the dynamic graph, including: Calculate the attention weights of each EEG lead node in the dynamic graph according to the fusion features of each EEG lead node in the dynamic graph. The calculation formula of the attention weights of the EEG lead nodes is as follows: Among them, represents the attention weight of the i-th EEG lead node, σ represents the Sigmoid activation function, represents the trainable projection matrix of the EEG lead node, represents the bias term of the EEG lead node, represents the spatio-temporal feature of the EEG lead node; Calculate the attention weights of the ear temperature sensor nodes in the dynamic graph according to the fusion features of the ear temperature sensor nodes in the dynamic graph. The calculation formula of the attention weights of the ear temperature sensor nodes is as follows: Among them, represents the attention weight of the ear temperature sensor node, represents the trainable projection matrix of the ear temperature sensor node, represents the bias term of the ear temperature sensor node, represents the spatio-temporal feature of the ear temperature sensor node; Fuse the spatio-temporal features of each node in the dynamic graph according to the attention weights of each EEG lead node in the dynamic graph and the attention weights of the ear temperature sensor nodes in the dynamic graph to obtain the fusion features of the dynamic graph. The calculation formula of the fusion features is as follows: Among them, H fused represents the fusion feature of the dynamic graph, V EEG represents each EEG lead segment in the dynamic graph, V TT represents the ear temperature sensor node in the dynamic graph.

9. The intraoperative patient anesthesia state evaluation system based on electroencephalogram signal and ear temperature monitoring according to claim 8, wherein Perform anesthesia depth regression and anesthesia state classification on the target patient according to the fusion features of the dynamic graph to obtain the CSI index and anesthesia state level of the target patient, including: Perform anesthesia depth regression on the fusion features of the dynamic graph to obtain the CSI index of the target patient. The calculation formula of the CSI index is as follows: Among them, represents the CSI index, and W r represents the weight matrix for the regression of anesthetic depth, and b r represents the bias term for the regression of anesthetic depth; Perform anesthesia state classification on the fusion features of the dynamic graph to obtain the anesthesia state level of the target patient. The calculation formula of the probability of the anesthesia state level is as follows: p = Softmax(W c H fused + b c ); Among them, p represents the probability of the anesthesia state level, and W c represents the weight matrix of the anesthesia state classification, and b c represents the bias term of the anesthesia state classification.

10. The intraoperative patient anesthesia state evaluation system based on EEG signal and ear temperature monitoring according to claim 9, characterized in that, Real-time feedback the CSI index and anesthesia state level of the target patient to the anesthesia infusion pump, and dynamically adjust the anesthesia drug infusion rate according to the fuzzy PID controller, including: Define the state space according to the CSI index and anesthesia state level of the target patient; Define the action space according to the adjustment proportional gain, adjustment integral gain, adjustment differential gain and adjustment fuzzy rule weight factor of the fuzzy PID controller; Determine the reward function according to the CSI index and anesthesia state level of the target patient. The calculation formula of the reward function is as follows: r t = -(a1CSI + a2ST); Among them, r t represents the reward function, a1 and a2 represent preset balance weights, CSI represents the CSI index of the target patient, and ST represents the anesthesia state level of the target patient; Construct an agent according to the state space, the action space and the reward function, and train the agent according to deep reinforcement learning.

Images