A multi-parameter fusion-based clinical anesthetic depth data monitoring method and device

Abstract

The application discloses a kind of clinical anesthetic depth data monitoring method and device based on multi-parameter fusion, it is related to anesthetic monitoring technical field, comprising: S1, electroencephalogram acquisition and causal analysis;S2, brain network construction and feature extraction;S3, anesthetic depth evaluation;S4, reinforcement learning decision;S5, closed-loop control execution.The present application calculates the directional transfer function between lead pair on the basis of multi-lead electroencephalogram signal, constructs the directed weighted brain function network reflecting real information flow direction, the global graph theory feature extracted can quantify the whole brain information integration efficiency, local graph theory feature can locate the key hub node of anesthetic drug action, finally, the brain network state representation vector formed by attention mechanism fusion, the direction, intensity and path efficiency of information flow between each brain area, such as multidimensional information, close to the neurobiological nature of anesthetic drug-induced unconsciousness, provide more abundant, more reliable state information basis for subsequent precise control.

Description

Technical Field

[0001] This invention relates to the field of anesthesia monitoring technology, specifically to a method and device for monitoring clinical anesthesia depth data based on multi-parameter fusion. Background Technology

[0002] Currently, monitoring the depth of anesthesia during general anesthesia surgery primarily relies on commercially available monitoring devices based on single-lead or limited-lead EEG signals, such as the bispectral index (BIS) and entropy index. These devices analyze the power spectrum of EEG signals, extract energy characteristics of specific frequency bands, and map them into dimensionless indices using empirical formulas for anesthesiologists' reference. In recent years, with the development of artificial intelligence technology, some studies have attempted to introduce deep learning models into anesthesia depth assessment, utilizing multimodal physiological data for fusion analysis to improve monitoring accuracy. Simultaneously, a few systems have begun exploring the use of pharmacokinetic models combined with current monitoring values ​​to predict future states and provide physicians with text-based regulatory suggestions to assist clinical decision-making.

[0003] However, the aforementioned existing technologies have shortcomings, namely, they have not formed an effective closed-loop mechanism at both the anesthesia depth perception level and the control decision level. Specifically:

[0004] At the perception level, existing methods generally use the spectral energy characteristics of local brain regions as the basis for judgment. Essentially, this is a "first-order statistical" analysis of EEG signals, which cannot reflect the inhibitory effect of anesthetic drugs on the information transmission and collaborative integration ability between different regions of the whole brain. The latter is precisely the core neural mechanism of loss of consciousness induced by general anesthesia.

[0005] At the control level, even if some solutions propose regulation suggestions based on scoring functions, they are essentially open-loop recommendation models that rely on fixed rules. The decision-making logic cannot adaptively evolve according to the individual patient's drug efficacy response during the operation, and the suggestions still need to be manually confirmed and implemented by the physician in the end, which cannot form a true "perception-decision-execution" closed-loop regulation.

[0006] The deficiencies at both levels mentioned above result in significant limitations of existing technologies in terms of monitoring accuracy, individual adaptability, and control timeliness, making it difficult to meet the clinical needs of modern precision anesthesia for intelligent systems.

[0007] In response to this problem, this application proposes a method and device for monitoring clinical anesthesia depth data based on multi-parameter fusion. Summary of the Invention

[0008] The purpose of this invention is to provide a clinical anesthesia depth data monitoring method and device based on multi-parameter fusion, so as to solve the problem that the existing technology has not formed an effective closed-loop mechanism at both the anesthesia depth perception level and the control decision level.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] Firstly, this application provides a method for monitoring clinical anesthesia depth data based on multi-parameter fusion, including:

[0011] Multi-lead EEG signals were acquired, and time-frequency transformation and inter-lead causal analysis were performed on the multi-lead EEG signals to obtain the orientation causal strength values ​​between each lead pair.

[0012] A time-varying weighted brain functional network is constructed based on the directional causal strength value, and the global graph theory features and local graph theory features of the time-varying weighted brain functional network are extracted. The brain network state representation vector is obtained by fusing the global graph theory features and local graph theory features.

[0013] The brain network state representation vector is input into the trained anesthesia depth assessment network, and the current anesthesia depth index is output.

[0014] The deviation between the current anesthesia depth index and the target anesthesia depth range is used as the state input to the reinforcement learning agent, which then outputs a drug delivery rate adjustment command based on the current state.

[0015] The target-controlled infusion device is driven to perform drug infusion according to the drug delivery rate adjustment command, and a new round of multi-lead EEG signals is continuously collected to form a closed-loop monitoring.

[0016] Furthermore, the step of performing time-frequency transformation and inter-lead causal analysis on the multi-lead EEG signals to obtain the orientational causal strength values ​​between each lead pair includes:

[0017] Multi-scale wavelet packet decomposition was performed on the EEG signal of each lead to extract the time-frequency components at each scale. For each pair of leads, the orientation transfer function value of the time-frequency component of one lead to the time-frequency component of another lead was calculated. The orientation transfer function values ​​at all scales were then weighted and fused according to the frequency contribution weight to obtain the orientation causal strength value between the lead pairs.

[0018] Furthermore, the value of the directional transfer function is determined in the following manner:

[0019] A vector autoregression model is constructed using the historical time-frequency sequence of lead 1 and the current time-frequency sequence of lead 2. The variance contribution ratio of lead 2 at the current time to the historical time-frequency value of lead 1 is calculated, and this variance contribution ratio is used as the directional transfer function value from lead 1 to lead 2.

[0020] Further, the step of constructing a time-varying weighted brain functional network based on the directional causal strength value, and extracting the global graph theory features and local graph theory features of the time-varying weighted brain functional network, includes:

[0021] A directed weighted network is constructed using each lead as a network node and the directed edge weights as directional causal strength values.

[0022] The directed weighted network is dynamically updated within a preset time window;

[0023] The average shortest path length and global clustering coefficient of the entire network are calculated as global graph theory features, and the betweenness centrality and in-degree strength of each node are calculated as local graph theory features.

[0024] Furthermore, the process of fusing the global graph theory features and local graph theory features to obtain the brain network state representation vector includes:

[0025] The betweenness centrality and in-degree strength of each node are encoded and concatenated into a local feature vector according to the node position, and the average shortest path length and the global clustering coefficient are concatenated into a global feature vector.

[0026] An attention mechanism is used to calculate the contribution weight of the global feature vector to each node in the local feature vector. The local feature vector is then weighted and aggregated according to the contribution weight and concatenated with the global feature vector to obtain the brain network state representation vector.

[0027] Further, the step of using the deviation between the current anesthesia depth index and the target anesthesia depth range as the state input reinforcement learning agent includes:

[0028] The difference between the current anesthesia depth index and the median value of the target interval is calculated as the baseline deviation value;

[0029] Calculate the rate of change of the current anesthesia depth index relative to the previous time step;

[0030] The basic deviation value and the rate of change are combined into a two-dimensional state vector and input into the reinforcement learning agent.

[0031] Furthermore, the reinforcement learning agent adopts an actor-critic architecture, wherein the critic network is used to evaluate the expected cumulative reward of executing the dosing rate adjustment instruction in the current state, and the actor network is used to output the dosing rate adjustment instruction according to the current state;

[0032] After each dosing rate adjustment command is executed, the actor network and critic network update their network parameters online based on changes in the actual anesthesia depth index.

[0033] Furthermore, the drug delivery rate adjustment command includes the adjustment direction and adjustment range;

[0034] The adjustment direction is to increase the infusion rate, decrease the infusion rate, or maintain the current rate;

[0035] The adjustment range is positively correlated with the magnitude of the base deviation value, and when the direction of the rate of change is opposite to the direction of the base deviation value, the adjustment range is multiplied by a decay coefficient less than 1.

[0036] Furthermore, after the target-controlled infusion device is driven to perform drug infusion according to the drug delivery rate adjustment command, it also includes:

[0037] Record the anesthesia depth index response value after each administration rate adjustment command is executed, and construct an administration-response time sequence record;

[0038] When two consecutive dosing rate adjustment commands are in opposite directions, the dosing-response timing record is marked as an oscillation event, and the upper limit of the adjustment amplitude of the subsequent reinforcement learning agent output is reduced.

[0039] Secondly, this application provides a clinical anesthesia depth data monitoring device based on multi-parameter fusion, comprising:

[0040] The EEG acquisition and causal analysis module is used to acquire multi-lead EEG signals, perform time-frequency transformation and inter-lead causal analysis on the multi-lead EEG signals, and obtain the directional causal strength value between each lead pair;

[0041] The brain network construction and feature extraction module is used to construct a time-varying weighted brain functional network based on the directional causal strength value, extract the global graph theory features and local graph theory features of the time-varying weighted brain functional network, and fuse the global graph theory features and local graph theory features to obtain the brain network state representation vector.

[0042] An anesthesia depth assessment module is used to input the brain network state representation vector into the trained anesthesia depth assessment network and output the current anesthesia depth index.

[0043] The reinforcement learning decision-making module is used to take the deviation between the current anesthesia depth index and the target anesthesia depth range as the state input to the reinforcement learning agent, and the reinforcement learning agent outputs the drug administration rate adjustment command according to the current state.

[0044] The closed-loop control execution module is used to drive the target-controlled infusion device to perform drug infusion according to the drug administration rate adjustment command, and to continuously collect a new round of multi-lead EEG signals to form closed-loop monitoring.

[0045] Compared with existing technologies, this invention provides a method and device for monitoring clinical anesthesia depth data based on multi-parameter fusion. This invention calculates the directional transfer function between lead pairs based on multi-lead EEG signals, constructing a directed weighted brain functional network reflecting the true flow of information. The extracted global graph theory features can quantify the efficiency of whole-brain information integration, while local graph theory features can locate key hub nodes of anesthetic drug action. Finally, the brain network state representation vector formed through attention mechanism fusion fully preserves multi-dimensional information such as the direction, intensity, and path efficiency of information flow between brain regions. This technical solution enables the system to perceive the decrease in communication efficiency between brain regions, the simplification of information transmission paths, and the inhibition of hub node functions. Compared with traditional spectral features that only reflect local activity, the perception dimension of this invention is closer to the neurobiological essence of drug-induced loss of consciousness, providing a richer and more reliable foundation of state information for subsequent precise control. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0047] Figure 1 A flowchart of a clinical anesthesia depth data monitoring method based on multi-parameter fusion is provided for an embodiment of the present invention;

[0048] Figure 2 This is a block diagram of a clinical anesthesia depth data monitoring device based on multi-parameter fusion, provided as an embodiment of the present invention. Detailed Implementation

[0049] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0050] As attached Figure 1 As shown:

[0051] Example 1:

[0052] A method for monitoring clinical anesthesia depth data based on multi-parameter fusion, comprising:

[0053] S1. Collect multi-lead EEG signals, perform time-frequency transformation and inter-lead causal analysis on the multi-lead EEG signals, and obtain the orientation causal strength value between each lead pair;

[0054] In step S1, the EEG signal of each lead is decomposed into multi-scale wavelet packets to extract the time-frequency components at each scale.

[0055] For each pair of leads, the orientation transfer function value of the time-frequency component of one lead to the time-frequency component of the other lead is calculated, and the orientation transfer function values ​​at all scales are weighted and fused according to the frequency contribution weight to obtain the orientation causal strength value between the lead pairs.

[0056] The value of the directional transfer function is determined in the following way:

[0057] A vector autoregression model is constructed using the historical time-frequency sequence of the first lead and the current time-frequency sequence of the second lead. The variance contribution ratio of the current time-frequency value of the second lead to the historical time-frequency value of the first lead is calculated, and this variance contribution ratio is used as the directional transfer function value from the first lead to the second lead.

[0058] Specifically, the patient's real-time EEG signals are first acquired using a multi-lead EEG acquisition device. Specifically, electrodes are arranged according to the international standard 10-20 lead system, preferably using an 8-channel or 16-channel EEG acquisition configuration, with a sampling frequency set to 250Hz to 1000Hz.

[0059] To ensure signal quality, baseline drift correction and power frequency notch processing were performed on the raw EEG signals during the acquisition process. Artifacts such as electrooculography (EOG) and electromyography (EMG) were removed using independent component analysis to obtain clean multi-lead EEG signals.

[0060] Time-frequency transformation and inter-lead causality analysis were performed on the acquired multi-lead EEG signals to obtain the orientational causality strength values ​​between each lead pair. The specific processing procedure is as follows:

[0061] First, the EEG signal of each lead is decomposed into multi-scale wavelet packets. The db4 or db8 wavelet from the Daubechies wavelet family is selected as the mother wavelet, and the number of decomposition layers is set to 4 to 6, thereby decomposing the original signal into time-frequency components with different bandwidths.

[0062] Compared with the traditional short-time Fourier transform, wavelet packet decomposition can provide more refined frequency resolution, especially in the delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), and beta (13–30 Hz) bands of interest for anesthesia depth monitoring, where it can obtain more stable time-frequency representations.

[0063] Furthermore, for each lead pair (i, j), the directional transfer function value of the time-frequency component of lead i with respect to the time-frequency component of lead j is calculated. The calculation of the directional transfer function is based on the theoretical framework of the vector autoregressive model:

[0064] A vector autoregression model is constructed using the historical time-frequency sequence of lead i and the current time-frequency sequence of lead j. The optimal order of the model is determined by the Akaike information criterion. After estimating the model parameters, the causal contribution of lead i to lead j is calculated. Specifically, it is the proportion of variance contribution explained by the historical time-frequency value of lead j at the current moment by the time-frequency value of lead i.

[0065] The larger this ratio, the stronger the directional influence of lead i on lead j. The directional transfer function values ​​calculated at all scales are weighted and fused according to the frequency contribution weights corresponding to each scale to finally obtain the directional causal strength value between the lead pairs.

[0066] Repeat the above calculation for all lead pairs to obtain a causal strength matrix containing complete directed connectivity information. This matrix reflects the dynamic changes in the flow of information and the intensity of interaction between different brain regions during the process of changing anesthesia.

[0067] Furthermore, the system preferably updates the causal strength matrix every 5 to 30 seconds to adapt to the dynamic evolution of brain functional connectivity under anesthesia. In clinical applications, when patients are awake, the directional connectivity between different brain regions exhibits high complexity and bidirectionality.

[0068] As anesthetic drugs take effect, the brain's functional network gradually simplifies, the directionality of information flow increases, and the intensity values ​​of certain pathways in the causal intensity matrix undergo regular changes. These changes are an important physiological basis for subsequent determination of the depth of anesthesia.

[0069] S2. Construct a time-varying weighted brain functional network based on the directional causal strength value, and extract the global graph theory features and local graph theory features of the time-varying weighted brain functional network. Then, fuse the global graph theory features and local graph theory features to obtain the brain network state representation vector.

[0070] In step S2, a directed weighted network is constructed with each lead as a network node and the directed edge weights as directional causal strength values.

[0071] The directed weighted network is dynamically updated within a preset time window;

[0072] The average shortest path length and global clustering coefficient of the entire network are calculated as global graph theory features, and the betweenness centrality and in-degree strength of each node are calculated as local graph theory features.

[0073] The betweenness centrality and in-degree strength of each node are encoded and concatenated into a local feature vector according to the node position, and the average shortest path length and the global clustering coefficient are concatenated into a global feature vector.

[0074] The attention mechanism is used to calculate the contribution weight of the global feature vector to each node in the local feature vector. The local feature vector is then weighted and aggregated according to the contribution weight and then concatenated with the global feature vector to obtain the brain network state representation vector.

[0075] Specifically, after obtaining the directional causal strength values ​​between each pair of leads, this invention constructs a time-varying weighted brain function network. Using each EEG lead as a network node, and the directional causal strength values ​​between each pair of leads calculated in the aforementioned steps as the weights of the directed edges, a directed weighted network is constructed.

[0076] This network fully preserves the strength and direction information of interactions between brain regions, and compared with traditional undirected correlation networks or coherent networks, it can more accurately characterize the differentiated effects of anesthetic drugs on brain information transmission pathways.

[0077] After the network is constructed, the system dynamically updates the directed weighted network within a preset time window.

[0078] Preferably, the time window length is set to 60 seconds and the window sliding step is 10 seconds, so that the network can update in real time according to the changes in EEG signals, ensuring that the network features can reflect the brain function state at the current moment.

[0079] Based on this, graph theory features of the network are extracted at both global and local scales. Regarding global graph theory features, the average shortest path length and global clustering coefficient of the entire network are calculated.

[0080] The average shortest path length reflects the efficiency of information transmission across the entire network. As the depth of anesthesia increases, the brain's ability to integrate information decreases, and the average shortest path length usually increases.

[0081] The global clustering coefficient reflects the trend of the network forming local information processing modules. Under anesthesia, the coefficient often shows a pattern of first increasing and then decreasing, which is closely related to the concentration and duration of action of anesthetic drugs.

[0082] In terms of local graph theory features, the betweenness centrality and in-degree strength of each node are calculated. Betweenness centrality measures the importance of a node as an information transmission hub in the network, while in-degree strength reflects the total amount of information received by the node from other nodes.

[0083] Under anesthesia, the betweenness centrality of certain key nodes (such as leads in the frontal lobe region) changes significantly, which is consistent with the inhibitory effect of anesthetic drugs on the function of the prefrontal cortex.

[0084] Furthermore, after completing the above feature extraction, the global graph theory features and local graph theory features are fused into a brain network state representation vector. The fusion process preferably employs the following strategy:

[0085] First, the betweenness centrality and in-degree strength of each node are concatenated into a local feature vector according to the spatial position of the node on the scalp. Then, the average shortest path length and the global clustering coefficient are concatenated into a global feature vector.

[0086] Then, an attention mechanism is introduced—using the global feature vector as the query vector and the features of each node in the local feature vector as the key vector, the attention score of the global features to each node is calculated, and the contribution weight of each node is obtained after softmax normalization.

[0087] This weight reflects the importance of each region in determining the depth of anesthesia under the current whole-brain state. Finally, the local feature vectors are weighted and aggregated according to their contribution weights, and then concatenated with the global feature vector to form the final brain network state representation vector.

[0088] This representation vector retains the network integration information at the global level while highlighting the local contributions of key nodes, and has a stronger ability to represent the depth of anesthesia than single-scale features.

[0089] S3. Input the brain network state representation vector into the trained anesthesia depth assessment network and output the current anesthesia depth index.

[0090] In step S3, the brain network state representation vector obtained in the previous steps is input into the pre-trained anesthesia depth assessment network, which outputs the anesthesia depth index at the current moment.

[0091] The preferred network for assessing anesthesia depth is a multilayer perceptron or a lightweight Transformer architecture. The multilayer perceptron architecture consists of an input layer, two hidden layers, and an output layer. The number of nodes in the hidden layers is set to 64 and 32, respectively. The ReLU activation function is used. The output layer uses a Sigmoid activation function to limit the output value to the range of 0 to 1, where 0 represents a fully awake state and 1 represents a deep anesthesia state.

[0092] The lightweight Transformer structure utilizes a self-attention mechanism to capture the long-range dependencies between components within the brain network state representation vector, which can further improve the evaluation accuracy.

[0093] Specifically, the anesthesia depth assessment network was pre-trained using supervised learning. The training data was constructed as follows:

[0094] Simultaneously, EEG signals and corresponding clinical anesthesia depth labels are collected from patients undergoing large-scale surgery. The clinical anesthesia depth labels are determined by experienced anesthesiologists based on clinical reference indicators such as the bispectral index (BIS) and entropy index, as well as the patient's intraoperative vital signs.

[0095] The training dataset needs to cover patient samples of different age groups, different types of surgery, and different anesthetic drug regimens to ensure that the network has good generalization ability.

[0096] During training, mean squared error is used as the loss function, the Adam optimizer is used for parameter updates, and an early stopping strategy is used to prevent overfitting. Finally, a lightweight evaluation model that can run in real time on embedded or bedside monitoring devices is obtained.

[0097] Furthermore, in actual intraoperative application, the brain network state representation vector at the current moment is input at regular intervals (such as 5 seconds or 10 seconds), and the network outputs the corresponding anesthesia depth index after forward inference calculation.

[0098] Since the brain network state representation vector itself already contains individualized brain functional connectivity information, this evaluation result has better individual adaptability compared to traditional methods based on fixed frequency band power thresholds.

[0099] Meanwhile, because brain network features are highly robust to signal noise, this assessment method can still maintain relatively stable output even if it is affected by electrocautery interference or electrode displacement during surgery.

[0100] S4. The deviation between the current anesthesia depth index and the target anesthesia depth range is used as the state input reinforcement learning agent, and the reinforcement learning agent outputs the drug administration rate adjustment command according to the current state.

[0101] In step S4, the difference between the current anesthesia depth index and the median value of the target interval is calculated as the basic deviation value;

[0102] Calculate the rate of change of the current anesthesia depth index relative to the previous time step;

[0103] The basic deviation value and the rate of change are combined into a two-dimensional state vector and input into the reinforcement learning agent.

[0104] The reinforcement learning agent adopts an actor-critic architecture, where the critic network is used to evaluate the expected cumulative reward of executing the dosing rate adjustment instruction in the current state, and the actor network is used to output the dosing rate adjustment instruction according to the current state.

[0105] The actor network and critic network update their network parameters online based on changes in the actual depth of anesthesia index after each dosing rate adjustment command is executed.

[0106] The drug delivery rate adjustment command includes the adjustment direction and adjustment range;

[0107] The adjustment direction is to increase the infusion rate, decrease the infusion rate, or maintain the current rate;

[0108] The adjustment range is positively correlated with the magnitude of the base deviation value, and when the direction of the rate of change is opposite to the direction of the base deviation value, the adjustment range is multiplied by a decay coefficient less than 1.

[0109] After the target-controlled infusion device is driven to perform drug infusion according to the drug delivery rate adjustment command, it also includes:

[0110] Record the anesthesia depth index response value after each administration rate adjustment command is executed, and construct an administration-response time sequence record;

[0111] When two consecutive dosing rate adjustment commands are in opposite directions, the dosing-response timing record is marked as an oscillation event, and the upper limit of the adjustment amplitude of the subsequent reinforcement learning agent output is reduced.

[0112] The target anesthesia depth range is preset based on the patient's basic information and the type of surgery, and the boundary value of the target anesthesia depth range is dynamically adjusted during the operation based on the long-term trend of the brain network state representation vector.

[0113] Specifically, after obtaining the current anesthesia depth index, this invention compares it with a preset target anesthesia depth range and uses the deviation value as a state input to the reinforcement learning agent.

[0114] The specific procedure is as follows: First, a target anesthesia depth range is preset, which reflects the range of anesthesia depth that the patient is expected to maintain during the current surgical stage.

[0115] During the initial setup, the range is retrieved from a pre-stored clinical knowledge base based on the patient's basic information (such as age, weight, ASA classification) and the type of surgery (such as cardiac surgery, abdominal surgery, orthopedic surgery, etc.). For example, for a typical adult patient undergoing abdominal surgery, the target range can be set between 0.4 and 0.7.

[0116] During the actual control process during the operation, the system collects the current anesthesia depth index in real time and calculates the difference between it and the median value of the target interval as the basic deviation value; at the same time, it records the rate of change of the current anesthesia depth index relative to the previous moment, that is, the amount of change of the anesthesia depth index per unit time.

[0117] The baseline deviation value and the rate of change are combined into a two-dimensional state vector. This two-dimensional state vector is a compact description of the current anesthesia state:

[0118] The baseline deviation indicates the urgency and direction of the intervention (whether it needs to be deepened or lightened), while the rate of change indicates trend information (the direction and speed at which the depth of anesthesia is drifting). The combination of these two factors enables the reinforcement learning agent to simultaneously perceive "where it is now" and "where it is going," thereby making more rational intervention decisions.

[0119] Subsequently, the reinforcement learning agent outputs instructions to adjust the drug delivery rate based on the current state. The reinforcement learning agent preferably employs an actor-critic architecture within the framework of a deep deterministic policy gradient algorithm.

[0120] The critic network is used to evaluate the expected cumulative return of performing a specific action in the current state. Its output is a scalar value, representing the expected long-term return obtained by continuously adjusting according to the current strategy from the current state. The actor network, on the other hand, directly maps specific dosing rate adjustment instructions based on the current state.

[0121] During the training phase, the critic network provides the policy gradient direction to the actor network, guiding the actor network to update parameters in the direction that can obtain higher cumulative rewards.

[0122] Furthermore, during the actual operation phase, after each execution of the drug administration rate adjustment command, the system performs lightweight online updates to the parameters of the actor network and the critic network based on changes in the actual anesthesia depth index.

[0123] Specifically, the state before execution, the action performed, the reward value observed after execution, and the new state after execution are combined into a four-tuple experience sample and stored in the experience replay pool; every set number of steps, a batch of samples is randomly sampled from the experience replay pool to perform mini-batch gradient updates of the network parameters.

[0124] This online learning mechanism enables the agent to continuously adapt to the individual pharmacodynamic characteristics of the patient during surgery, achieving truly personalized closed-loop control. The dosing rate adjustment command specifically includes two elements: adjustment direction and adjustment magnitude.

[0125] The adjustment direction can be one of three: increasing the infusion rate, decreasing the infusion rate, or maintaining the current rate. The adjustment magnitude is positively correlated with the magnitude of the baseline deviation value, that is, the larger the deviation, the larger the adjustment magnitude. At the same time, when the direction of the rate of change of state is opposite to the direction of the baseline deviation value (such as the depth of anesthesia being lower than the target value but recovering on its own), the adjustment magnitude is multiplied by a decay coefficient less than 1 to avoid over-adjustment.

[0126] S5. Drive the target-controlled infusion device to perform drug infusion according to the drug delivery rate adjustment command, and continuously collect a new round of multi-lead EEG signals to form closed-loop monitoring;

[0127] In step S5, the drug delivery rate adjustment command output by the reinforcement learning agent is sent to the interface module of the target-controlled infusion device. The target-controlled infusion device adjusts the infusion parameters according to the received command, driving the infusion pump to perform drug delivery at the corresponding rate.

[0128] The preferred drugs are commonly used anesthetics such as propofol or sevoflurane. The infusion rate is adjusted within a predetermined safety limit based on the type of drug and the patient's weight to prevent the infusion rate from exceeding the upper or lower safety limit due to abnormal instructions.

[0129] After the drug infusion parameters are adjusted, the blood concentration of the anesthetic drug in the patient's body gradually changes, which in turn causes characteristic changes in the electroencephalogram (EEG) signal. This change will be detected in the subsequent EEG acquisition process.

[0130] Specifically, the system continuously collects a new round of multi-lead EEG signals at a preset sampling period, and repeats the aforementioned steps S1 to S4, namely, collecting signals, calculating the orientation causal strength value, updating the brain functional network, extracting graph theory features, evaluating the anesthesia depth index, and reinforcing the learning agent to make new decisions, forming a complete closed-loop monitoring and control loop.

[0131] The optimal update frequency for this closed-loop circuit is once every 10 to 30 seconds, which ensures the system's timely response to changes in the anesthetic state and allows sufficient physiological time for drug distribution and onset of action in the body.

[0132] Through the aforementioned closed-loop mechanism, the system can continuously track changes in the patient's anesthesia depth throughout the entire surgical procedure and automatically adjust the drug infusion rate to maintain the anesthesia depth within the target range.

[0133] When the intensity of surgical stimulation changes (such as skin incision, visceral traction, etc.), the brain functional network connectivity reflected in the electroencephalogram (EEG) will change rapidly, and the anesthesia depth index will change accordingly. The reinforcement learning agent will quickly perceive the deviation and give corresponding adjustment instructions, thereby effectively preventing adverse events such as intraoperative awareness or excessive anesthesia.

[0134] Meanwhile, the system monitors and records the response to the depth of anesthesia after each administration. When two consecutive administration rate adjustment commands are in opposite directions, it indicates that the system may be in an oscillating control state. At this time, the system automatically marks it as an oscillation event and reduces the upper limit of the adjustment amplitude of the subsequent reinforcement learning agent output to avoid excessively frequent reverse adjustments.

[0135] Furthermore, the target anesthesia depth range can be dynamically adjusted during surgery based on the long-term trend of the brain network state representation vector.

[0136] For example, when the feature index representing the ability to integrate information in the brain network representation vector continues to drift in a positive direction, the system appropriately relaxes the boundary value of the target interval to make the anesthesia maintenance plan more adaptable.

[0137] Conversely, when characteristic indicators show abnormal fluctuations, the system appropriately tightens the boundary values ​​to enhance the alertness and responsiveness of the control. This dynamic adjustment mechanism enables the system to adapt to the differentiated needs for anesthesia depth at different surgical stages, further improving the flexibility and safety of clinical applications.

[0138] As shown above, this invention calculates the directional transfer function between lead pairs based on multi-lead EEG signals, constructing a directed weighted brain functional network that reflects the true flow of information. The extracted global graph theory features can quantify the efficiency of whole-brain information integration, while local graph theory features can locate key hub nodes for the action of anesthetic drugs. Finally, the brain network state representation vector formed by the fusion through the attention mechanism fully preserves multi-dimensional information such as the direction, intensity, and path efficiency of information flow between brain regions. This technical solution enables the system to perceive the decrease in communication efficiency between brain regions, the simplification of information transmission paths, and the inhibition of hub node functions—all direct physiological manifestations of drug action at the whole-brain network level. Compared to traditional spectral features that only reflect local activity, the perception dimension of this invention is closer to the neurobiological essence of loss of consciousness caused by anesthetic drugs, providing a richer and more reliable foundation of state information for subsequent precise regulation.

[0139] This invention combines the deviation value and rate of change of the anesthesia depth index from the target range into a state vector input reinforcement learning agent. Through an actor-critic architecture, the agent continuously updates its strategy online based on actual drug administration responses during the surgical process. Each surgery and each drug administration feedback optimizes subsequent decision-making quality, thus enabling the system to adapt in real-time to differences in drug efficacy among different patients and changes in intraoperative stimuli. Based on this, the drug administration rate adjustment command output by the agent directly drives the target-controlled infusion device to execute drug infusion, and a new round of EEG acquisition is then triggered, forming a closed-loop circuit. This mechanism automates the entire process of maintaining the anesthesia depth within the target range, significantly shortening the response time after state deviation and reducing the probability of adverse events such as intraoperative awareness and excessive anesthesia.

[0140] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for monitoring clinical anesthesia depth data based on multi-parameter fusion, characterized in that, include: Multi-lead EEG signals were acquired, and time-frequency transformation and inter-lead causal analysis were performed on the multi-lead EEG signals to obtain the orientation causal strength values ​​between each lead pair. A time-varying weighted brain functional network is constructed based on the directional causal strength value, and the global graph theory features and local graph theory features of the time-varying weighted brain functional network are extracted. The brain network state representation vector is obtained by fusing the global graph theory features and local graph theory features. The brain network state representation vector is input into the trained anesthesia depth assessment network, and the current anesthesia depth index is output. The deviation between the current anesthesia depth index and the target anesthesia depth range is used as the state input to the reinforcement learning agent, which then outputs a drug delivery rate adjustment command based on the current state. The target-controlled infusion device is driven to perform drug infusion according to the drug delivery rate adjustment command, and a new round of multi-lead EEG signals is continuously collected to form a closed-loop monitoring.

2. The clinical anesthesia depth data monitoring method based on multi-parameter fusion according to claim 1, characterized in that, The step of performing time-frequency transformation and inter-lead causal analysis on the multi-lead EEG signals to obtain the orientational causal strength values ​​between each lead pair includes: Multi-scale wavelet packet decomposition was performed on the EEG signal of each lead to extract the time-frequency components at each scale. For each pair of leads, the orientation transfer function value of the time-frequency component of one lead to the time-frequency component of another lead was calculated. The orientation transfer function values ​​at all scales were then weighted and fused according to the frequency contribution weight to obtain the orientation causal strength value between the lead pairs.

3. The clinical anesthesia depth data monitoring method based on multi-parameter fusion according to claim 2, characterized in that, The value of the directional transfer function is determined in the following way: A vector autoregression model is constructed using the historical time-frequency sequence of lead 1 and the current time-frequency sequence of lead 2. The variance contribution ratio of lead 2 at the current time to the historical time-frequency value of lead 1 is calculated, and this variance contribution ratio is used as the directional transfer function value from lead 1 to lead 2.

4. The clinical anesthesia depth data monitoring method based on multi-parameter fusion according to claim 1, characterized in that, The construction of a time-varying weighted brain functional network based on the directional causal strength value, and the extraction of global and local graph theory features of the time-varying weighted brain functional network, includes: A directed weighted network is constructed using each lead as a network node and the directed edge weights as directional causal strength values. The directed weighted network is dynamically updated within a preset time window; The average shortest path length and global clustering coefficient of the entire network are calculated as global graph theory features, and the betweenness centrality and in-degree strength of each node are calculated as local graph theory features.

5. The clinical anesthesia depth data monitoring method based on multi-parameter fusion according to claim 1, characterized in that, The brain network state representation vector obtained by fusing the global graph theory features and local graph theory features includes: The betweenness centrality and in-degree strength of each node are encoded and concatenated into a local feature vector according to the node position, and the average shortest path length and the global clustering coefficient are concatenated into a global feature vector. An attention mechanism is used to calculate the contribution weight of the global feature vector to each node in the local feature vector. The local feature vector is then weighted and aggregated according to the contribution weight and concatenated with the global feature vector to obtain the brain network state representation vector.

6. The clinical anesthesia depth data monitoring method based on multi-parameter fusion according to claim 1, characterized in that, The step of using the deviation between the current anesthesia depth index and the target anesthesia depth range as the state input to the reinforcement learning agent includes: The difference between the current anesthesia depth index and the median value of the target interval is calculated as the baseline deviation value; Calculate the rate of change of the current anesthesia depth index relative to the previous time step; The basic deviation value and the rate of change are combined into a two-dimensional state vector and input into the reinforcement learning agent.

7. The clinical anesthesia depth data monitoring method based on multi-parameter fusion according to claim 6, characterized in that, The reinforcement learning agent adopts an actor-critic architecture, where the critic network is used to evaluate the expected cumulative reward of executing the dosing rate adjustment instruction in the current state, and the actor network is used to output the dosing rate adjustment instruction according to the current state. After each dosing rate adjustment command is executed, the actor network and critic network update their network parameters online based on changes in the actual anesthesia depth index.

8. The clinical anesthesia depth data monitoring method based on multi-parameter fusion according to claim 6, characterized in that, The drug delivery rate adjustment command includes the adjustment direction and adjustment range; The adjustment direction is to increase the infusion rate, decrease the infusion rate, or maintain the current rate; The adjustment range is positively correlated with the magnitude of the base deviation value, and when the direction of the rate of change is opposite to the direction of the base deviation value, the adjustment range is multiplied by a decay coefficient less than 1.

9. The clinical anesthesia depth data monitoring method based on multi-parameter fusion according to claim 1, characterized in that, After the target-controlled infusion device is driven to perform drug infusion according to the drug delivery rate adjustment command, it also includes: Record the anesthesia depth index response value after each administration rate adjustment command is executed, and construct an administration-response time sequence record; When two consecutive dosing rate adjustment commands are in opposite directions, the dosing-response timing record is marked as an oscillation event, and the upper limit of the adjustment amplitude of the subsequent reinforcement learning agent output is reduced.

10. A clinical anesthesia depth data monitoring device based on multi-parameter fusion, characterized in that, include: The EEG acquisition and causal analysis module is used to acquire multi-lead EEG signals, perform time-frequency transformation and inter-lead causal analysis on the multi-lead EEG signals, and obtain the directional causal strength value between each lead pair; The brain network construction and feature extraction module is used to construct a time-varying weighted brain functional network based on the directional causal strength value, extract the global graph theory features and local graph theory features of the time-varying weighted brain functional network, and fuse the global graph theory features and local graph theory features to obtain the brain network state representation vector. An anesthesia depth assessment module is used to input the brain network state representation vector into the trained anesthesia depth assessment network and output the current anesthesia depth index. The reinforcement learning decision-making module is used to take the deviation between the current anesthesia depth index and the target anesthesia depth range as the state input to the reinforcement learning agent, and the reinforcement learning agent outputs the drug administration rate adjustment command according to the current state. The closed-loop control execution module is used to drive the target-controlled infusion device to perform drug infusion according to the drug administration rate adjustment command, and to continuously collect a new round of multi-lead EEG signals to form closed-loop monitoring.

Images