Anesthesia effect prediction method based on graph neural network

Abstract

The invention discloses an anesthesia effect prediction method based on a graph neural network. The method comprises the following steps: obtaining a standardized synchronization time sequence data set; forming a time sliding window sequence; generating an initial topology of the dynamic physiological index-drug effect variable diagram structure; generating a node initial feature representation set for each graph node, and initializing a graph neural network model parameter set; obtaining an updated dynamic physiological index-drug effect variable diagram structure; inputting the pharmacodynamic variable graph structure into a graph neural network model corresponding to the graph neural network model parameter set to obtain a node embedding representation set and an edge embedding representation set; outputting an induction period anesthesia effect prediction result set; and generating an interpretable output result set, wherein the interpretable output result set comprises a key variable link, a dominant influence factor and a drug administration adjustment suggestion. The rapid detection capability of the abnormal state in the induction period is improved.

Description

Technical Field

[0001] This invention relates to the field of induction anesthesia technology, and in particular to a method for predicting anesthetic effects based on graph neural networks. Background Technology

[0002] With the popularization and intelligent advancement of perioperative monitoring equipment, multimodal physiological signals and drug administration information of patients during the anesthesia induction period can be collected at high frequency and continuously. Clinically, there is an urgent need to improve the ability to predict the depth of anesthesia and risks based on big data and intelligent modeling.

[0003] Currently, most mainstream methods for evaluating the effect of anesthesia during the induction period rely on a single indicator or use machine learning models with static feature weights. These models treat various physiological parameters as independent of each other or use fixed correlation coefficients for modeling, ignoring the instantaneous coupling and strong nonlinear changes between various physiological indicators during the induction period after drug injection. Traditional models are difficult to depict the dramatic fluctuations in the correlation between indicators under the influence of drug efficacy, resulting in a delayed response to sudden events and insufficient ability to handle inter-individual differences. Summary of the Invention

[0004] One objective of this invention is to propose a method for predicting anesthetic effects based on graph neural networks, which improves the ability to rapidly detect abnormal states during the induction period.

[0005] A method for predicting anesthetic effects based on graph neural networks according to an embodiment of the present invention includes the following steps:

[0006] The raw data sets of multimodal physiological monitoring and drug administration records were collected and preprocessed to obtain a standardized and synchronized time-series data set;

[0007] A time sliding window is constructed based on a standardized and synchronized time series data set. The data segment within each time sliding window is used as the input segment for a single dynamic graph modeling, forming a time sliding window sequence.

[0008] Based on the variable change trend, co-fluctuation relationship or statistical coupling characteristics within each time sliding window sequence, the initial topology of the dynamic physiological index-pharmacological variable graph structure is generated.

[0009] On the initial topology of the dynamic physiological index-pharmacological variable graph structure, a set of learnable parameters for edge weights is assigned to each graph edge, and an initial feature representation set for each graph node is generated, while the parameter set of the graph neural network model is initialized.

[0010] The set of learnable parameters for edge weights is adaptively updated on a continuous time sliding window using a standardized and synchronized time series data set. When a sudden change in the correlation is detected, a topological update is performed on the dynamic physiological index-pharma effect variable graph structure to obtain the updated dynamic physiological index-pharma effect variable graph structure.

[0011] The updated dynamic physiological indicators-pharmacodynamic variables graph structure is input into the graph neural network model corresponding to the graph neural network model parameter set to obtain the node embedding representation set and the edge embedding representation set.

[0012] The spatiotemporal fusion feature representation set is generated by combining the node embedding representation set and the edge embedding representation set with the time series features in the standardized and synchronized time series data set. The spatiotemporal fusion feature representation set is input into the multi-task anesthesia effect prediction engine and outputs the induction period anesthesia effect prediction result set. The induction period anesthesia effect prediction result set includes the expected time to reach the target anesthesia depth, the probability of excessive anesthesia risk, the hemodynamic fluctuation risk score and other anesthesia-related outcome indicators.

[0013] Based on the predicted results of the induction period anesthesia effect and the edge weight distribution in the updated dynamic physiological indicators-pharmacological variables graph structure, an interpretable output set is generated. The interpretable output set includes key variable links, dominant influencing factors, and dosing adjustment suggestions.

[0014] Optionally, obtaining the standardized synchronized time-series data set includes:

[0015] The study collected raw data sets of multimodal physiological monitoring and drug administration records from patients before anesthesia induction and intubation. These raw data sets included bispectral index (BPI), EEG signal, systolic blood pressure, diastolic blood pressure, mean arterial pressure, heart rate, blood oxygen saturation, end-tidal carbon dioxide concentration, respiratory rate, and information on the type, dosage, and infusion rate of intravenously administered drugs. Simultaneously, the study collected patients' static baseline characteristics. The raw data sets of multimodal physiological monitoring and drug administration records were processed for time synchronization, noise removal, missing data completion, and unit standardization to obtain a standardized synchronized time-series data set. The standardized synchronized time-series data set maintained a one-to-one correspondence with the patients' static baseline characteristics.

[0016] Optionally, the formation of the time sliding window sequence includes:

[0017] Set the time sliding window length parameter and the sliding step size parameter;

[0018] The standardized synchronized time series data set is segmented based on the time sliding window length parameter and the sliding step parameter to form multiple time sliding windows in sequence;

[0019] Within each time sliding window, all multivariate feature vectors covered by the time sliding window are sequentially combined into a data segment;

[0020] Within each time sliding window, record all time indices covered by the time sliding window;

[0021] All data segments corresponding to the time sliding window are combined with the time index set in sequence to form a time sliding window sequence representation set.

[0022] Optionally, the generation of the initial topology of the dynamic physiological indicator-pharmacological variable graph structure includes:

[0023] Extract the data segment corresponding to each time sliding window sequence;

[0024] Each physiological monitoring variable, drug variable, and patient static baseline characteristic in the data segment is constructed as a graph node in the graph structure, and all graph nodes constitute a graph node set.

[0025] The cooperative change relationship characteristics of the corresponding variables of each pair of graph nodes in the graph node set are calculated within the current time sliding window. The cooperative change relationship characteristics are constructed by the change rate sequence of the variables and the variables of the previous moment within the time sliding window. The statistical coupling coefficient is used to represent the cooperative fluctuation relationship of the graph node pair within the k-th time sliding window.

[0026] Based on the statistical coupling coefficient, it is determined whether to establish a graph edge connection in the graph structure for each graph node pair. When the absolute value of the statistical coupling coefficient is greater than the set edge connection threshold, a graph edge connection is established between the graph node pairs. All graph node pairs that meet the conditions constitute a graph edge set.

[0027] By combining the set of graph nodes and the set of graph edges, the initial topology of the dynamic physiological indicators-pharmacological variables graph structure within the k-th time sliding window is generated.

[0028] Optionally, the initialization graph neural network model parameter set includes:

[0029] The initial topology of the dynamic physiological index-pharmacodynamic variable graph structure constructed within the k-th time sliding window is processed, and learnable edge weight parameters are introduced on each graph edge. The initial values ​​of the learnable edge weight parameters are jointly determined by the physiological coupling prior during the induction period and the local response regulatory factors.

[0030] Calculate the standard deviation and mean of the variable within the current time sliding window, and obtain the local response regulation factor of the variable pair by weighting the response heterogeneity index before and after the drug injection time point.

[0031] Learnable edge weight parameters are used as the initial weights of graph edges in the graph structure, and form a set of learnable edge weight parameters.

[0032] Based on the time-series performance of variables of each graph node in the graph node set within the current time sliding window, an initial feature representation vector of the node is constructed, and all initial feature representation vectors of the nodes form the initial feature representation set of the nodes.

[0033] Initialize the graph neural network model parameter set by inputting the set of learnable edge weight parameters, the set of initial node feature representations, and the graph neural network model parameter set into the graph neural network model.

[0034] Optionally, the obtained updated dynamic physiological index-pharmacodynamic variable graph structure includes:

[0035] Within each time sliding window, the standardized synchronized time series data set corresponding to the current time sliding window is used, combined with the learnable parameter set of edge weights and the initial feature representation set of nodes of the graph neural network model, to perform graph information propagation operation and obtain the edge representation output of each graph edge;

[0036] For each edge in the graph, the edge weight is updated by calculating the output result. The updated edge weights are then collected and updated to form an edge weight update set.

[0037] Determine whether there is a structural abrupt change based on the magnitude of the change in each edge weight in the edge weight update set;

[0038] Perform a topology update operation on graph edges that meet the structural mutation criteria:

[0039] When the weight of a graph edge increases after an update and there is no such edge connecting the current graph structure, a new graph edge is added between the corresponding graph node pairs.

[0040] If the weight of a graph edge decreases after an update and there are connecting edges in the current graph structure, remove that graph edge from the graph structure.

[0041] Apply all topology update operations to the initial topology of the original graph structure to generate a dynamic physiological indicator-drug efficacy variable graph structure with updated topology.

[0042] The updated graph structure retains the same set of graph nodes, while the graph edge set is filtered based on whether the updated edge weights exceed the edge retention threshold, ultimately resulting in a new graph edge set.

[0043] Optionally, the set of predicted results for the induction period anesthesia includes:

[0044] Within each time sliding window, the graph neural network model is used to calculate the node embedding representation set and the edge embedding representation set of the dynamic physiological index-pharmacodynamic variable graph structure after the topology is updated;

[0045] Extract the time series features of the standardized and synchronized time series data set within the current time sliding window;

[0046] The node embedding representation vector of each graph node is concatenated with its time series feature vector to obtain the spatiotemporal fusion feature representation of a single node.

[0047] The spatiotemporal fusion feature representations of all graph nodes are combined sequentially into a spatiotemporal fusion feature representation set. At the same time, the graph edge embedding representation set and the spatiotemporal fusion feature set of graph nodes are used as inputs to the multi-task anesthesia effect prediction engine.

[0048] In the multi-task anesthesia effect prediction engine, a multi-branch prediction task is performed on the input graph node spatiotemporal fusion feature set and graph edge embedding representation set. The output set of induction period anesthesia effect prediction results includes the estimated time to reach the target depth of anesthesia, the probability of excessive anesthesia risk, the hemodynamic fluctuation risk score, and other anesthesia-related outcome indicators.

[0049] Optionally, the output of the multi-task anesthesia effect prediction engine includes:

[0050] For each time sliding window, the spatiotemporal fusion feature representation set and the graph edge embedding representation set are processed through the shared feature extraction module to obtain the global fusion expression vector;

[0051] The global fusion expression vector is input into the multi-task prediction branch, each multi-task prediction branch corresponds to a sub-task of anesthetic effect, and each branch outputs the corresponding prediction result.

[0052] For each subtask, the corresponding prediction result and its corresponding true label are output, and the single-task loss is calculated. The weighted combination is then used to form a joint loss function.

[0053] The gradient descent optimization algorithm is used to iteratively update the parameters of the multi-task anesthesia effect prediction engine, so that the value of the joint loss function continuously decreases in all time sliding windows, and finally obtains a set of real-time updated prediction results of induction period anesthesia effect.

[0054] Optionally, the generation of the interpretable output result set includes:

[0055] Based on the set of predicted results of anesthetic effect during induction and the edge weight distribution in the updated dynamic physiological index-pharmacological variable graph structure, edges with edge weight values ​​higher than the set edge weight threshold are extracted to determine the key variable link set. The key variable link set is the set of directed paths of physiological monitoring variables and drug variables within the current time sliding window, where the edge weight value is greater than or equal to the edge weight threshold and is significantly associated with the trend of change of the predicted results of the target anesthetic effect.

[0056] Based on the rate of change, predicted output variation, and edge weight values ​​of each variable in the key variable link set within the spatiotemporal fusion feature representation set, dominant influencing factors are defined. Dominant influencing factors are variables that satisfy the following conditions:

[0057] The dominant influencing factors are the set of variables whose rate of change in the key variable link is higher than the rate of change threshold and is within the range of change in the predicted anesthetic effect, and whose edge weight values ​​are in the top 30% of all links.

[0058] Based on the key variable chain set, the dominant influencing factors, and the prediction results of the induction period anesthesia effect, combined with the side weight trend of drug variables and the status of the dominant influencing factors, dosing adjustment suggestions are generated.

[0059] Optionally, the dosing adjustment recommendation is a recommendation that meets the following conditions:

[0060] When the increase in the margin weight between the drug variable and the dominant influencing factor is greater than the increase threshold in a continuous window, and an increase in the risk of anesthesia depth is predicted at the same time, a recommendation to reduce the drug dosage is given.

[0061] When the decrease in the marginal weight between the drug variable and the dominant influencing factor is greater than the decrease threshold, and the depth of anesthesia is lower than the target threshold, a suggestion to appropriately increase the drug dosage is given.

[0062] When the side weights of physiological monitoring variables change abruptly and are associated with the risk of adverse events, a dynamic warning is given.

[0063] The beneficial effects of this invention are:

[0064] (1) This invention realizes real-time perception and expression of relationship mutations based on dynamic graphs of physiological indicators and drug efficacy variables. By introducing a dynamic graphing mechanism during the anesthesia induction period, the topological structure between physiological monitoring variables, drug variables and patient static basic characteristics is reconstructed in real time by a time sliding window. The statistical coupling feature and the edge weight dynamic adjustment algorithm enable the model to keenly capture sudden changes in the correlation between various physiological indicators after drug injection. This effectively solves the clinical need for traditional static models to reflect relationship mutations and nonlinear couplings driven by drug efficacy in real time, and improves the ability to quickly detect abnormal states during the induction period.

[0065] (2) This invention proposes a margin learning method that integrates historical priors and individualized dynamic regulation to achieve adaptive optimization of margins under individual differences and multi-drug combination scenarios. By integrating the physiological coupling priors of historical large clinical sample statistics with the current real-time local response regulation factors of patients, a steady-state prior plus dynamic perturbation dual-channel margin initialization mechanism is constructed. Combined with the adaptive update of learnable margin parameters within the window and mutation detection, it effectively addresses the differences in physiological response under different patient baseline states and drug combinations, so that the model margins can be dynamically fine-tuned according to individual characteristics and clinical feedback, which significantly improves the generalization and clinical transferability of the model.

[0066] (3) This invention realizes spatiotemporal fusion prediction and interpretable clinical auxiliary output with multiple branches and multiple outcomes. By expressing the spatiotemporal fusion features of nodes and edges, dynamic graph encoding and multivariate time series information are jointly input into the multi-task prediction engine, which can simultaneously output multiple anesthesia-related outcome indicators and generate real-time interpretable output and dosing adjustment suggestions based on dynamic edge weight distribution and dominant variable links. Attached Figure Description

[0067] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0068] Figure 1 This is a flowchart of a method for predicting anesthetic effects based on graph neural networks proposed in this invention. Detailed Implementation

[0069] Example 1:

[0070] refer to Figure 1 A method for predicting anesthetic effects based on graph neural networks includes the following steps:

[0071] The raw data sets of multimodal physiological monitoring and drug administration records were collected and preprocessed to obtain a standardized and synchronized time-series data set;

[0072] In this embodiment, a standardized and synchronized time-series data set is obtained, including:

[0073] The study collected raw data sets of multimodal physiological monitoring and drug administration records from patients before anesthesia induction and intubation. These raw data sets included bispectral index (BPI), EEG signal, systolic blood pressure, diastolic blood pressure, mean arterial pressure, heart rate, blood oxygen saturation, end-tidal carbon dioxide concentration, respiratory rate, and information on the type, dosage, and infusion rate of intravenously administered drugs. Simultaneously, the study collected patients' static baseline characteristics. The raw data sets of multimodal physiological monitoring and drug administration records were processed for time synchronization, noise removal, missing data completion, and unit standardization to obtain a standardized synchronized time-series data set. The standardized synchronized time-series data set maintained a one-to-one correspondence with the patients' static baseline characteristics.

[0074] In this embodiment, the formation of the time sliding window sequence includes:

[0075] Set the time sliding window length parameter and the sliding step size parameter;

[0076] The time sliding window length parameter controls the continuous time length covered by each time sliding window, while the sliding step size parameter controls the time offset between adjacent time sliding windows. Together, the time sliding window length parameter and the sliding step size parameter determine the degree of window overlap.

[0077] The standardized synchronized time series data set is segmented based on the time sliding window length parameter and the sliding step parameter to form multiple time sliding windows in sequence;

[0078] Each time sliding window corresponds to a starting time point, each time sliding window covers a continuous time period, the distance between the starting time points of each time sliding window is equal to the sliding step parameter, and the total number of time sliding windows is equal to starting from the starting time point of the standardized synchronized time series data set, moving backward by one sliding step parameter each time, until the last time sliding window can still completely contain a data segment with a length equal to the time sliding window length parameter.

[0079] Within each time sliding window, all multivariate feature vectors covered by the time sliding window are sequentially combined into a data segment;

[0080] Each data segment includes all standardized, synchronized multivariate time-series information within a time sliding window.

[0081] Within each time sliding window, record all time indices covered by the time sliding window;

[0082] The time index set is used to maintain the time correspondence between the node embedding representation set and the edge weight learnable parameter set in dynamic graph modeling, so that the data segment in each time sliding window corresponds to a specific time period.

[0083] All data segments corresponding to the time sliding window are combined with the time index set in sequence to form a time sliding window sequence representation set.

[0084] A time sliding window is constructed based on a standardized and synchronized time series data set. The data segment within each time sliding window is used as the input segment for a single dynamic graph modeling, forming a time sliding window sequence.

[0085] In this embodiment, the generation of the initial topology of the dynamic physiological index-pharmacological variable graph structure includes:

[0086] Extract the data segment corresponding to each time sliding window sequence;

[0087] The data segment consists of multivariate feature vectors at multiple consecutive time points. Each multivariate feature vector includes all physiological monitoring variables, drug variables, and patient static baseline characteristics.

[0088] Each physiological monitoring variable, drug variable, and patient static baseline characteristic in the data segment is constructed as a graph node in the graph structure, and all graph nodes constitute a graph node set.

[0089] The set of graph nodes remains consistent within each time sliding window, ensuring that each variable can uniquely correspond to a graph node.

[0090] The cooperative change relationship characteristics of the corresponding variables of each pair of graph nodes in the graph node set are calculated within the current time sliding window. The cooperative change relationship characteristics are constructed by the change rate sequence of the variables and the variables of the previous moment within the time sliding window. The statistical coupling coefficient is used to represent the cooperative fluctuation relationship of the graph node pair within the k-th time sliding window.

[0091] In Example 1, the method for calculating the rate of change sequence is as follows: subtract the variable from the variable at the previous time step from the variable at the current time step, calculate the average of the rate of change sequences of the two variables, measure the deviation between the rate of change at each time step and the average, calculate the statistical co-relationship characteristics between the rate of change sequences of the two variables, and use the statistical co-relationship characteristics as the statistical coupling coefficient.

[0092]

[0093] in, Represents variable x (i) Compared with the variable x at the previous moment (j) The statistical coupling coefficient within the k-th time sliding window. and They represent the variables x respectively (i) Compared with the variable x at the previous moment (j) The mean rate of change within the current time sliding window and Let represent the rate of change of the i-th and j-th variables at time t, defined as the value of the variable at the current time minus its value at the previous time, where t is the variable's value at the current time. k L represents the index of the starting time point of the k-th time sliding window. w The parameter represents the length of the time sliding window, and t represents the time index. It iterates through all time points within the time sliding window.

[0094] Based on the statistical coupling coefficient, it is determined whether to establish a graph edge connection in the graph structure for each graph node pair. When the absolute value of the statistical coupling coefficient is greater than the set edge connection threshold, a graph edge connection is established between the graph node pairs. All graph node pairs that meet the conditions constitute a graph edge set.

[0095] By combining the set of graph nodes and the set of graph edges, the initial topology of the dynamic physiological indicators-pharmacological variables graph structure within the k-th time sliding window is generated.

[0096] Based on the variable change trend, co-fluctuation relationship or statistical coupling characteristics within each time sliding window sequence, the initial topology of the dynamic physiological index-pharmacological variable graph structure is generated.

[0097] In this embodiment, initializing the graph neural network model parameter set includes:

[0098] The initial topology of the dynamic physiological index-pharmacodynamic variable graph structure constructed within the k-th time sliding window is processed, and learnable edge weight parameters are introduced on each graph edge. The initial values ​​of the learnable edge weight parameters are jointly determined by the physiological coupling prior during the induction period and the local response regulatory factors.

[0099] The induction-phase physiological coupling prior reflects the steady-state coupling strength of variable pairs in a large historical clinical sample, derived from statistical distribution information in the labeled database. The local response regulation factor measures the short-term dynamic sudden coupling degree of variable pairs in the current real-time patient data after the induction-phase event. The initial values ​​of the edge-weighted learnable parameters are obtained by weighting the induction-phase physiological coupling prior function and the local response regulation factor according to the fusion weight coefficient. All edge-weighted learnable parameters constitute the edge-weighted learnable parameter set.

[0100] Calculate the standard deviation and mean of the variable within the current time sliding window. Weight the variable pair by the response heterogeneity index before and after the drug injection time point to obtain the local response regulation factor of the variable pair, reflecting the sensitivity to abnormal coupling of intraoperative drug-efficacy response.

[0101]

[0102] in, and They represent the variables x respectively (i) The standard deviation and mean, γ within the k-th time sliding window ij The response heterogeneity index for the variable pair (i,j) before and after the drug injection time point during the induction period.

[0103] Learnable edge weight parameters are used as the initial weights of graph edges in the graph structure, and form a set of learnable edge weight parameters.

[0104] Based on the time-series performance of variables of each graph node in the graph node set within the current time sliding window, an initial feature representation vector of the node is constructed, and all initial feature representation vectors of the nodes form the initial feature representation set of the nodes.

[0105] The node initial feature representation vector consists of the time-series statistical features of the variables, the rate of change, and the peak response within the window. The node initial feature representation set describes the basic state information of the node.

[0106] Initialize the graph neural network model parameter set by inputting the set of learnable edge weight parameters, the set of initial node feature representations, and the graph neural network model parameter set into the graph neural network model.

[0107] The graph neural network model parameter set includes all learnable weights, bias terms, and nonlinear activation function parameters used for node information propagation, edge weight fusion, and feature aggregation in the graph structure. The graph neural network model parameter set is coupled with the edge weight learnable parameter set and the node initial feature representation set.

[0108] On the initial topology of the dynamic physiological index-pharmacological variable graph structure, a set of learnable parameters for edge weights is assigned to each graph edge, and an initial feature representation set for each graph node is generated, while the parameter set of the graph neural network model is initialized.

[0109] In this embodiment, the updated dynamic physiological index-pharmacodynamic variable graph structure is obtained, including:

[0110] Within each time sliding window, the standardized synchronized time series data set corresponding to the current time sliding window is used, combined with the learnable parameter set of edge weights and the initial feature representation set of nodes of the graph neural network model, to perform graph information propagation operation and obtain the edge representation output of each graph edge;

[0111] For each edge in the graph, the edge weight is updated by calculating the output result. The updated edge weights are then collected and updated to form an edge weight update set.

[0112] In Example 1, the edge weight update calculation method is as follows: the learnable parameters of the edge weights of the graph edges in the current time sliding window, the edge representation output results, and the initial feature representation vectors of the graph nodes are input into the adaptive update function, multiplied by the edge weight update learning rate, and then summed to obtain the updated edge weights. The edge weight update set dynamically reflects the changes in the coupling relationship between variables in the graph in the current time sliding window.

[0113] The adaptive update function assigns different weights based on the correlation between the edge representation output and node features, as well as the fluctuation trend of the target variable in the induction anesthesia monitoring. The weight adjustment factor is multiplied by the edge weight update learning rate and then combined with the current learnable edge weight parameters to obtain the updated edge weights of the graph. The weight adjustment factor is set according to the following principles:

[0114] When an edge represents a highly synchronized trend between the output and the target variable and a stronger correlation with the rapid response of the drug variable, the weight adjustment factor of that edge is increased; otherwise, it is decreased. The thresholds for high synchronization and rapid response are set according to time requirements.

[0115] Determine whether there is a structural abrupt change based on the magnitude of the change in each edge weight in the edge weight update set;

[0116] In Example 1, the method for calculating the change range of edge weights is to find the absolute value of the difference between the current updated edge weight and the edge weight before the update. When the change range of edge weights is greater than the set mutation judgment threshold, the edge of the graph is determined to be in a state of sudden coupling structure change.

[0117] Perform a topology update operation on graph edges that meet the structural mutation criteria:

[0118] When the weight of a graph edge increases after an update and there is no such edge connecting the current graph structure, a new graph edge is added between the corresponding graph node pairs.

[0119] If the weight of a graph edge decreases after an update and there are connecting edges in the current graph structure, remove that graph edge from the graph structure.

[0120] Apply all topology update operations to the initial topology of the original graph structure to generate a dynamic physiological indicator-drug efficacy variable graph structure with updated topology.

[0121] The updated graph structure retains the same set of graph nodes, while the graph edge set is filtered based on whether the updated edge weights exceed the edge retention threshold, ultimately resulting in a new graph edge set.

[0122] The set of learnable parameters for edge weights is adaptively updated on a continuous time sliding window using a standardized and synchronized time series data set. When a sudden change in the correlation is detected, a topological update is performed on the dynamic physiological index-pharma effect variable graph structure to obtain the updated dynamic physiological index-pharma effect variable graph structure.

[0123] In this embodiment, the set of prediction results for the induction period anesthesia effect includes:

[0124] Within each time sliding window, the graph neural network model is used to calculate the node embedding representation set and the edge embedding representation set of the dynamic physiological index-pharmacodynamic variable graph structure after the topology is updated;

[0125] The node embedding representation set represents the embedding vector representation of each graph node within the current time sliding window, while the edge embedding representation set represents the embedding vector representation of each graph edge within the current time sliding window.

[0126] Extract the time series features of the standardized and synchronized time series data set within the current time sliding window;

[0127] Time series features include variable trends, short-term volatility, abrupt change slope, and historical state values ​​of the target indicator. Time series features represent the multidimensional time series characteristics of the variable corresponding to each graph node within the current window.

[0128] The node embedding representation vector of each graph node is concatenated with its time series feature vector to obtain the spatiotemporal fusion feature representation of a single node.

[0129] Spatiotemporal fusion features can jointly reflect the graph embedding features and temporal features of variables.

[0130] The spatiotemporal fusion feature representations of all graph nodes are combined sequentially into a spatiotemporal fusion feature representation set. At the same time, the graph edge embedding representation set and the spatiotemporal fusion feature set of graph nodes are used as inputs to the multi-task anesthesia effect prediction engine.

[0131] In the multi-task anesthesia effect prediction engine, a multi-branch prediction task is performed on the input graph node spatiotemporal fusion feature set and graph edge embedding representation set. The output set of induction period anesthesia effect prediction results includes the estimated time to reach the target depth of anesthesia, the probability of excessive anesthesia risk, the hemodynamic fluctuation risk score, and other anesthesia-related outcome indicators.

[0132] The multi-task anesthesia effect prediction engine adopts a joint loss optimization mechanism. While outputting multiple prediction results, it performs joint optimization by weighting factors according to the importance of different sub-tasks, so that the model achieves the optimal balance between clinical risk, prediction accuracy and real-time performance.

[0133] The updated dynamic physiological indicators-pharmacodynamic variables graph structure is input into the graph neural network model corresponding to the graph neural network model parameter set to obtain the node embedding representation set and the edge embedding representation set.

[0134] In this embodiment, the output of the multi-task anesthesia effect prediction engine includes:

[0135] For each time sliding window, the spatiotemporal fusion feature representation set and the graph edge embedding representation set are processed through the shared feature extraction module to obtain the global fusion expression vector;

[0136] The global fusion expression vector is input into the multi-task prediction branch, each multi-task prediction branch corresponds to a sub-task of anesthetic effect, and each branch outputs the corresponding prediction result.

[0137] The subtask of anesthetic effects includes a branch for predicting the time to achieve the target depth of anesthesia, a branch for predicting the probability of over-deep anesthesia, a branch for predicting the risk score of hemodynamic fluctuations, and a branch for predicting other anesthesia-related outcome indicators. The predicted result of the time to achieve the target depth of anesthesia is in seconds, the predicted result of the probability of over-deep anesthesia is in percentage, the predicted result of the risk score of hemodynamic fluctuations is in dimensionless score between 0 and 1, and the predicted result of other anesthesia-related outcome indicators is in dimensionless predicted value under the standard rating scale.

[0138] For each subtask, the corresponding prediction result and its corresponding true label are output, and the single-task loss is calculated. The weighted combination is then used to form a joint loss function.

[0139] The single-task loss for the time to achieve the target depth of anesthesia is obtained by calculating the absolute value of the difference between the predicted value and the actual value. The single-task loss for the probability of excessive anesthesia risk is obtained by calculating the cross-entropy loss. The single-task loss for the hemodynamic fluctuation risk score is obtained by calculating the square of the difference between the predicted value and the actual value. The single-task loss for other anesthesia-related outcome indicators is obtained by calculating the square of the Euclidean distance between the predicted vector and the actual vector. The weighting factor is used to dynamically balance the contribution of each sub-task to the joint loss. The joint loss function guides the multi-task prediction engine to achieve the optimal balance between multi-task prediction accuracy, balance and clinical needs.

[0140] The gradient descent optimization algorithm is used to iteratively update the parameters of the multi-task anesthesia effect prediction engine, so that the value of the joint loss function continuously decreases in all time sliding windows, and finally obtains a set of real-time updated prediction results of induction period anesthesia effect.

[0141] The spatiotemporal fusion feature representation set is generated by combining the node embedding representation set and the edge embedding representation set with the time series features in the standardized and synchronized time series data set. The spatiotemporal fusion feature representation set is input into the multi-task anesthesia effect prediction engine and outputs the induction period anesthesia effect prediction result set. The induction period anesthesia effect prediction result set includes the expected time to reach the target anesthesia depth, the probability of excessive anesthesia risk, the hemodynamic fluctuation risk score and other anesthesia-related outcome indicators.

[0142] In this embodiment, the generation of the interpretable output result set includes:

[0143] Based on the set of predicted results of induction period anesthesia effect and the edge weight distribution in the updated dynamic physiological index-pharmacological variable graph structure, edges with edge weight values ​​higher than the set edge weight threshold are extracted to determine the key variable link set. The key variable link set is the set of directed paths of physiological monitoring variables and drug variables within the current time sliding window, where the edge weight value is greater than or equal to the edge weight threshold and is significantly associated with the trend of change of the target anesthesia effect prediction result.

[0144] Based on the rate of change, predicted output variation, and edge weight values ​​of each variable in the key variable link set within the spatiotemporal fusion feature representation set, dominant influencing factors are defined. Dominant influencing factors are variables that satisfy the following conditions:

[0145] The dominant influencing factors are the set of variables whose rate of change in the key variable link is higher than the rate of change threshold and is within the range of change in the predicted anesthetic effect, and whose edge weight values ​​are in the top 30% of all links.

[0146] Based on the key variable chain set, the dominant influencing factors, and the prediction results of the induction period anesthesia effect, combined with the side weight trend of drug variables and the status of the dominant influencing factors, dosing adjustment suggestions are generated.

[0147] Based on the predicted results of the induction period anesthesia effect and the edge weight distribution in the updated dynamic physiological indicators-pharmacological variables graph structure, an interpretable output set is generated. The interpretable output set includes key variable links, dominant influencing factors, and dosing adjustment suggestions.

[0148] In this embodiment, the recommended dosing adjustments are recommendations that meet the following conditions:

[0149] When the increase in the margin weight between the drug variable and the dominant influencing factor is greater than the increase threshold in a continuous window, and an increase in the risk of anesthesia depth is predicted at the same time, a recommendation to reduce the drug dosage is given.

[0150] When the decrease in the marginal weight between the drug variable and the dominant influencing factor is greater than the decrease threshold, and the depth of anesthesia is lower than the target threshold, a suggestion to appropriately increase the drug dosage is given.

[0151] When the side weights of physiological monitoring variables change abruptly and are associated with the risk of adverse events, a dynamic warning is given.

[0152] Example 2:

[0153] The cardiothoracic surgery department of Hospital A performed a coronary artery bypass grafting surgery on a 63-year-old male patient weighing 76 kg. He had a history of hypertension and type 2 diabetes for more than 20 years and his recent cardiac function assessment was NYHA class II. Due to his advanced age and underlying diseases, there was a high risk of hemodynamic fluctuations during anesthesia induction and difficulties in managing deep and light anesthesia during the operation. The anesthesiologist used the system of this invention and compared it with the traditional BIS single threshold alarm and static regression model.

[0154] The anesthesia team connected the patient to multimodal monitoring equipment according to standard procedures, including BIS electroencephalography, non-invasive blood pressure, ECG, pulse oximetry, and end-tidal carbon dioxide. At the same time, the multimodal physiological monitoring and drug administration information synchronous acquisition module during induction was activated. The patient's induction regimen was propofol (1.5 mg / kg) and midazolam (0.04 mg / kg) intravenously, remifentanil target-controlled infusion, and atracurium relaxation was administered sequentially.

[0155] The data acquisition module records BIS, EEG, systolic blood pressure, diastolic blood pressure, MAP, heart rate, SpO2, EtCO2, respiratory rate, and actual drug dosage and rate from the start of intravenous injection to the completion of cannulation. It also combines the patient's static baseline characteristics such as BMI, age, gender, past chronic diseases, and drug allergies. All signals are recorded synchronously at 2-second intervals and automatically perform noise reduction, interpolation, and standardization to ensure data quality and consistency.

[0156] The model baseline window was established 10 seconds before the induction began, and the following data segments were collected (partial data examples):

[0157] Table 1 provides examples of some data.

[0158]

[0159] The system of this invention performs segmented processing on a standardized and synchronized time series data set using a 3-second sliding window (1-second step). Within each window, the system automatically generates dynamic graph nodes containing various physiological variables, drug variables, and basic features, as well as initial edge weights based on co-variables, through a dynamic graph construction algorithm. In the initial stage, most variable edge weights are low. The model detects an increase in drug rate, and MAP and BIS show signs of decline. Abrupt changes in edge weights indicate strong nonlinear coupling between drug and physiological variables, and the system automatically increases the edge weights.

[0160] During the 3-10 second period after drug injection, traditional static models and BIS single-threshold early warning systems failed to identify the impending sudden drop in MAP and a sharp decline in BIS. The dynamic graph neural network model of this invention, based on the dynamic edge weight adjustment and edge weight update mechanism, timely captures the strong coupling between BIS, MAP, and HR multivariates, and outputs through the prediction engine: the target depth of anesthesia is expected to be reached in 8 seconds, the probability of blood pressure drop is 78%, and it is recommended to prepare for blood pressure escalation intervention in advance. The static weight prediction error of the traditional method is as high as 22 mmHg, which is insufficient advance and only alarms after the actual blood pressure drop occurs; while the MAP fluctuation prediction advance of the model of this invention is an average of 6.7 seconds, and the error is controlled within 7 mmHg.

[0161] In the 15 seconds before intubation, the patient's MAP actually dropped by as much as 28 mmHg, and BIS dropped to 41. This invention accurately predicts the time point when the target BIS is reached, identifies the risk of deep anesthesia in advance, and automatically marks the decline in HR and mild abnormal fluctuations in SpO2 as the dominant influencing factors. The interpretable output is: Key variable link: propofol → MAP → BIS → HR, dominant influencing factor: dynamic changes in propofol dose and MAP, and suggests reducing propofol dose or staged intravenous injection.

[0162] The anesthesiologist adjusted the propofol injection rate in advance as advised, which promptly alleviated the downward trend in MAP and prevented significant hypoxia or sudden drop in heart rate.

[0163] The training set was constructed using real case data from two hospitals, Hospital A and Hospital B, from 2024 to 2025 (N=1075 cases). The independent test set consisted of 36 consecutive surgeries performed in the cardiac surgery and hepatobiliary surgery departments of Hospital A in March and April 2025. The performance of this invention compared with traditional BIS threshold warning and static regression methods in terms of abnormal blood pressure during induction, prediction of anesthesia depth, and early warning of intraoperative adverse events is shown in Table 2 below.

[0164] Table 2 Comparison data between the present invention and conventional inventions

[0165]

[0166] The specific case sampling data is as follows:

[0167] Case A, a 54-year-old female with a BMI of 23 and ASA II, was given propofol 1.7 mg / kg and remifentanil 0.09 μg / kg / min during induction. Her baseline BIS was 95. This invention alerted her to the excessive depth of anesthesia 8 seconds in advance and suggested slowing down the injection rate. Ultimately, her BIS dropped to 51. The traditional model did not provide an early alert, and her actual BIS dropped to 36, indicating a delayed low anesthesia risk alarm.

[0168] Case B, a 65-year-old male with a BMI of 28 and a history of diabetes, experienced a decrease in MAP from 112 to 82 and HR from 80 to 62 after induction. The model of this invention issued a low blood pressure risk warning 11 seconds in advance and suggested reducing the dosage of propofol and remifentanil. Traditional methods failed to identify this in time, and a brief period of hypoxia occurred before intubation.

[0169] In the training samples, the dynamic graph neural network model used 1075 historical surgical data to collect data at a 2-second resolution. Each case was divided into an average of 96 groups of windows, including node feature vectors, dynamic edge weights, and time series labels. The model adopted an 8:2 cross-validation ratio. The average mean square error of MAP fluctuation prediction on the training set was 9.3 mmHg, and on the test set it was 7.1 mmHg.

[0170] In the interpretability analysis, the accuracy rate of identifying key variable links and dominant influencing factors was higher than 87%, and the dosing adjustment recommendations were consistent with the opinions of clinical experts in 92% of cases.

[0171] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting anesthetic effects based on graph neural networks, characterized in that, Includes the following steps: The raw data sets of multimodal physiological monitoring and drug administration records were collected and preprocessed to obtain a standardized and synchronized time-series data set; A time sliding window is constructed based on a standardized and synchronized time series data set. The data segment within each time sliding window is used as the input segment for a single dynamic graph modeling, forming a time sliding window sequence. Based on the variable change trend, co-fluctuation relationship or statistical coupling characteristics within each time sliding window sequence, the initial topology of the dynamic physiological index-pharmacological variable graph structure is generated. On the initial topology of the dynamic physiological index-pharmacological variable graph structure, a set of learnable parameters for edge weights is assigned to each graph edge, and an initial feature representation set for each graph node is generated, while the parameter set of the graph neural network model is initialized. The set of learnable parameters for edge weights is adaptively updated on a continuous time sliding window using a standardized and synchronized time series data set. When a sudden change in the correlation is detected, a topological update is performed on the dynamic physiological index-pharma effect variable graph structure to obtain the updated dynamic physiological index-pharma effect variable graph structure. The updated dynamic physiological indicators-pharmacodynamic variables graph structure is input into the graph neural network model corresponding to the graph neural network model parameter set to obtain the node embedding representation set and the edge embedding representation set. The spatiotemporal fusion feature representation set is generated by combining the node embedding representation set and the edge embedding representation set with the time series features in the standardized and synchronized time series data set. The spatiotemporal fusion feature representation set is input into the multi-task anesthesia effect prediction engine and outputs the induction period anesthesia effect prediction result set. The induction period anesthesia effect prediction result set includes the expected time to reach the target anesthesia depth, the probability of excessive anesthesia risk, the hemodynamic fluctuation risk score and other anesthesia-related outcome indicators. Based on the predicted results of the induction period anesthesia effect and the edge weight distribution in the updated dynamic physiological indicators-pharmacological variables graph structure, an interpretable output set is generated. The interpretable output set includes key variable links, dominant influencing factors, and dosing adjustment suggestions.

2. The method for predicting anesthetic effects based on graph neural networks according to claim 1, characterized in that, The obtained standardized and synchronized time-series data set includes: The study collected raw data sets of multimodal physiological monitoring and drug administration records from patients before anesthesia induction and intubation. These raw data sets included bispectral index (BPI), EEG signal, systolic blood pressure, diastolic blood pressure, mean arterial pressure, heart rate, blood oxygen saturation, end-tidal carbon dioxide concentration, respiratory rate, and information on the type, dosage, and infusion rate of intravenously administered drugs. Simultaneously, the study collected patients' static baseline characteristics. The raw data sets of multimodal physiological monitoring and drug administration records were processed for time synchronization, noise removal, missing data completion, and unit standardization to obtain a standardized synchronized time-series data set. The standardized synchronized time-series data set maintained a one-to-one correspondence with the patients' static baseline characteristics.

3. The method for predicting anesthetic effects based on graph neural networks according to claim 1, characterized in that, The formation of the time-sliding window sequence includes: Set the time sliding window length parameter and the sliding step parameter; The standardized synchronized time series data set is segmented based on the time sliding window length parameter and the sliding step parameter to form multiple time sliding windows in sequence; Within each time sliding window, all multivariate feature vectors covered by the time sliding window are sequentially combined into a data segment; Within each time sliding window, record all time indices covered by the time sliding window; All data segments corresponding to the time sliding window are combined sequentially with the time index set to form a time sliding window sequence representation set.

4. The method for predicting anesthetic effects based on graph neural networks according to claim 1, characterized in that, The generation of the initial topology of the dynamic physiological index-pharmacological variable graph structure includes: Extract the data segment corresponding to each time sliding window sequence; Each physiological monitoring variable, drug variable, and patient static baseline characteristic in the data segment is constructed as a graph node in the graph structure, and all graph nodes constitute a graph node set. The cooperative change relationship characteristics of the corresponding variables of each pair of graph nodes in the graph node set are calculated within the current time sliding window. The cooperative change relationship characteristics are constructed by the change rate sequence of the variables and the variables of the previous moment within the time sliding window. The statistical coupling coefficient is used to represent the cooperative fluctuation relationship of the graph node pair within the k-th time sliding window. Based on the statistical coupling coefficient, it is determined whether to establish a graph edge connection in the graph structure for each graph node pair. When the absolute value of the statistical coupling coefficient is greater than the set edge connection threshold, a graph edge connection is established between the graph node pairs. All graph node pairs that meet the conditions constitute a graph edge set. By combining the set of graph nodes and the set of graph edges, the initial topology of the dynamic physiological indicators-pharmacological variables graph structure within the k-th time sliding window is generated.

5. The method for predicting anesthetic effects based on graph neural networks according to claim 1, characterized in that, The initialization graph neural network model parameter set includes: The initial topology of the dynamic physiological index-pharmacodynamic variable graph structure constructed within the k-th time sliding window is processed, and learnable edge weight parameters are introduced on each graph edge. The initial values ​​of the learnable edge weight parameters are jointly determined by the physiological coupling prior during the induction period and the local response regulatory factors. Calculate the standard deviation and mean of the variable within the current time sliding window, and obtain the local response regulation factor of the variable pair by weighting the response heterogeneity index before and after the drug injection time point. Learnable edge weight parameters are used as the initial weights of graph edges in the graph structure, and form a set of learnable edge weight parameters. Based on the time-series performance of variables of each graph node in the graph node set within the current time sliding window, an initial feature representation vector of the node is constructed, and all initial feature representation vectors of the nodes form the initial feature representation set of the nodes. Initialize the graph neural network model parameter set by inputting the set of learnable edge weight parameters, the set of initial node feature representations, and the graph neural network model parameter set into the graph neural network model.

6. The method for predicting anesthetic effects based on graph neural networks according to claim 1, characterized in that, The updated dynamic physiological index-pharmacological variable graph structure includes: Within each time sliding window, the standardized synchronized time series data set corresponding to the current time sliding window is used, combined with the learnable parameter set of edge weights and the initial feature representation set of nodes of the graph neural network model, to perform graph information propagation operation and obtain the edge representation output of each graph edge; For each edge in the graph, the edge weight is updated by calculating the output result. The updated edge weights are then collected and updated to form an edge weight update set. Determine whether there is a structural abrupt change based on the magnitude of the change in each edge weight in the edge weight update set; Perform a topology update operation on graph edges that meet the structural mutation criteria: When the weight of a graph edge increases after an update and there is no such edge connecting the current graph structure, a new graph edge is added between the corresponding graph node pairs. If the weight of a graph edge decreases after an update and there are connecting edges in the current graph structure, remove that graph edge from the graph structure. Apply all topology update operations to the initial topology of the original graph structure to generate a dynamic physiological indicator-drug efficacy variable graph structure with updated topology. The updated graph structure retains the same set of graph nodes, while the graph edge set is filtered based on whether the updated edge weights exceed the edge retention threshold, ultimately resulting in a new graph edge set.

7. The method for predicting anesthetic effects based on graph neural networks according to claim 1, characterized in that, The set of prediction results for the induction period anesthesia effect includes: Within each time sliding window, the graph neural network model is used to calculate the node embedding representation set and the edge embedding representation set of the dynamic physiological index-pharmacodynamic variable graph structure after the topology is updated; Extract the time series features of the standardized and synchronized time series data set within the current time sliding window; The node embedding representation vector of each graph node is concatenated with its time series feature vector to obtain the spatiotemporal fusion feature representation of a single node. The spatiotemporal fusion feature representations of all graph nodes are combined sequentially into a spatiotemporal fusion feature representation set. At the same time, the graph edge embedding representation set and the spatiotemporal fusion feature set of graph nodes are used as inputs to the multi-task anesthesia effect prediction engine. In the multi-task anesthesia effect prediction engine, a multi-branch prediction task is performed on the input graph node spatiotemporal fusion feature set and graph edge embedding representation set. The output set of induction period anesthesia effect prediction results includes the estimated time to reach the target depth of anesthesia, the probability of excessive anesthesia risk, the hemodynamic fluctuation risk score, and other anesthesia-related outcome indicators.

8. The method for predicting anesthetic effects based on graph neural networks according to claim 1, characterized in that, The output of the multi-task anesthesia effect prediction engine includes: For each time sliding window, the spatiotemporal fusion feature representation set and the graph edge embedding representation set are processed by the shared feature extraction module to obtain the global fusion expression vector; The global fusion expression vector is input into the multi-task prediction branch, each multi-task prediction branch corresponds to a sub-task of anesthetic effect, and each branch outputs the corresponding prediction result. For each subtask, the corresponding prediction result and its corresponding true label are output, and the single-task loss is calculated. The weighted combination is then used to form a joint loss function. The gradient descent optimization algorithm is used to iteratively update the parameters of the multi-task anesthesia effect prediction engine, so that the value of the joint loss function continuously decreases in all time sliding windows, and finally obtains a set of real-time updated prediction results of induction period anesthesia effect.

9. The method for predicting anesthetic effects based on graph neural networks according to claim 1, characterized in that, The generation of the interpretable output result set includes: Based on the set of predicted results of anesthetic effect during induction and the edge weight distribution in the updated dynamic physiological index-pharmacological variable graph structure, edges with edge weight values ​​higher than the set edge weight threshold are extracted to determine the key variable link set. The key variable link set is the set of directed paths of physiological monitoring variables and drug variables within the current time sliding window, where the edge weight value is greater than or equal to the edge weight threshold and is significantly associated with the trend of change of the predicted results of the target anesthetic effect. Based on the rate of change, predicted output variation, and edge weight values ​​of each variable in the key variable link set within the spatiotemporal fusion feature representation set, dominant influencing factors are defined. Dominant influencing factors are variables that satisfy the following conditions: The dominant influencing factors are the set of variables whose rate of change in the key variable link is higher than the rate of change threshold and is within the range of change in the predicted anesthetic effect, and whose edge weight values ​​are in the top 30% of all links. Based on the key variable chain set, the dominant influencing factors, and the prediction results of the induction period anesthesia effect, combined with the side weight trend of drug variables and the status of the dominant influencing factors, dosing adjustment suggestions are generated.

10. The method for predicting anesthetic effects based on graph neural networks according to claim 1, characterized in that, The recommended dosing adjustments are recommendations that meet the following criteria: When the increase in the margin weight between the drug variable and the dominant influencing factor is greater than the increase threshold in a continuous window, and an increase in the risk of anesthesia depth is predicted at the same time, a recommendation to reduce the drug dosage is given. When the decrease in the marginal weight between the drug variable and the dominant influencing factor is greater than the decrease threshold, and the depth of anesthesia is lower than the target threshold, a suggestion to appropriately increase the drug dosage is given. When the side weights of physiological monitoring variables change abruptly and are associated with the risk of adverse events, a dynamic warning is given.

Images