Real-time early warning system and method for dynamic adverse events of intraoperative patient

Abstract

The invention discloses an intraoperative patient dynamic adverse event real-time early warning system and method, and relates to the technical field of intelligent diagnosis and treatment data processing, and the real-time early warning system comprises an intraoperative data unified preprocessing module which is used for preprocessing clinical variables and outputting structured triples; the variable structure modeling module is used for carrying out feature extraction, and the variable graph feature sequence output time sequence modeling module is used for modeling dynamic information of historical and future contexts at the same time and outputting a context feature sequence H = {ht}; the feature fusion and risk prediction module is used for fusing gt and ht. The risk score sequence prediction interpretability module for outputting continuous time steps is used for carrying out feature attribution analysis on prediction output. The dynamic adverse event real-time early warning system and method are innovative in structure, accurate in prediction, high in interpretability and suitable for deployment, can adapt to complex data characteristics in an operation, and have high real-time performance, high robustness and clinical interpretability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent diagnosis and treatment data processing, and particularly relates to an intraoperative patient dynamic adverse event real-time early warning system and method. BACKGROUND

[0002] During intensive care (ICU) and surgical anesthesia, the patient's vital signs change with high dynamics and suddenness, such as sudden drop in blood pressure, arrhythmia, and sharp decrease in oxygenation index, and the clinician needs to make decisions and take intervention measures in a very short time. Traditional risk assessment methods, such as SAPS score and APACHE score, mainly rely on preoperative or early admission static indicators, and cannot reflect the real-time change trend of the patient's physiological state during surgery or critical period.

[0003] In actual clinical work, monitoring equipment can continuously record the patient's multi-source physiological data, such as invasive / non-invasive blood pressure, heart rate, blood oxygen saturation, respiratory rate, electroencephalogram monitoring, anesthetic gas concentration, drug administration time and dosage, etc. However, these data usually have inconsistent sampling frequencies, a large number of missing values, and come from different devices and systems, forming a complex irregular longitudinal data stream, making it difficult for doctors to make continuous and accurate risk assessment by eye or experience.

[0004] For example, in intraoperative hypotension events, if timely warning and intervention cannot be made before blood pressure drops significantly, it may lead to low perfusion of important organs, causing postoperative acute kidney injury or neurological complications. Similarly, for high-risk patients, the physiological response after the use of anesthetic drugs is extremely individualized, and cannot be effectively warned by fixed thresholds or single-point indicators.

[0005] In recent years, medical artificial intelligence has gradually emerged in the application of intraoperative risk management, and some studies have attempted to use deep learning models to process intraoperative continuous monitoring data to achieve early identification of hypotension, arrhythmia and other events. Although existing research has made preliminary achievements in ICU and intraoperative risk prediction, there are still significant deficiencies in dealing with the complex structure of real intraoperative monitoring data and meeting the high timeliness requirements of clinical decision-making. This is manifested in the following aspects:

[0006] 1. Inadaptation to the irregular structure of intraoperative monitoring data

[0007] Most current models assume that the input data is regular and complete time series, which is seriously inconsistent with the actual situation during surgery. During surgery, various monitoring data (such as invasive blood pressure, anesthetic gas concentration, drug dosage, etc.) have different recording frequencies, uneven time points, and a large number of missing values and asynchrony. Existing methods mostly use simple linear interpolation or forced alignment, ignoring the important information carried by the time interval, which can easily lead to misjudgment of the clinical risk trend.

[0008] 2. Lack of mechanism to model "time sensitivity"

[0009] Intraoperative events have high time dependence, such as changes in reaction within 5 minutes after drug administration have critical clinical significance. However, traditional models often do not introduce explicit time-aware mechanisms, cannot identify interval information of variable updates, and cannot simulate key dynamics such as time delay and short-term trends, resulting in a lack of response capability of model prediction to intraoperative sudden changes.

[0010] 3. Ignoring the synergy and intervention effect between monitoring variables

[0011] There are physiological correlations between clinical variables, such as blood pressure affected by drugs and respiratory rate related to sedation level. Existing models generally treat various monitoring indicators as independent inputs, do not build interaction networks between variables, lack the ability to express "drug action chain" or "pathological pathway synergy change", and are difficult to simulate real pathological processes.

[0012] 4. Only one-way understanding of time series, unable to present state evolution trend

[0013] Most models use a one-way time series structure from "past to present", and cannot consider the influence of "future state trend on current assessment". For example, the fluctuation trend before intraoperative hypotension or the early postoperative recovery trajectory are clinically significant for determining whether the current intervention is effective. The lack of context modeling capability limits the model's understanding of state evolution patterns.

[0014] 5. Unable to highlight the importance of key time periods

[0015] Not all time points during surgery are equally important for risk prediction. Current methods generally use equal weight to process the entire time series data, lack of mechanism to highlight key time slices (such as before drug onset, 5 minutes before disease deterioration), making the model insensitive to sudden states, limiting its early warning value.

[0016] 6. Lack of model transparency, making it difficult for doctors to trust its use

[0017] Most existing deep learning models are "black box structures" and do not integrate effective explainability techniques such as variable attribution and visualization analysis. Doctors cannot determine which variables, trends, or clinical facts the model's risk score is based on in actual work, making it difficult to establish clinical trust, resulting in the model being difficult to truly enter the intraoperative decision-making process.

[0018] 7. Complex reasoning and slow response, difficult to apply to intraoperative real-time scenarios

[0019] Many existing methods rely on complex structure and large parameter models (such as Transformer structure or large-scale graph network), although the theoretical performance is good, but in the intraoperative environment, the conventional server is difficult to support its second-level reasoning. Model loading is slow, response delay is high, and it is difficult to meet the strict time limit requirement of "fast evaluation and timely intervention" in the surgical process.

[0020] In summary, although some researches in recent years have tried to apply data-driven methods to improve risk identification ability, and have made preliminary progress in intraoperative risk prediction, in terms of data processing accuracy, model interpretability, real-time deployment and system realizability, it still cannot meet the clinical needs in the current operating room environment, cannot realize real-time dynamic evaluation at the minute level, and lacks sufficient transparency and stability, and lacks a complete system-level technical solution to support clinical translation application. SUMMARY

[0021] Based on the above problems, the purpose of the present application is to solve the problems of poor adaptability, weak real-time performance, insufficient modeling ability and low interpretability of the existing adverse event risk prediction methods in the actual clinical environment, and to provide an intraoperative patient dynamic adverse event real-time warning system and method, which can adapt to the complex data characteristics in the operation, and has high real-time performance, strong robustness and clinical interpretability.

[0022] The technical scheme adopted by the present application to achieve its purpose is an intraoperative patient dynamic adverse event real-time warning system, comprising an intraoperative data unified preprocessing module, a variable structure modeling module, a time series modeling module, a feature fusion and risk prediction module, and a prediction interpretability module, wherein:

[0023] The intraoperative data unified preprocessing module is used to receive the monitoring record sequence of a plurality of variables from a plurality of clinical devices, denoted as variable feature X, uniformly resample all variables to a fixed time step, and construct a mask matrix M(t, d) for marking whether there is an observation value of the variable at time t, and finally output the structured triple {X, M, Γ}, wherein the time embedding Γ is the time-aware embedding result integrated by the time decay embedding vector of each variable at each time step;

[0024] The variable structure modeling module is used to receive the variable feature X and the time embedding Γ, and splice and fuse each dimension feature at each time step to form an initial variable representation matrix H (0) , use graph structure modeling, and use graph convolutional neural network GCN for feature extraction to obtain the node feature representation matrix H (1) outputted by the graph convolutional neural network GCN after the first layer convolution calculation, and finally output the variable graph feature sequence

[0025] The time series modeling module is configured to receive the variable feature sequence A bidirectional long short-term memory network (BiLSTM) is adopted to model dynamic information of historical and future contexts, and an output of each time step is denoted as wherein and denote forward and backward LSTM outputs respectively, and a final output context feature sequence H = {h t};

[0026] The feature fusion and risk prediction module is configured to fuse variable structure features g t output by the variable structure modeling module and context features h t output by the time series modeling module, adopt a gating mechanism, and output a risk score per minute through a sigmoid function to output a risk score sequence of continuous time steps

[0027] The prediction explainability module is configured to perform feature attribution analysis on a prediction output by using an integral gradient method (IG), compare a change amplitude of a model output in a process in which a variable is smoothly transitioned from a reference input x' to an actual input x, and determine a contribution degree of each feature to .

[0028] Further, the variable features wherein t = 1,..., T denotes a time step, and d = 1,..., D denotes a variable dimension, and the variable at least includes blood pressure, heart rate, respiratory rate, blood oxygen saturation, electroencephalogram monitoring, anesthetic gas concentration, medication time, and dosage;

[0029] The time embedding Γ is in the following form:

[0030]

[0031] In the formula, is an exponential decay embedding, which is in the following form:

[0032]

[0033] In the formula, β is a learnable parameter, used to adjust an influence strength of a time interval on a model representation, is a time interval vector, reflecting an interval distance between a current time point and a latest observation of the variable, and is defined as follows:

[0034]

[0035] In the formula, τ is a historical time point index less than a current time t, used to find a time point at which the variable was last observed.

[0036] Further, the variable structure modeling module uses graph structure modeling, specifically:

[0037] Each variable is abstracted as a node in the graph, and the adjacency matrix Encode the prior relationship between variables or the correlation learned dynamically, and use a learnable adjacency matrix to represent:

[0038]

[0039] Degree matrix:

[0040] Where A is the prior relationship matrix between variables, which can be obtained by physiological knowledge or data-driven learning, I is the unit matrix representing the self-loop connection, and the degree matrix After normalization, it is used for graph convolution operation;

[0041] Graph convolutional neural network GCN update formula:

[0042]

[0043] In the formula: H (l) is the node feature matrix of the lth layer; W (l) is the weight matrix of the lth layer; and σ is the activation function.

[0044] The variable structure modeling module uses graph convolutional neural network GCN for feature extraction, and the specific calculation method is:

[0045]

[0046] In the formula: indicates the adjacency matrix plus the self-loop, is its degree matrix, W (0) is a learnable transformation weight, and σ(·) represents the activation function.

[0047] Further, the time series modeling module is also used to establish global dependencies between all time steps using a multi-head self-attention mechanism, specifically, the LSTM output is used as the input of the query, key and value, and the calculation is as follows:

[0048]

[0049] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0050] In the formula: Q=h t W Q , K=h t W​K V = h t W V , W O is a linear transformation matrix for multi-head attention output, used for dimension transformation on the spliced result;

[0051] W Q is a query (Query) mapping matrix, used to map the input hidden state h t to the query vector Q; W K is a key (Key) mapping matrix, used to map the input hidden state h t to the key vector K; W V is a value (Value) mapping matrix, used to map the input hidden state h t to the value vector V. They are all learnable parameter matrices, which will be automatically updated during the training process.

[0052] Further, the variable structure feature g t is fused with the context feature h t , specifically, a fusion weight z t is generated by a sigmoid function, and the fusion method is as follows,

[0053] z t = σ (W z [g t ; h t ]+b z )

[0054]

[0055] In the formula: is the fused feature representation, W z is a learnable weight, b z is a bias term, and represents element-wise multiplication.

[0056] The sigmoid function is used to output the risk score per minute Specifically, after the fused feature representation is converted by the fully connected layer, the sigmoid function is used to output,

[0057]

[0058] In the formula, the output layer parameter W0 is a weight mapping matrix from features to a prediction space, and b0 is the corresponding bias term.

[0059] Further, the prediction explainability module is used to perform feature attribution analysis on the prediction output by using the integral gradient method IG, specifically:

[0060] The contribution value of the i-th feature is calculated under the model prediction function F(·) based on the model input x and the reference input x', and the integral expression is as follows:

[0061]

[0062] In the formula, x represents the actual input feature vector of the model at a certain time step, and contains the numerical value of all monitoring variables at the time point; x' represents the reference input, which is the reference feature vector for contribution calculation, IG i (x) represents the contribution value of the i-th input feature when generating the current prediction result , reflecting the positive or negative influence and its size of the feature on the prediction result; α is the interpolation coefficient of the integral path, and the value range is [0, 1], which is used to represent the linear interpolation progress from the reference input x' to the actual input x; dα is the integral infinitesimal, which represents the calculation process of gradually accumulating the gradient in the process of α changing from 0 to 1.

[0063] The application also provides an intraoperative patient dynamic adverse event real-time warning method, comprising the following steps:

[0064] S1, receiving a monitoring record sequence of a plurality of variables from a clinical multi-device, denoted as variable feature X, uniformly resampling all variables to a fixed time step, and constructing a mask matrix M(t, d) for marking whether there is an observation value of the variable at time t, finally outputting the structured triple {X, M, Γ}, wherein the time embedding Γ is the time-aware embedding result integrated by the time decay embedding vectors of all variables at each time step;

[0065] S2, according to the variable feature X and the time embedding Γ, splicing and fusing each dimension feature at each time step to form an initial variable representation matrix H (0) , using graph structure modeling, and using graph convolutional neural network GCN for feature extraction to obtain the node feature representation matrix H (1) obtained after the first layer convolution calculation of the graph convolutional neural network GCN, finally outputting the variable graph feature sequence

[0066] S3, according to the variable graph feature sequence using bidirectional long short-term memory network BiLSTM to model the dynamic information of the history and future context at the same time, and the output of each time step is denoted as wherein and respectively represent the forward and backward LSTM outputs, and finally output the context feature sequence H={h t};

[0067] S4, the variable structure feature gt with contextual features h t fusion, using a gating mechanism, outputting a risk score per minute through a sigmoid function final output a sequence of risk scores at continuous time steps

[0068] S5, using integral gradient method IG for feature attribution analysis on the predicted output, by comparing the change in model output during the process of variable smoothing transition from reference input x' to actual input x, determine the contribution of each feature to .

[0069] Further, the variable features where t=1,…,T represents the time step, d=1,…,D represents the variable dimension, and the variables include at least blood pressure, heart rate, respiratory rate, blood oxygen saturation 2, electroencephalogram monitoring, anesthetic gas concentration, medication time and dosage;

[0070] The time embedding Γ is as follows:

[0071]

[0072] In the formula, is an exponential decay embedding, which is as follows:

[0073]

[0074] In the formula, β is a learnable parameter, used to adjust the influence strength of the time interval on the model representation, is a time interval vector, reflecting the interval distance between the current time point and the last observation of the variable, which is defined as follows:

[0075]

[0076] In the formula, τ is the index of the historical time point less than the current time t, used to find the time when the variable was last observed.

[0077] The graph structure modeling used in step S2 is as follows:

[0078] Each variable is abstracted as a node in the graph, and the prior relationship between variables or the correlation dynamically learned is encoded through an adjacency matrix learnable adjacency matrix:

[0079]

[0080] Degree matrix:

[0081] Where A is the prior relationship matrix between variables, which can be obtained by physiological knowledge or data-driven learning, I is the identity matrix representing the self-loop connection, and the combination degree matrix After normalization, it is used for graph convolution operation;

[0082] The graph convolution neural network GCN update formula is:

[0083]

[0084] In the formula: H (l) is the lth layer node feature matrix; W (l) is the lth layer weight matrix; and sigma is an activation function;

[0085] The variable structure modeling module adopts a graph convolution neural network GCN for feature extraction, and the specific calculation method is:

[0086]

[0087] In the formula: denotes an adjacency matrix plus a self-loop, is its degree matrix, W (0) is a learnable transformation weight, and sigma(·) denotes an activation function

[0088] Further, in the step S3, a multi-head self-attention mechanism is also used to establish global dependencies between all time steps, specifically, the LSTM output is taken as the input of query, key and value, and the calculation is as follows:

[0089]

[0090] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0091] In the formula: Q=h t W Q , K=h t W K , V=h t W V , W O is a linear transformation matrix of multi-head attention output, used for dimension transformation on the spliced result;

[0092] W Q is a query (Query) mapping matrix, used for mapping the input hidden state h t to the query vector Q; W K is a key (Key) mapping matrix, used for mapping the input hidden state h t to the key vector K; and W VValue mapping matrix, used to map the input hidden state h t to the value vector V. They are all learnable parameter matrices, which are automatically updated during the training process.

[0093] Further, in step S4, the variable structure feature g t is fused with the context feature h t , specifically, the fusion proportion z t is generated by the sigmoid function, and the fusion method is as follows,

[0094] z t = σ(W z [g t ; h t ]+b z )

[0095]

[0096] In the formula: is the fused feature representation, W z is a learnable weight, b z is a bias term, and represents element-wise multiplication.

[0097] The sigmoid function is used to output the risk score per minute Specifically, after the fused feature representation is converted by the fully connected layer, the sigmoid function is used to output,

[0098]

[0099] In the formula, the output layer parameter W0 is the weight mapping matrix of the feature to the prediction space, and b0 is the corresponding bias term.

[0100] In step S5, the integral gradient method IG is used to perform feature attribution analysis on the prediction output, specifically, based on the model input x and the reference input x', the attribution value of the i-th feature is calculated under the model prediction function F(·), and the integral expression is as follows:

[0101]

[0102] In the formula, x represents the actual input feature vector of the model at a certain time step, containing the numerical values of all monitoring variables at that time point; x' represents the reference input, which is the reference feature vector for attribution calculation; IG i (x) represents the i-th input feature in generating the current prediction result The attribution value of the time reflects the positive or negative influence of the feature on the prediction result and its size; alpha is the interpolation coefficient of the integral path, the value range is [0, 1], and is used to represent the linear interpolation progress from the reference input x' to the actual input x; dalpha is the integral infinitesimal, which represents the calculation process of gradually accumulating the gradient in the process of alpha changing from 0 to 1.

[0103] The beneficial effects of the present application are:

[0104] The present application aims at the many challenges faced in intraoperative patient adverse event prediction, including irregularity of monitoring data, strong time dependence, asynchronous variable sampling, uninterpretable prediction, and weak generalization ability of the model in multiple scenarios, and constructs an intelligent dynamic risk early warning method with structural innovation, accurate prediction, strong interpretability and suitability for deployment. The method shows excellent performance in multiple dimensions, specifically in the following aspects:

[0105] (1) The present application realizes the dynamic and continuous early warning of intraoperative adverse events at the minute level, significantly improving the real-time response ability to sudden conditions. Unlike traditional static risk scoring methods that are based on only preoperative or postoperative 24-hour single-point data, the present model supports real-time rolling input of multi-source data with a very short preset time step, for example, 1 minute, and continuously outputs updated risk scores, which can timely identify high-risk states such as hypotension, respiratory depression and drug overreaction. The model was verified in the INSPIRE database based on large-scale real clinical intraoperative data, and achieved excellent performance in dynamic prediction tasks, with an area under the receiver operating characteristic curve (AUROC) of more than 0.91 and an area under the precision-recall curve (AUPRC) of more than 0.41, significantly better than traditional LSTM and static scoring systems, with outstanding real-time and forward-looking intervention value.

[0106] (2) The time-aware preprocessing mechanism proposed in the present application can effectively integrate intraoperative multi-source, asynchronous and irregularly sampled data, introduce a mask matrix and a time interval vector on the basis of preserving the time sequence structure, avoid the information distortion caused by simple filling or forced interpolation of missing values in traditional methods, and improve the adaptability and overall robustness of the model to complex data structures in real intraoperative conditions.

[0107] (3) The model combines graph neural networks with bidirectional time series modeling structure (BiLSTM) and Transformer attention mechanism, and for the first time expresses the synergistic relationship between variables and the context dependence in the time dimension in the intraoperative risk modeling scenario. Through modeling of variable dependence structure, the model can better identify the drug action path, the linkage trend between physiological variables, and focus on the key time window combined with the attention mechanism, thereby significantly improving the accuracy and sensitivity of risk state identification. Experimental comparison shows that in the intraoperative 12 to 48 hour death prediction task, the AUROC index of the model is 2.1 to 4.7 percentage points higher than that of the traditional LSTM on average.

[0108] (4) To adapt to the dynamic changes of different patient characteristics and surgical stages, the application introduces a gating fusion mechanism in the model structure, which can automatically adjust the fusion weight between graph structure features and time sequence features, thereby maintaining prediction stability in complex clinical scenarios such as different variable update frequencies and slight data distribution drift. In cross-subset evaluation, the model's AUROC remains above 0.81 when migrating between different surgical types in the INSPIRE database, fully verifying its strong generalization ability among different clinical subgroups.

[0109] (5) The model of the application embeds the Integrated Gradients explanation algorithm, providing clear variable attribution output for each risk prediction result, supporting doctors in tracing the source of risk score and identifying key influencing factors such as lactate elevation, systolic blood pressure drop, and sedative drug dosage. Actual application examples show that the model prediction results have good temporal consistency and physiological rationality with physiological indicator changes, improving the doctor's trust and acceptance of the model, and having good clinical usability.

[0110] (6) The model structure design of the application is lightweight, with a total parameter amount controlled at about 720,000, memory occupation less than 3MB, and inference speed up to 35 times per second on ordinary CPU devices, suitable for deployment in operating room workstations, edge computing terminals, or hospital servers in resource-constrained environments, effectively solving the practical contradiction of "strong model performance but difficult deployment", and having excellent engineering landing performance.

[0111] (7) The model architecture of the application has high compatibility and expandability, and can perform single-task adverse event prediction, as well as support simultaneous death prediction, postoperative complication judgment, hospitalization time estimation, and other multi-task joint modeling, providing a good technical foundation and platform support for building an integrated intelligent risk assessment system covering the entire perioperative period. BRIEF DESCRIPTION OF DRAWINGS

[0112] Figure 1 Figure 1 is a structure diagram of a real-time early warning system according to Embodiment 1 of the application.

[0113] Figure 2 Figure 2 is a flowchart illustrating a real-time early warning method according to an embodiment of the present application. DETAILED DESCRIPTION

[0114] In order to more clearly understand the above-mentioned purposes, features and advantages of the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0115] Embodiment 1

[0116] Figure 1 An embodiment of the intraoperative real-time early warning system for dynamic adverse events of a patient in the present application is shown, which includes an intraoperative data unified preprocessing module, a variable structure modeling module, a time sequence modeling module, a feature fusion and risk prediction module, and a prediction interpretability module, wherein:

[0117] The intraoperative data unified preprocessing module is used for receiving data streams of a plurality of variables from a plurality of devices in a clinic, and inputting a monitoring record sequence including a plurality of variables, denoted as variable features Wherein t = 1, …, T represents a time step, d = 1, …, D represents a variable dimension, and the variable at least includes blood pressure, heart rate, respiratory rate, blood oxygen saturation SpO2, brain electrical monitoring BIS, anesthetic gas concentration, medication time and dosage; since the sampling frequencies of various variables are different, the time records are not uniform and there are missing values, this module first uniformly resamples all variables to a fixed 1-minute time step to form an equal-interval time axis, and on this basis, a mask matrix M(t, d) is constructed to mark whether there is an observation value of the variable at time t, if there is, the corresponding element is 1, otherwise 0.

[0118] For the observation time point interval of each variable, a time interval vector is introduced Its definition is as follows:

[0119]

[0120] In the formula, τ is a historical time point index less than the current time t, which is used to find the time when the variable was last observed.

[0121] This vector reflects the interval distance between the current time point and the last observation of the variable. In order to pass the time decay signal to the model, this module further calculates the exponential decay embedding Its form is as follows:

[0122]

[0123] Where β is a learnable parameter for adjusting the influence strength of the time interval on the model representation.

[0124] Finally, the module outputs the structured triplets {X, M, Γ}, where Γ is the time-aware embedding result integrated by the time decay embedding vectors of all variables at each time step:

[0125]

[0126] The embedding vector is concatenated with the original variable feature and input into the subsequent variable structure modeling module and time series modeling module.

[0127] The variable structure modeling module is configured to receive the preprocessed variable feature X and the time embedding Γ as inputs, concatenate and fuse the features of each dimension at each time step to form an initial variable representation matrix H of dimension Dxd. (0) Considering the complex interaction between different variables during the operation, such as the dependence between medication and blood pressure, respiratory rate and sedation state, the module uses graph structure modeling to abstract each variable as a node in the graph, and encodes the prior relationship between variables or the correlation learned dynamically through an adjacency matrix The adjacency matrix A can be a fixed graph calculated based on Pearson correlation coefficient, mutual information, etc., or a dynamic graph learned end-to-end by the model. In this embodiment, a learnable adjacency matrix is used to represent:

[0128]

[0129] where A is the prior relationship matrix between variables, which can be obtained through physiological knowledge or data-driven learning, I is the identity matrix representing the self-loop connection, and the degree matrix D is used for graph convolution operation after normalization;

[0130] The graph convolution neural network GCN update formula is:

[0131]

[0132] In the formula, H (l) is the node feature matrix of the lth layer, W (l) is the weight matrix of the lth layer, and σ is the ReLU activation function.

[0133] To enhance the variable structure modeling capability, this embodiment uses a graph convolutional neural network GCN to extract features, and the calculation method is:

[0134]

[0135] In the formula: represents the adjacency matrix with a self-loop, is its degree matrix, and W (0) ​where σ(·) denotes an activation function, H (1) denotes the node feature representation matrix obtained by the first layer of graph convolution, where each row corresponds to the feature vector of a variable node after fusing the information of its neighbor nodes. Compared with the initial input feature matrix H (0) , it has contained high-order correlation information after one graph structure propagation and feature transformation. This result will be used as the input of subsequent graph convolution layers or other modules to further extract the high-order dependency between variables.

[0136] Through multi-layer graph convolution stacking operations, the final variable graph feature sequence is output as the input of the subsequent time series modeling module.

[0137] The time series modeling module is configured to receive the variable graph feature sequence as input to capture the evolution trend of the patient state in the time dimension. A bidirectional long short-term memory network (BiLSTM) is used to model the dynamic information of the historical and future context simultaneously, and the output of each time step is denoted as where and denote the forward and backward LSTM outputs, respectively.

[0138] To further enhance the model's attention ability to key time segments (such as the 5 minutes before the blood pressure drops, the reaction window after medication, etc.), the time series modeling module is also configured to use a multi-head self-attention mechanism to establish global dependencies between all time steps. Specifically, the LSTM output is used as the input of the query, key, and value, and the calculation is as follows:

[0139]

[0140] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0141] where: Q=h t W Q , K=h t W K , V=h t W V , and W O is a linear transformation matrix for the multi-head attention output, used for dimension transformation on the concatenated result.

[0142] W Q is a query (Query) mapping matrix used to map the input hidden state h t to the query vector Q; and W Kis the Key mapping matrix, which is used to map the input hidden state h t to the key vector K; W V is the Value mapping matrix, which is used to map the input hidden state h t to the value vector V. They are all learnable parameter matrices, which are automatically updated during the training process.

[0143] By focusing on abnormal fluctuations at different time scales through multiple attention heads in parallel, the model can learn complex preoperative trends, intraoperative abnormalities, and postoperative recovery trajectories, etc.

[0144] The final output of this module is the context feature sequence H = {h t}, which is used to fuse with the variable structure features and generate risk scores.

[0145] The feature fusion and risk prediction module is used to fuse the variable structure features g t output by the variable structure modeling module with the context features h t output by the time series modeling module, taking into account the differences in expression content and modeling angle. To achieve dynamic weight control, a gating mechanism is used to generate a fusion ratio z t through a sigmoid function, and the fusion method is as follows:

[0146] z t = σ(W z [g t ; h t ]+b z )

[0147]

[0148] In the formula: is the fused feature representation, W z is a learnable weight, b z is a bias term, and ⊙ represents element-wise multiplication.

[0149] The fused feature representation is converted through a fully connected layer, and a sigmoid function is used to output the risk score per minute

[0150]

[0151] In the formula, the output layer parameter W0 is the weight mapping matrix from the feature to the prediction space, and b0 is the corresponding bias term.

[0152] The final output of this module is the risk score sequence at consecutive time steps , which is used for clinical early warning.

[0153] In the model training phase, the trained model is a time-aware bidirectional attention long short-term memory network, which is improved on the basis of the baseline LSTM model.

[0154] Training timing and method:

[0155] After dividing the ICU cases (MIMIC-IV and eICU-CRD databases) into training set, validation set and test set (ratio 7:2:1) by patient, the model is trained on the training set, the validation set is used to select the best model, and the test set is used for performance evaluation and generalization verification.

[0156] For static prediction tasks (such as death prediction 12 hours after ICU admission), fixed time point triggered prediction and parameter learning in the training phase;

[0157] For dynamic prediction tasks (such as hourly prediction of 24-hour mortality risk), continuous prediction is performed at each time step during the entire ICU hospitalization, and the model is optimized in the training phase to adapt to the continuously updated input.

[0158] To deal with the problem of uneven ratio of positive and negative samples, a weighted cross-entropy loss function is introduced for optimization:

[0159]

[0160] In the formula, α is the loss weight of positive samples, which is automatically calculated from the training set; represents the prediction value of the i-th sample at the current time step, that is, the risk probability (range 0-1) output by the model; since represents the prediction value at time step t, which is used to emphasize the time index of the time series, and Both have the same essential meaning, the difference lies in the use of subscript symbols: i is more biased towards sample number, t is more biased towards time index, and they can be interchanged in single patient time series prediction.

[0161] L represents the value of the loss function (Loss function), and the weighted cross-entropy loss is used to measure the gap between the predicted value of the model and the true label, and the positive class weight α is introduced to alleviate the problem of sample class imbalance.

[0162] During the training process, the optimizer will perform backpropagation according to the size of L, update the model parameters, and make the prediction result closer to the true label y, and finally improve the prediction accuracy of the model. The specific role is as the objective function (Objective function) of model parameter optimization, and the training process is to constantly minimize L.

[0163] Although the single time step prediction value has been given in the foregoing inference process The calculation method, but the purpose of the model training phase is not only to generate a single time. The key lies in continuously iteratively optimizing model parameters based on a large number of training samples. Specifically, during training, the system calculates predicted values ​​at multiple time steps. Corresponding real label y t The model employs a weighted cross-entropy loss L, which is then converted into a gradient signal via backpropagation. This gradient signal is used to update the learnable parameters in various sub-modules, including the graph neural network, bidirectional temporal modeling structure, and gating fusion mechanism. This process gradually improves the model's prediction accuracy and generalization ability for different patients and risk states at different time points. The trained and optimized model can stably output high-confidence, continuous-time-step risk score sequences during the inference phase. Used for real-time clinical early warning and decision support.

[0164] The predictive interpretability module is used to perform feature attribution analysis on the predicted output using the integral gradient method, thereby enhancing the transparency of the model output and the doctor's understanding of the risk score results. By comparing the magnitude of change in the model output during the smooth transition of variables from the reference input x′ to the actual input x, the module determines the relationship between each feature and the actual input x. The extent of their contribution is as follows:

[0165] Based on the model input x and the reference input x′, the attribution value of the i-th feature is calculated under the model prediction function F(·), and the integral expression is as follows:

[0166]

[0167] In the formula, x represents the actual input feature vector of the model at a certain time step, containing the values ​​of all monitored variables (such as blood pressure, heart rate, blood oxygen saturation SpO2, medication dosage, etc.) at that time point; x′ represents the reference input, serving as the baseline feature vector for attribution calculation. In this embodiment, it can be a zero tensor or the mean vector of each feature, representing the input state of "no effective information". Both of these, along with the aforementioned prediction results, are... The relationship is: The model prediction function F(·) is calculated based on the input x. The integral gradient method determines the feature pairs by comparing the magnitude of the change in the model output during the smooth transition from the reference input x′ to the actual input x. The extent of their contribution.

[0168] IG i (x) represents the i-th input feature in generating the current prediction result. The attribution value reflects the positive or negative influence and its size of the feature on the prediction result. When the attribution value is positive, it means that the feature will increase the prediction risk score. When the attribution value is negative, it means that the feature will decrease the risk score. The greater the absolute value is, the more significant the influence is. Alpha is an interpolation coefficient of the integral path, and the value range is [0, 1], which is used to represent the linear interpolation progress from the reference input x' to the actual input x; dalpha is an integral infinitesimal, which represents the calculation process of gradually accumulating the gradient in the process of alpha changing from 0 to 1.

[0169] In actual calculation, the integral operation is realized by numerical approximation. After the attribution result is visualized, it is displayed in the form of a heat map beside the per-minute risk score result, which helps doctors track the key variables (such as lactate rise, SBP drop, recent drug dose increase, etc.) that affect the prediction, thereby improving the clinical interpretability and credibility of the model.

[0170] The embodiment adopts a lightweight deep neural network structure, with a total parameter quantity of about 720,000 and a single inference time of not more than 0.03 seconds, which can run on a normal hospital server or edge device without GPU. The input data adopts a sliding window mechanism, and the input sequence of the recent period (15 minutes) is updated every minute, continuously inputting new data to the system and automatically rolling out new risk scores. The model provides a standard RESTful API interface, which supports the connection with a hospital information system (HIS), a central monitoring system and an anesthesia workstation. The system can be extended and integrated into an intraoperative risk board platform, supports multi-task parallel prediction, including intraoperative hypotension, postoperative complications, hospitalization time, etc., and provides comprehensive support for intraoperative intelligent risk management.

[0171] Embodiment 2

[0172] Figure 2 An embodiment of the method for real-time early warning of dynamic adverse events of a patient in surgery is shown, which comprises the following steps:

[0173] S1, receiving a monitoring record sequence of a plurality of variables from a plurality of clinical devices, denoted as variable features Wherein t = 1, …, T represents a time step, d = 1, …, D represents a variable dimension, and the variables at least include blood pressure, heart rate, respiratory rate, blood oxygen saturation 2, electroencephalogram monitoring, anesthetic gas concentration, medication time and dose; all the variables are uniformly resampled to a fixed 1-minute time step to form an equidistant time axis, and on this basis, a mask matrix M(t, d) is constructed to mark whether there is an observation value of the variable at time t. If there is, the corresponding element is 1, otherwise it is 0.

[0174] For the observation time interval of each variable, a time interval vector is introduced The definition is as follows:

[0175]

[0176] where τ is the index of the historical time point less than the current time t, used to find the time point when the variable was last observed.

[0177] The vector reflects the interval distance between the current time point and the last observation of the variable. To pass the time decay signal to the model, the module further calculates the exponential decay embedding which is in the form of

[0178]

[0179] where β is a learnable parameter to adjust the influence strength of the time interval on the model representation.

[0180] Finally, the structured triplets {X, M, Γ} are output, where Γ is the time-aware embedding result integrated from the time decay embedding vectors of all variables at each time step:

[0181]

[0182] S2, according to the variable feature X and the time embedding Γ, the dimensional features of each time step are spliced and fused to form an initial variable representation matrix H with dimensions Dxd (0) , using graph structure modeling, each variable is abstracted as a node in the graph, and the adjacency matrix encodes the prior relationship between variables or the correlation learned dynamically. The adjacency matrix A can be a fixed graph calculated based on Pearson correlation coefficient, mutual information, etc., or a dynamic graph learned end-to-end by the model. In this embodiment, a learnable adjacency matrix is used to represent:

[0183]

[0184] Degree matrix:

[0185] where A is the prior relationship matrix between variables, which can be obtained through physiological knowledge or data-driven learning, I is the identity matrix representing the self-loop connection, and the degree matrix is used for graph convolution operation after normalization;

[0186] The graph convolutional neural network GCN update formula is:

[0187]

[0188] where H (l) is the node feature matrix of the lth layer; W (l) is the weight matrix of the lth layer; and σ is the ReLU activation function.

[0189] To enhance the variable structure modeling capability, the embodiment adopts a graph convolutional neural network GCN (Graph Convolutional Network) for feature extraction, and the calculation method is as follows:

[0190]

[0191] In the formula: denotes an adjacency matrix with a self-loop, is a degree matrix, and W (0) is a learnable transformation weight, and σ(·) denotes an activation function, H (1) denotes a node feature representation matrix obtained by the graph convolutional neural network GCN after first layer convolution calculation, each row of the matrix corresponds to a feature vector of a variable node after fusing information of neighbor nodes; through multi-layer graph convolution superposition operation, a variable graph feature sequence is finally output

[0192] S3, according to the variable graph feature sequence A bidirectional long short-term memory network BiLSTM is adopted to model dynamic information of historical and future contexts simultaneously, and the output of each time step is denoted as wherein and denote forward and backward LSTM outputs, respectively;

[0193] A multi-head self-attention mechanism is also adopted to establish global dependencies between all time steps, specifically, the LSTM output is taken as input of query, key and value, and the calculation is as follows:

[0194]

[0195] MultiHead (Q, K, V) = Concat (head1,..., head h ) W o

[0196] In the formula: Q = h t W Q , K = h t W K , V = h t W v , and W o is a linear transformation matrix of multi-head attention output, used for dimension transformation on the spliced result;

[0197] W Q is a query (Query) mapping matrix, used for mapping the input hidden state h t to a query vector Q; W KThis is a key mapping matrix used to map the input hidden state h. t Mapped to key vector K; W V This is a value mapping matrix used to map the input hidden state h. t These are mapped to value vectors V. They are all learnable parameter matrices that are automatically updated during training.

[0198] The final output context feature sequence H = {h} t};

[0199] S4. The variable structural feature g t With contextual features h t The fusion process employs a gating mechanism, generating the fusion weight z using the sigmoid function. t The fusion method is as follows:

[0200] z t =σ(W z [g t h t ]+b z )

[0201]

[0202] In the formula: For the fused feature representation, W z For learnable weights, b z For bias terms, ⊙ indicates element-wise multiplication;

[0203] Feature representation after fusion After transformation by a fully connected layer, a risk score per minute is output using the sigmoid function.

[0204]

[0205] In the formula, the output layer parameter W0 is the weight mapping matrix from features to the prediction space, and b0 is the corresponding bias term.

[0206] The final output is a risk score sequence with continuous time steps. Used for clinical early warning

[0207] S5. Use the integral gradient method to perform feature attribution analysis on the predicted output. By comparing the changes in the model output during the smooth transition of variables from the reference input x′ to the actual input x, determine the feature pairs. The degree of contribution. Specifically, based on the model input x and the reference input x′, the attribution value of the i-th feature is calculated under the model prediction function F(·), and the integral expression is as follows:

[0208]

[0209] where x represents the actual input feature vector of the model at a certain time step, containing the values of all monitoring variables at that time point; x' represents the reference input, which is the benchmark feature vector for the attribution calculation; IG i (x) represents the attribution value of the i-th input feature when generating the current prediction result , reflecting the positive or negative impact of the feature on the prediction result and its size; a is the interpolation coefficient of the integral path, with a value range of [0, 1], used to represent the linear interpolation progress from the reference input x' to the actual input x; d a is the integral infinitesimal, representing the calculation process of gradually accumulating the gradient in the process of a from 0 to 1.

Claims

1. A real-time early warning system for dynamic adverse events in patients during surgery, characterized in that, It includes a unified intraoperative data preprocessing module, a variable structure modeling module, a time series modeling module, a feature fusion and risk prediction module, and a prediction interpretability module, among which: The intraoperative data unified preprocessing module is used to receive monitoring record sequences of multiple variables from multiple clinical devices, denoted as variable features X, uniformly resample all variables to a fixed time step, and construct a mask matrix M(t,d) to mark whether the variable has an observation value at time t. Finally, the structured triple {X,M,Γ} is output, where the time embedding Γ is the time-aware embedding result integrated by the time decay embedding vector of all variables at each time step. The variable structure modeling module receives the variable features X and the time embedding Γ, and concatenates and fuses the features of each dimension at each time step to form an initial variable representation matrix H. (0) We use graph structure modeling and graph convolutional neural network (GCN) for feature extraction to obtain the node feature representation matrix H, which represents the node feature representation matrix obtained after the first convolutional layer of the GCN. (1) Through multi-layer graph convolution stacking operations, the final output is a variable graph feature sequence. The time-series modeling module is used to receive the variable graph feature sequence. A bidirectional long short-term memory (BiLSTM) network is used to model the dynamic information of both historical and future contexts. The output at each time step is denoted as... in and These represent the forward and backward LSTM outputs, respectively, and the final output context feature sequence H = {h} t }; The feature fusion and risk prediction module is used to integrate the variable structure features g generated by the variable structure modeling module. t Context features h output by the time series modeling module t The system integrates data and employs a gating mechanism, outputting a risk score every minute via a sigmoid function. The final output is a risk score sequence with continuous time steps. The predictive interpretability module is used to perform feature attribution analysis on the predicted output using the integral gradient method. By comparing the changes in the model output during the smooth transition of variables from the reference input x′ to the actual input x, it determines the feature pairs. The extent of their contribution.

2. The real-time early warning system for dynamic adverse events in patients during surgery according to claim 1, characterized in that: The variable characteristics Where t = 1, ..., T represents the time step, and d = 1, ..., D represents the variable dimension. The variables include at least blood pressure, heart rate, respiratory rate, blood oxygen saturation, electroencephalogram monitoring, anesthetic gas concentration, medication time and dosage. The time embedding Γ takes the following form: In the formula, For exponentially decaying embedding, its form is as follows: In the formula, β is a learnable parameter used to adjust the strength of the influence of the time interval on the model representation. This is a time interval vector, reflecting the distance between the current time point and the most recent observation of the variable, and is defined as follows: In the formula, τ is the index of a historical time point less than the current time t.

3. The real-time early warning system for dynamic adverse events in patients during surgery according to claim 1, characterized in that, The variable structure modeling module uses graph structure modeling specifically as follows: Each variable is abstracted as a node in a graph, using an adjacency matrix. Prior relationships or correlations learned dynamically between encoded variables are represented using a learnable adjacency matrix: Degree matrix: Where A is the prior relation matrix between variables, I is the identity matrix representing self-loop connections, and the associativity matrix is... After normalization, it is used for graph convolution operations; The update formula for Graph Convolutional Neural Network (GCN): Where: H (l) W is the feature matrix of the nodes in the l-th layer; (l) Let σ be the weight matrix of the l-th layer; σ is the activation function. The variable structure modeling module uses a graph convolutional neural network (GCN) for feature extraction. The specific calculation method is as follows: In the formula: This represents the adjacency matrix with self-loops. For its degree matrix, W (0) The transform weights are learnable, and σ(·) represents the activation function.

4. The real-time early warning system for dynamic adverse events in patients during surgery according to claim 1, characterized in that: The time series modeling module is also used to establish global dependencies across all time steps using a multi-head self-attention mechanism. Specifically, the LSTM output is used as the input for the query, key, and value, and the calculation is as follows: MultiHead(Q,K,V)=Concat(head1,...,head h )W O In the formula: Q = h t W Q K = h t W K V = h t W v W 0 W is the linear transformation matrix for the multi-head attention output. Q To query the mapping matrix, W K Let W be the key mapping matrix. V The value mapping matrix.

5. The real-time early warning system for dynamic adverse events in patients during surgery according to claim 1, characterized in that: The variable structural feature g t With contextual features h t The fusion process specifically involves generating the fusion weight z using the sigmoid function. t The fusion method is as follows: z t =σ(W z [g t ;h t ]+b z ) In the formula: For the fused feature representation, W z For learnable weights, b z For bias terms, ⊙ indicates element-wise multiplication; The risk score is output every minute using the sigmoid function. Specifically, the fused feature representation After transformation by a fully connected layer, the output is obtained using the sigmoid function. In the formula, the output layer parameter W0 is the weight mapping matrix from features to the prediction space, and b0 is the corresponding bias term.

6. The real-time early warning system for dynamic adverse events in patients during surgery according to claim 1, characterized in that, The prediction interpretability module is used to perform feature attribution analysis on the prediction output using the integral gradient method, specifically as follows: Based on the model input x and the reference input x′, the attribution value of the i-th feature is calculated under the model prediction function F(·), and the integral expression is as follows: In the formula, x represents the actual input feature vector of the model at a certain time step, containing the values ​​of all monitored variables at that time point; x′ represents the reference input, serving as the baseline feature vector for attribution calculation, IG i (x) represents the i-th input feature in generating the current prediction result. The attribution value reflects the positive or negative impact of the feature on the prediction result and its magnitude; α is the interpolation coefficient of the integral path, with a value range of [0,1], used to represent the linear interpolation progress from the reference input x′ to the actual input x; dα is the integral infinitesimal element, representing the calculation process of gradually accumulating the gradient as α changes from 0 to 1.

7. A method for real-time early warning of dynamic adverse events in patients during surgery, characterized in that, Includes the following steps: S1. Receive monitoring record sequences of multiple variables from multiple clinical devices, denoted as variable features X. Resample all variables to a fixed time step and construct a mask matrix M(t,d) to mark whether the variable has an observation value at time t. Finally, output the structured triple {X,M,Γ}, where the time embedding Γ is the time-aware embedding result integrated by the time decay embedding vector of all variables at each time step. S2. Based on the variable features X and the time embedding Γ, the features of each dimension at each time step are concatenated and fused to form an initial variable representation matrix H. (0) We use graph structure modeling and graph convolutional neural network (GCN) for feature extraction to obtain the node feature representation matrix H, which represents the node feature representation matrix obtained after the first convolutional layer of the GCN. (1) Through multi-layer graph convolution stacking operations, the final output is a variable graph feature sequence. S3. Based on the variable graph feature sequence A bidirectional long short-term memory (BiLSTM) network is used to model the dynamic information of both historical and future contexts. The output at each time step is denoted as... in and These represent the forward and backward LSTM outputs, respectively, and the final output context feature sequence H = {h} t }; S4. The variable structural feature g t With contextual features h t The system integrates data and employs a gating mechanism, outputting a risk score every minute via a sigmoid function. The final output is a risk score sequence with continuous time steps. S5. Use the integral gradient method to perform feature attribution analysis on the predicted output. By comparing the changes in the model output during the smooth transition of variables from the reference input x′ to the actual input x, determine the feature pairs. The extent of their contribution.

8. The method for real-time early warning of dynamic adverse events in patients during surgery according to claim 7, characterized in that: The variable characteristics Where t = 1, ..., T represents the time step, and d = 1, ..., D represents the variable dimension. The variables include at least blood pressure, heart rate, respiratory rate, blood oxygen saturation, EEG monitoring, anesthetic gas concentration, medication time and dosage. The time embedding Γ takes the following form: In the formula, For exponentially decaying embedding, its form is as follows: In the formula, β is a learnable parameter used to adjust the strength of the influence of the time interval on the model representation. This is a time interval vector, reflecting the distance between the current time point and the most recent observation of the variable, and is defined as follows: In the formula, τ is the index of a historical time point less than the current time t; The graph structure modeling used in step S2 specifically involves: Each variable is abstracted as a node in a graph, using an adjacency matrix. Prior relationships or correlations learned dynamically between encoded variables are represented using a learnable adjacency matrix: Degree matrix: Where A is the prior relation matrix between variables, I is the identity matrix representing self-loop connections, and the associativity matrix is... After normalization, it is used for graph convolution operations; The update formula for Graph Convolutional Neural Network (GCN): Where: H (l) W is the feature matrix of the nodes in the l-th layer; (l) Let σ be the weight matrix of the l-th layer; σ is the activation function. The variable structure modeling module uses a graph convolutional neural network (GCN) for feature extraction. The specific calculation method is as follows: In the formula: This represents the adjacency matrix with self-loops. For its degree matrix, W (0) The transform weights are learnable, and σ(·) represents the activation function.

9. A method for real-time early warning of dynamic adverse events in patients during surgery according to claim 7, characterized in that: In step S3, a multi-head self-attention mechanism is also used to establish global dependencies across all time steps. Specifically, the LSTM output is used as the input for the query, key, and value, and the calculation is as follows: MultiHead(Q,K,V)=Concat(head1,...,head h )W O In the formula: Q = h t W Q K = h t W K V = h t W v W 0 W is the linear transformation matrix for the multi-head attention output. Q To query the mapping matrix, W K Let W be the key mapping matrix. V The value mapping matrix.

10. A method for real-time early warning of dynamic adverse events in patients during surgery according to claim 7, characterized in that: In step S4, the variable structural feature g t With contextual features h t The fusion is performed by generating the fusion weight z using the sigmoid function. t The fusion method is as follows: z t =σ(W z [g t ;h t ]+b z ) In the formula: For the fused feature representation, W z For learnable weights, b z For bias terms, ⊙ indicates element-wise multiplication; The risk score is output every minute using the sigmoid function. Specifically, the fused feature representation After transformation by a fully connected layer, the output is obtained using the sigmoid function. In the formula, the output layer parameter W0 is the weight mapping matrix from features to the prediction space, and b0 is the corresponding bias term; In step S5, the integral gradient method is used to perform feature attribution analysis on the predicted output. Specifically, based on the model input x and the reference input x′, the attribution value of the i-th feature is calculated under the model prediction function F(·). The integral expression is as follows: In the formula, x represents the actual input feature vector of the model at a certain time step, containing the values ​​of all monitored variables at that time point; x′ represents the reference input, serving as the baseline feature vector for attribution calculation, IG i (x) represents the i-th input feature in generating the current prediction result. The attribution value reflects the positive or negative impact of the feature on the prediction result and its magnitude; α is the interpolation coefficient of the integral path, with a value range of [0,1], used to represent the linear interpolation progress from the reference input x′ to the actual input x; dα is the integral infinitesimal element, representing the calculation process of gradually accumulating the gradient as α changes from 0 to 1.

Images