An interpretable early warning method for sepsis based on monitoring data

By introducing the Actor-Critic framework and Hilbert-Schmidt independence criterion in the sepsis early warning model, combined with the integrated gradient algorithm, the problem of insufficient interpretability of the deep learning model is solved, and the accuracy of sepsis risk is achieved and interpretability warning improve the generalization ability and prediction accuracy of the model.

CN117936075BActive Publication Date: 2024-11-15CHONGQING INST OF GREEN & INTELLIGENT TECH CHINESE ACAD OF SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410087228.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-19
Publication Date
2024-11-15
Estimated Expiration
2044-01-19

AI Technical Summary

Technical Problem

The existing deep learning-based early warning models for sepsis lack interpretability and fail to provide clear predictive outcomes and decision support, resulting in difficulties for clinicians in treatment.

Method used

Using the Actor-Critic framework based on monitoring data to reinforce learning networks and combine the benchmark model of Hilbert-Schmidt independence criterion, the interpretability warning of sepsis risk is achieved by improving the orthogonalization of reward functions and features. And visual analysis is performed using integrated gradient algorithm.

Benefits of technology

It improves the generalization ability and interpretability of the model, provides accurate warnings on sepsis risk, reduces surgical risks, and provides timely and accurate decision-making support for medical workers.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117936075B_ABST
    Figure CN117936075B_ABST
Patent Text Reader

Abstract

The present invention is an interpretable early warning method for sepsis based on monitoring data, which belongs to the field of medical big data. It is implemented by an interpretable sepsis model based on monitoring data; the interpretable sepsis model based on monitoring data is composed of a reinforcement learning network of an Actor-Critic framework in parallel with a benchmark model. The steps of the method are: S1: construct a historical data set by statistical static features and time series data; S2: establish and train a reinforcement learning network of an Actor-Critic framework; S3: establish and train a benchmark model; S4: establish and optimize a joint loss function; S5: predict the risk value of sepsis in patients; S6: perform classified early warning; S7: visualize and interpret the early warning. The method model of the present invention has good generalization ability, interpretability and prediction accuracy, and can efficiently assist medical workers in realizing timely and accurate early warning of sepsis in patients and reduce the occurrence of surgical risks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to an interpretable early warning method for sepsis based on monitoring data, belongs to the field of medical big data, and is particularly suitable for an interpretable early warning method for sepsis based on monitoring data. Background Art

[0002] Sepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection: due to the uncertainty of pathogenesis, the diversity of symptoms, and the complexity of treatment, accurate prediction and decision support have always been a difficult problem. It is also one of the most challenging problems in the field of disease prediction and one of the global health priorities. According to the World Health Organization's estimate in 2022, about 30 million people worldwide suffer from sepsis each year, of which 6 million die from it, with a mortality rate of up to 20%. Sepsis is acute and serious, and must be treated immediately when it occurs. Every hour of delay in treatment will increase the mortality rate by about 6%. However, in addition to general guidelines, there is no intelligent decision support tool for the best treatment of sepsis that can be explained so far. If the occurrence of sepsis can be warned in advance, accurately and explainably, the risk of sepsis can be effectively reduced.

[0003] Many research teams have been seeking to use artificial intelligence (AI) to identify patients with early symptoms of sepsis, such as the Epic Sepsis Model (ESM)[1], which is widely used in the United States. Researchers analyzed 38,455 patients hospitalized at the University of Michigan and found that the sensitivity of ESM was 33%, the specificity was 83%, the positive predictive value was 12%, the negative predictive value was 95%, and the area under the curve was 0.63. A recent study published in Nature Communications[2] combined natural language processing (NLP) analysis of doctors' clinical notes with structured (Electronic Medical Record, EMR) data to improve the ability to predict sepsis risk. Compared with clinicians, the artificial intelligence algorithm model can increase the early detection rate of sepsis by 32%, and has a high prediction accuracy of 94%, 87% and 87% in the first 12 hours before the patient is diagnosed, which can save golden time for clinical treatment. Duke University Hospital's Sepsis Watch[3] identifies cases based on dozens of variables (including vital signs, laboratory test results and medical history). Its training data includes 50,000 patient records and more than 32 million data points. Information is extracted from the patient's medical records every 5 minutes. It can identify early cases of sepsis and issue alarms, thus providing intensive real-time analysis that human doctors cannot provide. However, these studies lack multi-center experimental verification, and the model is based on "black box" deep learning technology. The prediction results are not interpretable and transparent enough to guide clinicians to make effective decisions. Due to the high-risk nature of medical applications, it is not enough to have only a deep model that can produce the correct answer without explanation. In current research, interpretability in medicine is mainly divided into post-hoc interpretability analysis and ad hoc interpretability modeling.

[0004] Therefore, a deep learning method that can perform real-time online analysis and is interpretable to assist in the early warning of sepsis risk is of great research and application value and can greatly protect people's lives.

[0005] [1].Habib AR,Lin AL,Grant RW:The Epic Sepsis Model Falls Short-TheImportance of External Validation.JAMA internal medicine 2021,181(8):1040-1041.

[0006] [2].Goh KH,Wang L,Yeow AYK,Poh H,Li K,Yeow JJL,Tan GYH: Artificialintelligence in sepsis early prediction and diagnosis using unstructured data in healthcare.Nature communications 2021,12(1):711.

[0007] [3].Kuan WS, Ibrahim I, Leong BS, Jain S, Lu Q, Cheung YB, Mahadevan M: Emergency Department Management of Sepsis Patients: A Randomized, Goal-Oriented, Noninvasive Sepsis Trial. Annals of emergency medicine 2016, 67(3): 367-378.e363. Summary of the invention

[0008] In view of this, the present invention provides an interpretable early warning method for sepsis based on monitoring data, which aims to analyze and process monitoring data based on a reinforcement learning model, accurately warn patients of sepsis risk, and provide interpretable results. Overfitting is avoided by improving the reward function, and the Hilbert-Schmidt Independent Criterion (HSIC) is introduced to orthogonalize the feature space, thereby improving the generalization ability and interpretability of the model.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] The method for explaining sepsis early warning based on monitoring data is characterized in that it is implemented by an explainable sepsis model based on monitoring data; the explainable sepsis model based on monitoring data is composed of a reinforcement learning network of an Actor-Critic framework in parallel with a benchmark model; the reinforcement learning network of the Actor-Critic framework is composed of an Actor network and a Critic network in series; the Actor network is composed of a convolutional neural network 1 (CNN-1) connected in series with a random sampler, the convolutional neural network 1 is used to generate sampling probabilities, and the random sampler extracts features according to the generated sampling probabilities; the Critic network is a convolutional neural network 2 (CNN-2), which is used to predict the outcome of sepsis using the extracted features; the benchmark model is composed of a temporal convolutional network (TCN) and a fully connected layer network (Full-connected network) in parallel; the Actor network and the input end of the benchmark model are connected as data; the output end of CNN-1 in the Actor network is connected to the input end of the fully connected layer network; the loss function of the reinforcement learning network of the Actor-Critic framework is connected to the loss function of the benchmark model at the output end, and they work together to adjust the parameters of CNN1.

[0011] The specific steps of this method are as follows:

[0012] S1: Statistics of static characteristics of different patients and monitoring devices to monitor patients and obtain time series data of vital signs {x t |t=1,…,T} to construct historical data set;

[0013] S2: Establish a reinforcement learning network with an Actor-Critic framework, and use historical data sets for training to optimize network parameters;

[0014] S3: Establish a benchmark model, use the Hilbert-Schmidt Independent Criterion (HSIC) and cross entropy to establish the loss function of the benchmark model, and use the historical data set data for training;

[0015] S4: Combine the loss function of the reinforcement learning network of the Actor-Critic framework with the loss function of the benchmark model to establish a joint loss function, use historical data sets for training, and optimize the Actor network parameters of the reinforcement learning network of the Actor-Critic framework;

[0016] S5: Based on the patient's static characteristics and the time series data of the patient's vital signs obtained by the monitoring device in real time, the reinforcement learning network of the Actor-Critic framework optimized in step S4 is used to predict the time period The patient's risk of sepsis;

[0017] S6: Use classifier to classify time periods Classify and warn the risk values ​​of sepsis in patients within the range;

[0018] S7: The integrated gradient algorithm is used for visualization and interpretability analysis of early warning of sepsis in patients.

[0019] in, x i,t is the value of the i-th monitoring data at time t, 1≤i≤d; T is the last moment of monitoring; s and d are positive integer dimensions.

[0020] Furthermore, the working process of the reinforcement learning network of the Actor-Critic framework described in step S2 is:

[0021] S201: Calculate the statistical features of the time series data at each moment using statistical methods, and concatenate the static features with the statistical features at that moment as input features of the reinforcement learning network of the Actor-Critic framework;

[0022] S202: CNN1 generates a d+s-dimensional probability vector based on the input features;

[0023] S203: The random sampler randomly selects some features from the input features according to the probability vector as the input of CNN2, and outputs the predicted probability of sepsis of the patient p1∈[0,1];

[0024] S204: Combine the manually labeled time series data with sepsis labels of 0 or 1, and use cross entropy to establish the probability p1 of patient sepsis and the loss function L of sepsis labels critic (p1, label);

[0025] S205: Using L critic (p1, label), optimize the parameters of the Critic network by combining manually labeled time series data;

[0026] The statistical characteristics include: mean and variance {Dx t |t=1,…,T}; where

[0027] Further, the working process of the benchmark model described in step S3 is:

[0028] S301: Calculate the statistical characteristics of the time series data at each moment using statistical methods;

[0029] S302: Use the time convolution network (TCN) to extract the time features of each moment of the time series data

[0030] S303: Input the statistical features, time features and static features into the fully connected layer, and output the probability p2∈[0,1] of the patient's sepsis;

[0031] S304: Combine the manually labeled time series data with sepsis labels of 0 or 1, and use cross entropy to establish the probability p2 of patient sepsis and the loss function L of sepsis labels cross (p2, label);

[0032] S305: Analyze the statistical characteristics and time characteristics using HSIC to obtain the HSIC loss function L HISC ;

[0033] S306: Calculate the loss function L of the benchmark model base =L cross (p2,label)+L HISC , and combined with manually labeled time series data to optimize the TCN and fully connected layer parameters of the benchmark model.

[0034] Furthermore, the HSIC described in step S305 is used to measure the distribution difference between the statistical feature x and the time feature y, and its loss function is: Among them: Cross-covariance operator μ x =E x φ(x), is the tensor product, φ(x), For a given nonlinear mapping of x and y, E x 、E y is the mean operator about x and y; E xy is the expectation operator for the joint distribution of x and y; for any matrix A = [a i,j ],have

[0035] Furthermore, the joint loss function described in step S4 is: L = L critic (p1,label)+L base .

[0036] Furthermore, the classifier described in step S6 is a support vector machine, specifically: the time period The risk value of sepsis in patients within the time period is taken as a vector, and the elements are assigned weights and classified; before application, the time period needs to be used The classifier is trained using manually labeled time series data.

[0037] Further, step S7 is specifically as follows: first, using an integrated gradient algorithm, the static characteristics of the patient and the attribute characteristics of the time series data of the patient's vital signs monitored by the monitoring device are visualized through gradient visualization; then, the prediction results of the model are explained by analyzing the contribution of a single feature of the input data to the risk value of sepsis in the patient.

[0038] Preferably, the Critic network can directly output a binary classification warning result of the patient's sepsis risk, which can reduce the classifier processing in step S5 and improve efficiency.

[0039] The beneficial effects of the present invention are as follows: the present invention provides an interpretable early warning method for sepsis based on monitoring data, takes static data and dynamic data as the starting point, introduces the Hilbert-Schmidt independence criterion on the basis of the traditional reinforcement learning algorithm, realizes the orthogonalization of the feature space, improves the generalization ability, interpretability and prediction accuracy of the model, improves the reward function to avoid overfitting, and realizes visualization and interpretability by using an integrated gradient algorithm. The method of the present invention can efficiently assist medical workers in realizing timely and accurate early warning of sepsis in patients and reduce the occurrence of surgical risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to make the purpose and technical solution of the present invention, the present invention provides the following drawings for explanation:

[0041] Figure 1 : is a framework diagram of the sepsis interpretability model based on monitoring data in the present invention; wherein, the solid arrows represent the transmission data, and the dot-dashed arrows represent the training;

[0042] Figure 2 is a flow chart of the method of the present invention;

[0043] Figure 3 This is a diagram of the benchmark model architecture in Example 1 of the present invention;

[0044] Figure 4 : is a training result diagram of the joint loss function in Example 1 of the present invention; wherein the horizontal axis is the number of training iterations, and the vertical axis is the value of the joint loss function;

[0045] Figure 5 The figure is a training result diagram of the accuracy in Example 1 of the present invention; wherein the horizontal axis is the number of training iterations, and the vertical axis is the accuracy value (unit: %);

[0046] Figure 6 The diagram is a visualization and explanatory result diagram output in Example 1 of the present invention; wherein (a) is a heat map of the time series data input of the visualized vital signs; (b) is a heat map of the visualized static feature input; (c) is the interpretable result output, and the horizontal axis is 30 attribute data of corresponding serial numbers. DETAILED DESCRIPTION

[0047] Example 1: A sepsis dataset was collected for surgical patients in a hospital in a certain city from 2014 to 2018. The number of patients in the dataset is 1,843, including 405 sepsis-positive patients and 1,438 sepsis-negative patients (manual annotation has been completed); 80% of the data is used for the training set and 20% for the test set. The patient data includes preoperative examination and demographic data, i.e., the static characteristics of the patient, and the specific characteristic attributes are shown in Table 1; and the patient's intraoperative monitoring data, i.e., the time series data of vital signs, and the specific characteristic attributes are shown in Table 2, where each attribute corresponds to 2 data.

[0048] Table 1 Static characteristic attributes of patients

[0049]

[0050]

[0051] Table 2 Time series data attributes of patients' vital signs

[0052] Serial number Property Name 21 Heart rate 2 Systolic blood pressure 3 Diastolic blood pressure 4 Blood oxygen saturation 5 Central venous pressure

[0053] For the above patients, the risk warning and explainability of sepsis are carried out. The present invention provides a "sepsis explainable warning method based on monitoring data", combined with Figure 1, which is realized by an interpretable sepsis model based on monitoring data; the interpretable sepsis model based on monitoring data is composed of a reinforcement learning network of an Actor-Critic framework in parallel with a benchmark model; the reinforcement learning network of the Actor-Critic framework is composed of an Actor network and a Critic network in series; the Actor network is composed of a convolutional neural network 1 (CNN-1) in series with a random sampler, the convolutional neural network 1 is used to generate sampling probabilities, and the random sampler extracts features according to the generated sampling probabilities; the Critic network is a convolutional neural network 2 (CNN-2), which is used to predict the results of sepsis using the extracted features; the benchmark model is composed of a TCN and a fully connected layer network in parallel; the input ends of the Actor network and the benchmark model are connected to data; the output end of CNN-1 in the Actor network is connected to the input end of the fully connected layer network; the loss function of the reinforcement learning network of the Actor-Critic framework is connected to the loss function of the benchmark model at the output end, and they work together to adjust the parameters of CNN1.

[0054] The sepsis interpretability model based on monitoring data was implemented in PyTorch, and all experiments were performed on a server with an Ubuntu 16.04LTS operating system, an Intel Xeon e5-2650 V4 processor, and an Nvidia GTX 1080Ti GPU (64G memory), and the performance of the model was evaluated by a confusion matrix. In order to better demonstrate the characteristics of the method of the present invention, this example performed accuracy (ACC), area under the Roc curve (AUC), and average accuracy score (APR) analysis based on manual annotation and prediction results to test the accuracy of the early warning classification of the method of the present invention.

[0055] Combination Figure 2 , the specific steps are as follows:

[0056] Step 1: Establish a reinforcement learning network of the Actor-Critic framework, and use the training set data for training to optimize the network parameters.

[0057] The Actor network has 30 input nodes, 30 output nodes, 100 hidden nodes and a 2-layer fully connected neural network. The Critic network has the same structure as the Actor network, except that the output layer has 2 nodes, indicating whether the patient has sepsis warning. During training, 100 iterations of the back propagation algorithm were used, the initial learning rate was set to 0.01, and it was dynamically adjusted through exponential learning in the later stage, and the batch size was set to 32.

[0058] Specifically,

[0059] (1) Using statistical methods to calculate the statistical features of time series data at each moment, and concatenating the static features with the statistical features at that moment as the input features of the reinforcement learning network of the Actor-Critic framework;

[0060] (2) CNN1 generates a d+s-dimensional probability vector based on the input features;

[0061] (3) The random sampler randomly selects some features from the input features according to the probability vector as the input of CNN2, and outputs the predicted probability of sepsis p1∈[0,1] of the patient;

[0062] (4) Combined with the sepsis label (0 or 1) of the manually labeled time series data, the cross entropy is used to establish the probability p1 of patient sepsis and the loss function L of the sepsis label critic (p1,label);

[0063] (5) Using L critic (p1, label), optimize the parameters of the Critic network by combining manually labeled time series data;

[0064] (6) Wherein, the statistical characteristics include: mean and variance {Dx t |t=1,…,T}; where

[0065]

[0066] Step 2: Establish a benchmark model, such as Figure 3 As shown, the Hilbert-Schmidt independence criterion and cross entropy are used to establish the loss function of the benchmark model, and the training set data is used for training.

[0067] Specifically,

[0068] (1) Use statistical methods to calculate the statistical characteristics of time series data at each moment;

[0069] (2) Use the temporal convolutional network (TCN) to extract the temporal features of the time series data at each moment

[0070] Among them, TCN is a 10-layer network.

[0071] (3) Input the statistical features, temporal features and static features into the fully connected layer, and output the probability of sepsis p2∈[0,1];

[0072] (4) Combined with the sepsis label (0 or 1) of the manually labeled time series data, the cross entropy is used to establish the probability p2 of patient sepsis and the loss function L of the sepsis label cross (p2, label);

[0073] (5) Analyze the statistical characteristics and time characteristics using HSIC to obtain the HSIC loss function L HISC ;

[0074] (6) Calculate the loss function L of the baseline model base =L cross (p2,label)+L HISC , and combined with manually labeled time series data to optimize the TCN and fully connected layer parameters of the benchmark model.

[0075] Furthermore, the HSIC described in step S305 is used to measure the distribution difference between the statistical feature (denoted as x) and the time feature (denoted as y), and its loss function is: Among them: Cross-covariance operator μ x =E x φ(x), is the tensor product, φ(x), For a given nonlinear mapping of x and y, E x 、E y is the mean operator about x and y; for any matrix A = [a i,j ],have

[0076] Step 3: Combine the loss function of the reinforcement learning network of the Actor-Critic framework with the loss function of the benchmark model to establish a joint loss function, use the training set for training, and optimize the Actor network parameters of the reinforcement learning network of the Actor-Critic framework.

[0077] The joint loss function is: L = L critic (p1,label)+L base .

[0078] Joint loss function L base like Figure 4 The training accuracy results are shown in Figure 5As shown in the figure. As the model is trained, the accuracy and loss value of the 10th iteration have converged; the accuracy of the baseline model converges to 97% after 10 iterations; the accuracy of the reinforcement learning network fluctuates to a certain extent in the early stage, but based on the prediction results of the baseline model, the parameters of the reinforcement learning network can be adjusted in reverse to make the model more accurate. As can be seen from the figure, after 80 iterations of training, the accuracy of the reinforcement learning network has also stabilized.

[0079] Step 4: Use the test set as input and use the reinforcement learning network of the Actor-Critic framework optimized in step 3 to predict the time period Patients at risk for sepsis.

[0080] In order to better demonstrate the method of the present invention, this embodiment introduces models commonly used in the medical field for comparative experiments, specifically including: Gaussian Bayes model (NB), k-nearest neighbor model (KNN), logistic regression (LR), random forest (RF), decision tree (DT), multilayer perceptron (MLP), GBST, ADBOOST method, and the experimental results are shown in Table 3.

[0081] Table 3 Comparative experimental results

[0082] method AUC APR ACC NB 0.95 0.97 0.97 KNN 0.95 0.97 0.95 LR 0.95 0.95 0.95 RF 0.95 0.95 0.95 DT 0.95 0.95 0.95 MLP 0.95 0.95 0.95 GBDT 0.95 0.95 0.95 ADBOOST 0.95 0.95 0.95 Method of the present invention 0.99 0.99 0.99

[0083] Step 5: Use the integrated gradient algorithm to perform visualization and interpretable analysis on the early warning of sepsis in patients.

[0084] First, we use the integrated gradient algorithm to visualize the static characteristics of the patient and the attribute characteristics of the time series data of the patient's vital signs monitored by the monitoring device through gradient visualization; then, we explain the prediction results of the model by analyzing the contribution of a single feature of the input data to the patient's sepsis risk value. For a certain patient, the visualization and interpretability results are shown in the figure below: Figure 6 As shown; through analysis (c), it can be seen that the data of 13 attributes are closely related to the occurrence of sepsis in the patient, which can guide medical workers to analyze the factors related to these attributes and take corresponding measures.

[0085] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. An interpretable early warning method for sepsis based on monitoring data, characterized in that: The present invention is realized by an interpretable sepsis model based on monitoring data; the interpretable sepsis model based on monitoring data is composed of a reinforcement learning network of an Actor-Critic framework in parallel with a benchmark model; the reinforcement learning network of the Actor-Critic framework is composed of an Actor network and a Critic network in series; the Actor network is composed of a convolutional neural network 1 (CNN-1) in series with a random sampler, the convolutional neural network 1 is used to generate sampling probabilities, and the random sampler extracts features according to the generated sampling probabilities; the Critic network is a convolutional neural network 2 (CNN-2), which is used to predict the results of sepsis using the extracted features; the benchmark model is composed of a temporal convolutional network (TCN) and a fully connected layer network in parallel; the input ends of the Actor network and the benchmark model are connected as data; the output end of CNN-1 in the Actor network is connected to the input end of the fully connected layer network; the loss function of the reinforcement learning network of the Actor-Critic framework is connected to the loss function of the benchmark model at the output end, and they work together to adjust the parameters of CNN1; The method comprises the following steps: S1: Statistics of static characteristics of different patients and monitoring devices to monitor patients and obtain time series data of vital signs {x t |t=1,…,T} to construct historical data set; S2: Establish a reinforcement learning network with an Actor-Critic framework, and use historical data sets for training to optimize network parameters; S3: Establish a benchmark model, use the Hilbert-Schmidt independence criterion and cross entropy to establish the loss function of the benchmark model, and use the historical data set data for training; S4: Combine the loss function of the reinforcement learning network of the Actor-Critic framework with the loss function of the benchmark model to establish a joint loss function, use historical data sets for training, and optimize the Actor network parameters of the reinforcement learning network of the Actor-Critic framework; S5: Based on the patient's static characteristics and the time series data of the patient's vital signs obtained by the monitoring device in real time, the reinforcement learning network of the Actor-Critic framework optimized in step S4 is used to predict the time period The patient's risk of sepsis; S6: Use classifier to classify time periods Classify and warn the risk values ​​of sepsis in patients within the range; S7: Visualization and interpretability analysis of early warning of sepsis in patients using integrated gradient algorithms; in, x i,t is the value of the i-th monitoring data at time t, 1≤i≤d; T is the last moment of monitoring; s and d are positive integer dimensions; The working process of the reinforcement learning network of the Actor-Critic framework described in step S2 is: S201: Calculate the statistical features of the time series data at each moment using statistical methods, and concatenate the static features with the statistical features at that moment as input features of the reinforcement learning network of the Actor-Critic framework; S202: CNN1 generates a d+s-dimensional probability vector based on the input features; S203: The random sampler randomly selects some features from the input features according to the probability vector as the input of CNN2, and outputs the predicted probability of sepsis of the patient p1∈[0,1]; S204: Combine the manually labeled time series data with sepsis labels of 0 or 1, and use cross entropy to establish the probability p1 of patient sepsis and the loss function L of sepsis labels critic (p1,label); S205: Using L critic (p1, label), optimize the parameters of the Critic network by combining manually labeled time series data; The statistical characteristics include: mean and variance {Dx t |t=1,…,T}; where The working process of the benchmark model described in step S3 is: S301: Calculate the statistical characteristics of the time series data at each moment using statistical methods; S302: Use the time convolutional network to extract the time features of each moment of the time series data S303: Input the statistical features, time features and static features into the fully connected layer, and output the probability p2∈[0,1] of the patient's sepsis; S304: Combine the manually labeled time series data with sepsis labels of 0 or 1, and use cross entropy to establish the probability p2 of patient sepsis and the loss function L of the sepsis label cross (p2, label); S305: Analyze the statistical characteristics and time characteristics using the Hilbert-Schmidt Independent Criterion (HSIC) to obtain the HSIC loss function L HISC ; S306: Calculate the loss function L of the benchmark model base =L cross (p2,label)+L HISC , and combined with manually labeled time series data to optimize the TCN and fully connected layer parameters of the benchmark model; The HSIC described in step S305 is used to measure the distribution difference between the statistical feature x and the time feature y, and its loss function is: Among them: Cross-covariance operator μ x =E x φ(x), is the tensor product, φ(x), For a given nonlinear mapping of x and y, E x 、E y is the mean operator about x and y; E xy is the expectation operator for the joint distribution of x and y; for any matrix A = [a i,j ],have Step S7 is specifically as follows: first, using the integrated gradient algorithm, the static characteristics of the patient and the attribute characteristics of the time series data of the patient's vital signs monitored by the monitoring device are visualized through gradient visualization; then, the prediction results of the model are explained by analyzing the contribution of a single feature of the input data to the risk value of sepsis in the patient.

2. The method for interpretable early warning of sepsis based on monitoring data according to claim 1, characterized in that: The joint loss function described in step S4 is: L = L critic (p1,label)+L base .

3. The method for interpretable early warning of sepsis based on monitoring data according to claim 1, characterized in that: The classifier described in step S6 is a support vector machine, specifically: the time period The risk value of sepsis in patients within the time period is taken as a vector, and the elements are assigned weights and classified; before application, the time period needs to be used The classifier is trained using manually labeled time series data.

4. The method for interpretable early warning of sepsis based on monitoring data according to claim 1, characterized in that: The Critic network can directly output a binary classification warning result of the patient's sepsis risk, which can reduce the classifier processing in step S5 and improve efficiency.