Electronic medical record analysis method, system and terminal for myocardial infarction or heart failure patient

Abstract

The invention discloses an electronic medical record analysis method, system and terminal for a myocardial infarction or heart failure patient, and the method comprises the steps: obtaining an electronic medical record of the patient, inputting the electronic medical record of the patient into a spatial-temporal dynamic graph neural network model, and obtaining a plurality of time series diagrams; obtaining sparse adjacency matrixes and node attribute information in the plurality of time sequence diagrams, and performing spatial information learning processing on the sparse adjacency matrixes and the node attribute information to obtain a plurality of multivariate time sequence diagrams; performing down-sampling processing on the multiple multivariate time sequence diagrams to obtain multiple multivariate time sequence features; and performing integrated learning processing on the spatial-temporal dynamic graph neural network model according to the multiple multivariate time sequence features, and outputting an electronic medical record analysis result corresponding to the electronic medical record of the patient. By constructing the space-time dynamic graph neural network model, the space-time information corresponding to the physiological parameters in the electronic medical record of the patient can be dynamically extracted, and the accuracy of outputting the analysis result of the electronic medical record is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to an electronic medical record analysis method, system, terminal and computer-readable storage medium for patients with myocardial infarction or heart failure. Background Art

[0002] Heart failure (HF) is one of the common complications after acute myocardial infarction (AMI). HF is a serious and terminal stage of cardiovascular disease, especially myocardial infarction, with high disability and mortality. However, how to analyze HF caused by MI through the electronic medical records of patients with MI or HF has not yet been determined.

[0003] To solve the above problems, the existing technology generally extracts static physiological variables in the patient's electronic medical records through a neural network model and then analyzes them. However, it does not consider the interaction and dynamic changes between the variables in the electronic medical records, resulting in the obtained electronic medical record analysis results being inaccurate and unable to meet user needs.

[0004] Therefore, the prior art still needs to be improved and developed. Summary of the invention

[0005] The main purpose of the present invention is to provide an electronic medical record analysis method, system, terminal and computer-readable storage medium for patients with myocardial infarction or heart failure, aiming to solve the problem that the prior art does not take into account the interaction and dynamic changes between variables in the electronic medical records of patients with myocardial infarction or heart failure, resulting in inaccurate electronic medical record analysis results.

[0006] To achieve the above object, the present invention provides an electronic medical record analysis method for patients with myocardial infarction or heart failure, the electronic medical record analysis method for patients with myocardial infarction or heart failure comprising the following steps:

[0007] Obtaining electronic medical records of patients, and inputting the electronic medical records of patients into a spatiotemporal dynamic graph neural network model to obtain a plurality of time series graphs;

[0008] Acquire a plurality of sparse adjacency matrices and node attribute information in the time series graphs, perform spatial information learning processing on the sparse adjacency matrices and the node attribute information, and obtain a plurality of multivariate time series graphs;

[0009] Downsampling the plurality of multivariate time series graphs to obtain a plurality of multivariate time series features;

[0010] The spatiotemporal dynamic graph neural network model is subjected to integrated learning processing according to the multiple multivariate time series features, and the electronic medical record analysis results corresponding to the patient's electronic medical record are output.

[0011] Optionally, the method for analyzing the electronic medical records of patients with myocardial infarction or heart failure, wherein the step of obtaining the electronic medical records of the patients and inputting the electronic medical records of the patients into a spatiotemporal dynamic graph neural network model to obtain a plurality of time series graphs, specifically includes:

[0012] Obtaining a patient's electronic medical record, inputting the patient's electronic medical record into a spatiotemporal dynamic graph neural network model, and extracting a multivariate time series in the patient's electronic medical record through the spatiotemporal dynamic graph neural network model;

[0013] Performing multiple segmentation processing on the multivariate time series to obtain multiple segmented multivariate time series;

[0014] The plurality of segmented multivariate time series are grouped and processed according to the time slots they are in, to obtain a plurality of time sequence diagrams.

[0015] Optionally, in the method for analyzing electronic medical records of patients with myocardial infarction or heart failure, the spatiotemporal dynamic graph neural network model comprises a temporal dynamic graph isomorphic network module;

[0016] The step of acquiring a plurality of sparse adjacency matrices and node attribute information in the time series graphs, and performing spatial information learning processing on the sparse adjacency matrices and the node attribute information to obtain a plurality of multivariate time series graphs specifically includes:

[0017] Inputting the plurality of time sequence graphs into the time dynamic graph isomorphic network module, and obtaining the sparse adjacency matrix and node attribute information in the plurality of time sequence graphs through the time dynamic graph isomorphic network module;

[0018] The sparse adjacency matrix and node attribute information are subjected to aggregation processing, dynamic time merging processing, self-circular addition processing and feature connection processing to obtain a plurality of multivariate time series graphs.

[0019] Optionally, the electronic medical record analysis method for patients with myocardial infarction or heart failure, wherein the step of obtaining a sparse adjacency matrix in a plurality of the time series graphs specifically includes:

[0020] Acquire connection weight information between different nodes in the plurality of time sequence graphs, and construct a graph adjacency matrix according to the connection weight information;

[0021] The graph adjacency matrix is ​​initialized to obtain a sparse adjacency matrix.

[0022] Optionally, the electronic medical record analysis method for patients with myocardial infarction or heart failure, wherein the initialization processing of the graph adjacency matrix to obtain a sparse adjacency matrix specifically includes:

[0023] Obtaining a plurality of random sampling values ​​in the time sequence graphs, and performing filling processing on the graph adjacency matrix according to the random sampling values ​​to obtain a filled graph adjacency matrix;

[0024] The fill-in graph adjacency matrix is ​​subjected to sparse conversion processing to obtain a sparse adjacency matrix.

[0025] Optionally, in the method for analyzing electronic medical records of patients with myocardial infarction or heart failure, the spatiotemporal dynamic graph neural network model includes a graph pooling module;

[0026] The downsampling of the plurality of multivariate time series graphs to obtain a plurality of multivariate time series features specifically includes:

[0027] Inputting the plurality of multivariate time series graphs into the graph pooling module, and obtaining the plurality of node numbers in the plurality of multivariate time series graphs through the graph pooling module;

[0028] A convolution operation is used to downsample the plurality of node numbers to obtain a plurality of multivariate time series features.

[0029] Optionally, in the method for analyzing electronic medical records of patients with myocardial infarction or heart failure, the spatiotemporal dynamic graph neural network model comprises an average pooling layer and a fully connected layer;

[0030] The step of performing integrated learning processing on the spatiotemporal dynamic graph neural network model according to the plurality of multivariate time series features and outputting the electronic medical record analysis result corresponding to the electronic medical record of the patient specifically includes:

[0031] Inputting the plurality of multivariate time series features into the average pooling layer and the fully connected layer in sequence to obtain a plurality of multivariate time series results;

[0032] An integrated learning strategy is used to obtain data features corresponding to multiple time scales in multiple multivariate time series results, and an integrated learning process is performed on the spatiotemporal dynamic graph neural network model according to the data features corresponding to the multiple time scales;

[0033] The spatiotemporal dynamic graph neural network model after ensemble learning processing outputs the electronic medical record analysis results corresponding to the patient's electronic medical record.

[0034] In addition, to achieve the above-mentioned purpose, the present invention further provides an electronic medical record analysis system for patients with myocardial infarction or heart failure, wherein the electronic medical record analysis system for patients with myocardial infarction or heart failure comprises:

[0035] An electronic medical record input module, used to obtain the patient's electronic medical record, and input the patient's electronic medical record into the spatiotemporal dynamic graph neural network model to obtain multiple time sequence graphs;

[0036] A spatial information learning module, used for acquiring a plurality of sparse adjacency matrices and node attribute information in the time series graphs, performing spatial information learning processing on the sparse adjacency matrices and the node attribute information, and obtaining a plurality of multivariate time series graphs;

[0037] A downsampling processing module, used for performing downsampling processing on the plurality of multivariate time series graphs to obtain a plurality of multivariate time series features;

[0038] The medical record analysis result output module is used to perform integrated learning processing on the spatiotemporal dynamic graph neural network model according to the multiple multivariate time series features, and output the electronic medical record analysis results corresponding to the patient's electronic medical record.

[0039] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and an electronic medical record analysis program for patients with myocardial infarction or heart failure stored in the memory and run on the processor, and when the electronic medical record analysis program for patients with myocardial infarction or heart failure is executed by the processor, the steps of the electronic medical record analysis method for patients with myocardial infarction or heart failure are implemented as described above.

[0040] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an electronic medical record analysis program for patients with myocardial infarction or heart failure, and when the electronic medical record analysis program for patients with myocardial infarction or heart failure is executed by a processor, the steps of the electronic medical record analysis method for patients with myocardial infarction or heart failure are implemented as described above.

[0041] In the present invention, the patient's electronic medical record is obtained, and the patient's electronic medical record is input into a spatiotemporal dynamic graph neural network model to obtain multiple time series graphs; the sparse adjacency matrix and node attribute information in the multiple time series graphs are obtained, and the sparse adjacency matrix and the node attribute information are subjected to spatial information learning processing to obtain multiple multivariate time series graphs; the multiple multivariate time series graphs are subjected to downsampling processing to obtain multiple multivariate time series features; the spatiotemporal dynamic graph neural network model is subjected to integrated learning processing according to the multiple multivariate time series features, and the electronic medical record analysis results corresponding to the patient's electronic medical record are output. The present invention can dynamically extract the spatiotemporal information corresponding to the physiological parameters in the patient's electronic medical record by constructing a spatiotemporal dynamic graph neural network model, and can accurately capture the dynamic dependencies and interactions between the physiological parameters in the patient's electronic medical record, significantly improving the accuracy of the output of the electronic medical record analysis results, and providing more reliable support for the hospital's clinical decision-making for patients with myocardial infarction or heart failure. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of a preferred embodiment of the electronic medical record analysis method for patients with myocardial infarction or heart failure of the present invention;

[0043] Figure 2 It is a schematic diagram of the overall framework of STDGNN of a preferred embodiment of the electronic medical record analysis method for patients with myocardial infarction or heart failure of the present invention;

[0044] Figure 3 It is a schematic diagram of the TDGIN framework of a preferred embodiment of the electronic medical record analysis method for patients with myocardial infarction or heart failure of the present invention;

[0045] Figure 4 It is a schematic diagram of the TGPool framework of a preferred embodiment of the electronic medical record analysis method for patients with myocardial infarction or heart failure of the present invention;

[0046] Figure 5 It is a schematic diagram of a complete list of physiological variables used for experimental analysis in a preferred embodiment of the electronic medical record analysis method for patients with myocardial infarction or heart failure of the present invention;

[0047] Figure 6 It is a schematic diagram of the data processing process of a preferred embodiment of the electronic medical record analysis method for patients with myocardial infarction or heart failure of the present invention;

[0048] Figure 7 It is a schematic diagram of the baseline model of the preferred embodiment of the electronic medical record analysis method for patients with myocardial infarction or heart failure of the present invention and the AUROC score benchmark of STDGNN across data sets;

[0049] Figure 8 It is a structural diagram of a preferred embodiment of the electronic medical record analysis system for patients with myocardial infarction or heart failure of the present invention;

[0050] Fig. 9 It is a structural diagram of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0052] The present invention relates to the field of electronic medical record analysis of in-hospital patients with myocardial infarction or heart failure, and in particular to an application that uses a deep learning model to analyze patient medical records in a real clinical environment and generate analysis results for reference by doctors. At present, there are two major challenges in predicting the severity of in-hospital patients with myocardial infarction or heart failure. First, the sampling frequency of many physiological parameters is low, resulting in a limited number of parameters that can be used for early detection and analysis of the severity of myocardial infarction / heart failure patients. Therefore, the complexity of human dynamic systems and the lack of sufficient information or data make such multi-step advance prediction a difficult and challenging task. Secondly, the use of multivariate time series (MTS) in electronic medical records for prediction and analysis remains a complex and long-term area of ​​concern. The core feature of MTS is that each individual time series not only contains its own dynamic change pattern, but also interacts with other series and exhibits rich temporal correlations. For example, an increased heart rate may be associated with an increase in blood pressure, and this relationship may change over time and with the patient's condition. This inherent dynamism and interconnectedness makes the information contained in MTS data extremely rich, but also increases the difficulty of analysis. However, current work on predicting the severity of myocardial infarction or heart failure in patients rarely considers the dynamics of time series and the correlation between different variables. Therefore, in response to the above technical problems, the present invention proposes a new type of graph neural network - Spatio-Temporal Dynamic Graph Neural Network (STDGNN), which dynamically extracts the spatiotemporal information of physiological parameters from the electronic medical records of in-hospital patients to obtain the severity of myocardial infarction or heart failure of patients, which can help hospitals or doctors to conduct further analysis. Specifically, STDGNN constructs a graphical representation of multiple physiological parameters and uses temporal convolution and graph neural network frameworks to learn the characteristics of each variable. In addition, the present invention develops a dynamic graph mechanism based on MTS data, which integrates multiple time periods of different lengths and adaptively learns the dynamic evolution of the graph structure through gradient descent. This mechanism aims to enhance the model's ability to capture dynamic dependencies between adjacent variables, thereby improving the effectiveness of graph learning in the context of complex medical data.

[0053] One of the most widely used scoring systems for predicting the severity of in-hospital patients is the Acute Physiology and Chronic Health Evaluation II (APACHE II) algorithm, which uses 15 variables, including physiological measurements, to provide a snapshot of the patient's recent condition. With the popularity of electronic health records (EHRs) and the deployment of patient workflow management software, the calculation process of APACHE II has become automated. However, the main disadvantage of the algorithm is that it fails to utilize available historical patient information. Although subsequent versions have been optimized, improved their deficiencies and added details, they have also increased the number of variables required. Therefore, due to its high complexity, the APACHE II algorithm is still not widely used in actual clinical practice.

[0054] Another widely used scoring system is the Simplified AcutePhysiology Score II (SAPS-II). SAPS-II assesses the severity of a patient's illness using demographic information, vital signs, and laboratory test results, and simplifies the calculation process of APACHE II. Because the SAPS-II score has an S-shaped relationship with severity, it can be used to predict the severity of in-hospital patients. SAPS-III is a subsequent version of SAPS-II that uses 17 variables, including physiological and disease manifestation variables, to further improve prediction accuracy.

[0055] The Oxford Acute Severity of Illness Score (OASIS) is also a system used to assess the severity of illness in hospital patients, especially within the first 24 hours after admission. OASIS predicts in-hospital severity based on multiple clinical variables such as demographics, comorbidities, vital signs, and laboratory results. Another scoring system is the Sequential Organ Failure Assessment (SOFA), which is used to calculate the risk of organ failure, and although it is not designed to predict severity, it is positively correlated with in-hospital severity.

[0056] It is important to note that these critical care scoring systems are designed for a broad patient population and are not specifically targeted at patients with heart failure. Current scoring systems for assessing heart failure focus primarily on detecting the risk and severity of heart failure or cardiovascular disease; however, these scoring systems are not suitable for directly predicting the severity of myocardial infarction or heart failure.

[0057] In addition, the essence of traditional scoring systems is based on capturing the linear relationship between predictor variables and outcomes, while the relationship between physiological parameters and severity is usually very complex and may not be strictly linear. Capturing nonlinear relationships can significantly enhance the predictive ability of severity.

[0058] With the popularization of EHR and the continuous advancement of computing technology, machine learning (ML) and deep learning (DL) methods have gradually been applied to the prediction of the severity of heart failure patients. For example, a variety of machine learning models, including Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Light Gradient Boosting Machine (LGBM), are used to predict the severity of heart failure patients. Or use the Extreme Gradient Boosting (XGBoost) model to predict the severity of heart failure patients, and combine SHapley Additive exPlanations (SHAP) to explain the prediction results of the model. On the same dataset, the same model can also be used to further predict the stratified severity of heart failure patients, and external validation is added. The Long Short-Term Memory (LSTM) model can also be used to analyze the impact of various combinations of physiological parameters on the severity of heart failure, or an interpretable deep learning model based on Transformer can be used to model and predict the occurrence of heart failure.

[0059] Multivariate Time Series (MTS) models have made significant progress in recent years. Although some landmark time series models have not been specifically used to predict the severity of heart failure, they have shown great potential in this task. For example, the Recurrent Neural Network (RNN) can model and remember sequence data by introducing recurrent connections, thereby effectively capturing temporal dependencies and performing well when dealing with missing data in the sequence. The Residual Network (ResNet) solves the problems of gradient vanishing and model degradation in deep neural network training by stacking residual blocks. The improved ResNetPlus introduces attention mechanisms and cross-layer connections, further enhancing the model's attention to and integration of important features in the sequence.

[0060] Other representative time series models include InceptionTime and InceptionTimePlus, which enhance the model's ability in long-term dependency and local feature understanding by capturing multi-scale features. In addition, Multilayer Perceptron (MLP) and its improved Gated Multilayer Perceptron (GMLP) better capture the global dependency of the sequence through the self-attention mechanism. TimesNet combines Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN), and captures local features and global dependencies in time series through the temporal attention mechanism, further improving the prediction ability. Using the Transformer architecture and introducing spatial pyramid pooling, the model's ability to capture complex patterns is significantly enhanced, making it a prediction technology for time series.

[0061] In recent years, Graph Neural Network (GNN) has also developed rapidly in the application of multivariate time series (MTS). GNN processes graph information through a local aggregation mechanism, and the embedding vector of each node is calculated by recursively aggregating and transforming the information of its neighbors. Over the years, multiple variants of GNN have been proposed, such as Graph Convolutional Network (GCN), Graph Attention Network (GAT), and Graph Isomorphism Network (GIN). These variants perform well in processing dependencies in time series data. For example, a graph learning layer is used to capture the hidden spatial relationship between time series data. Further, by capturing the correlation between sequences and the temporal dependency in the spectral domain, the prediction accuracy of MTS is effectively improved.

[0062] Despite the excellent performance of these models, most graphical models still only construct a single static graph for the entire time series data, ignoring the dynamic process that changes over time. Especially in the in-hospital environment, the human physiological state is often in an unstable state, so capturing the dynamic dependencies between physiological variables is particularly important for accurate prediction.

[0063] Among the existing methods for predicting the severity of in-hospital patients, although some traditional scoring systems and models based on machine learning and deep learning can provide certain predictive capabilities, they have multiple limitations. First, traditional severity scoring systems, such as APACHE II, SAPS-II, and SOFA, mainly rely on static physiological variables and fail to effectively capture the complex and dynamic temporal correlations between variables. These systems assume a linear relationship between the predictor variables and the severity, while ignoring the possible nonlinear and dynamic change patterns between physiological parameters and severity. In addition, existing MTS-based methods often only deal with changes in a single time series, paying less attention to the interaction and time dependency between different variables. The core feature of MTS is that each time series not only contains its own dynamic changes, but also has a complex interconnected relationship with other series, which has not been fully explored and utilized in existing technical methods. Therefore, existing technologies still have great deficiencies in capturing the dynamic dependencies between physiological variables of in-hospital patients and the interactions between multiple variables.

[0064] In view of the shortcomings of the prior art, the present invention proposes a new spatio-temporal dynamic graph neural network (STDGNN), which aims to improve the performance of the prediction model of the severity of myocardial infarction or heart failure in terms of capturing time series dynamics and multivariate interactions. By dynamically extracting the spatio-temporal information of physiological parameters from the electronic medical records of in-hospital patients, the present invention constructs a graphical representation of multiple physiological variables, and uses the framework of temporal convolution and graph neural network to effectively capture the characteristic changes of each variable. The core is the design of a dynamic graph mechanism based on MTS, which can adaptively learn the graph structure that evolves over time, and capture the dynamic dependencies and interactions of physiological parameters in combination with multiple time scales. Through this method, the present invention can more accurately capture the complex relationship between the physiological parameters of patients, significantly improve the ability to predict the severity of myocardial infarction or heart failure in patients, and provide more reliable support for clinical decision-making in the hospital.

[0065] The electronic medical record analysis method for patients with myocardial infarction or heart failure described in the preferred embodiment of the present invention is as follows: Figure 1 As shown, the electronic medical record analysis method for patients with myocardial infarction or heart failure includes the following steps:

[0066] Step S10: Obtain the patient's electronic medical record, and input the patient's electronic medical record into the spatiotemporal dynamic graph neural network model to obtain multiple time sequence graphs.

[0067] Specifically, the patient's electronic medical record is obtained, the patient's electronic medical record is input into a spatiotemporal dynamic graph neural network model, and the multivariate time series in the patient's electronic medical record is extracted through the spatiotemporal dynamic graph neural network model; the multivariate time series is segmented multiple times to obtain multiple segmented multivariate time series; the multiple segmented multivariate time series are grouped according to the time slots in which they are located to obtain multiple time series diagrams.

[0068] like Figure 2 As shown, the present invention designs a spatiotemporal dynamic graph neural network (STDGNN). In order to capture the dynamic changes caused by the local time period, the input MTS (i.e., the multivariate time series in the electronic medical record of the patient of the present invention) is segmented multiple times (multiple segmentations generate multiple dynamic graphs, and fixed-length time segments are used for segmentation, such as the total length of the multivariate time series is 24, which can be divided into two segments 12, three segments 8, and four segments 6), and they are grouped into different time slots. For each group, the time series graph is constructed according to their respective time slots, and each variable represents a node in the time series graph.

[0069] Step S20, obtaining a plurality of sparse adjacency matrices and node attribute information in the time series graphs, performing spatial information learning processing on the sparse adjacency matrices and the node attribute information, and obtaining a plurality of multivariate time series graphs.

[0070] The present invention utilizes the characteristic information of graph nodes and a learnable parameter module to initialize a set of graph adjacency matrices (the graph adjacency matrix enables the time series graph to be calculated in the graph neural network, and all parameters in the graph adjacency matrix will be continuously updated as the model is trained, and a sparse adjacency matrix is ​​subsequently obtained by initializing the graph adjacency matrix) to express the initial relationship between different nodes. At the same time, the elements of the graph adjacency matrix will be iteratively updated through the graph learning module.

[0071] Among them, the present invention designs TDGIN (i.e., the time dynamic graph isomorphic network module in the present invention) and TGPool (i.e., the graph pooling module in the present invention) together to form a graph learning module. Each group of generated time sequence graphs is sent to TDGIN for learning, and TGPool is used for graph pooling. TDGIN can learn the dynamic associations between different nodes through the dynamic changes of the time sequence graph.

[0072] Specifically, a plurality of the timing graphs are input into the time dynamic graph isomorphic network module, and the connection weight information between different nodes in the plurality of the timing graphs is obtained through the time dynamic graph isomorphic network module, and a graph adjacency matrix is ​​constructed according to the connection weight information; random sampling values ​​in the plurality of the timing graphs are obtained, and the graph adjacency matrix is ​​filled according to the random sampling values ​​to obtain a filled graph adjacency matrix; and the filled graph adjacency matrix is ​​sparsely converted to obtain a sparse adjacency matrix.

[0073] Given the complexity of the human body system, the impact of various physiological parameters on the disease is deeply correlated and constantly changing, so it is particularly important to construct a suitable timing diagram to solve this problem. The present invention regards each physiological parameter as a node in the timing diagram, and when switching time slots, provides the model with information about changes in connection weights between different nodes through different timing diagrams. The present invention uses a graph adjacency matrix to represent the hidden dependencies between variables (that is, nodes in the timing diagram). For the initialization of the graph adjacency matrix A, a learnable weight matrix W is first set, and then the weight matrix W is filled with random sampling values ​​from a uniform distribution [-limit, limit]. Among them, limit is the random sampling value, N s is the source node, N t is the target node. Then, the diagonal elements of the graph adjacency matrix A are set to -∞ to eliminate self-loops (the present invention captures the association between variables (that is, nodes and nodes in the time series graph), and the nodes of the self-loop are connected to themselves, so the self-loop does not need to exist). Finally, by implementing Top-k sparsification, the most important k node connection relationships of each node are retained, and the remaining values ​​are set to zero. In this way, the graph adjacency matrix A can be converted into a sparse form to reduce computational costs. The formula is as follows: A sparse =Top-k(A, k), where A sparse is a sparse adjacency matrix, Top-k(·) is the sparsification process, A is the graph adjacency matrix, and k is the number of retained nodes.

[0074] Furthermore, node attribute information in multiple time series graphs is obtained, and the sparse adjacency matrix and node attribute information are aggregated, dynamically merged, self-circulated, and added, and feature connected to obtain multiple multivariate time series graphs.

[0075] For the TDGIN (i.e., the time dynamic graph isomorphic network module in the present invention) provided in the present invention: the spatial relationship of different physiological parameters changing over time can be explored through the interaction generated by the node features in the timing graph. GNN is powerful in processing graph structure data, focusing on information transmission and feature aggregation between nodes. Existing GNN methods have shown excellent learning capabilities in processing static graph data. However, for dynamic graphs composed of multiple timing graphs converted by MTS, existing methods are not suitable for learning dynamic changes in the time dimension in the timing graph. Therefore, in the present invention, an improved GIN method is proposed, called the time dynamic graph isomorphic network (TDGIN), such as Figure 3 As shown, its purpose is to extract spatiotemporal information in dynamic graphs.

[0076] Each time series graph contains the same nodes, so the feature vectors corresponding to the nodes constitute the feature vector set (where v is a graph node, is the eigenvector corresponding to the first graph node in the eigenvector set, is the number of nodes, is a positive integer). The feature vector set is composed of MTS sequence slices (i.e., cut by time periods, represents a slice, d is the length of the slice). The structural information of the time series graph is combined with the attribute information of the node to learn the spatial information in the graph, where the initialized sparse adjacency matrix A sparse Represents the structural information of the graph. The formula is as follows: Where N is the total number of nodes; in, is the final graph, i.e., the multivariate time series graph in the present invention, is the transposed matrix of the sparse adjacency matrix; where Yes A sparse A is the standardization of , i is the row vector of the sparse adjacency matrix, and j is the vertical vector of the sparse adjacency matrix. sparse is a learnable parameter that can be continuously updated during the training process. In the dynamic graph processing step, for each node, the feature vector set will be added to its previous timing diagram The purpose of this is to integrate the feature information of the previous time slot. Then, on this basis, self-loops are added to each graph to reduce the information loss of the node itself. represents dynamic time merging; represents self-loop addition, ∈ is a correction factor, which is used to reflect the effect of time or parameter changes on impact.

[0077] After that, a layer of MLP is designed to learn the feature changes of each node in the dynamic graph (obtained from the time series graph). MLP contains two temporal convolutional layers to capture the dynamic changes in the time dimension, and then connects the features of all time slots in the dynamic graph to obtain the output of TDGIN (i.e., MTS graph, multivariate time series graph). The core formula of TDGIN is as follows:

[0078] in, represents the GIN output of node v in layer k, s∈N* is the number of time slots, and s>1, represents the weight of the edge from node i to node j in the graph at different times.

[0079] Step S30: down-sample the plurality of multivariate time series graphs to obtain a plurality of multivariate time series features.

[0080] The present invention provides TGPool (i.e., the graph pooling module in the present invention). TGPool is a differentiable method with learnable dynamic change information in the time dimension. Integrating it into the layered pooling method can avoid the flat pooling problem. The results of each group are obtained after average pooling and the fully connected layer.

[0081] For the TGPool set in the present invention: Pooling is a common operation in deep learning and is necessary after the classifier generates new features. The essence of pooling is the downsampling of features, but the graph is limited to non-Euclidean data structures that are not easy to operate. General pooling methods, such as maximum pooling, mean pooling, and sum pooling, aggregate a vector or even a feature matrix into multiple features, which leads to potential flat pooling problems and thus loses a lot of information. In addition, the present invention constructs an MTS graph, which retains the dynamic information of the time series. However, there is currently no mainstream pooling method that can handle this problem. Therefore, the present invention proposes a novel time series graph pooling method, named TGPool, such as Figure 4 It incorporates temporal features into graph pooling, alleviates the problem of flat pooling through layered pooling, and retains the features of MTS as much as possible.

[0082] Specifically, the plurality of multivariate time series graphs are input into the graph pooling module, and the plurality of node numbers in the plurality of multivariate time series graphs are obtained through the graph pooling module; the plurality of node numbers are downsampled by using a convolution operation to obtain a plurality of multivariate time series features.

[0083] The present invention uses convolution to downsample the number of nodes, where the number of nodes is the channel of the convolution layer. In order not to destroy the receptive field window, the convolution kernel size of this convolution is set to the same as the temporal convolution kernel size in TDGIN. For example, the shape of the graph input to the graph pooling module is (b, c, n, f), where b is the batchsize, c is the number of graphs, n is the number of nodes, and f is the feature of the node. The multivariate time series graph is converted into the first and second dimensions to ensure that the convolution operation downsamples the nodes. in and n out Set them as the number of input and output channels of the convolution respectively. The specific formula is as follows: in, is the final feature, K means the convolution kernel, so K c is the convolution kernel of the corresponding graph, K f is the convolution kernel corresponding to the node feature, Indicates that the input tensor is at position (b, n in , c+i, f+j), Represents the element position of the convolution kernel (n out , n in, i, j).

[0084] Step S40: performing integrated learning processing on the spatiotemporal dynamic graph neural network model according to the multiple multivariate time series features, and outputting the electronic medical record analysis results corresponding to the patient's electronic medical record.

[0085] In order to synthesize the dynamic changes of different time series graphs, the present invention uses ensemble learning to integrate all group networks together, and jointly determine the final output of the model according to the convergence degree of different local dynamic information of the time series. In this way, the prediction accuracy and robustness of the model are improved, and overfitting is reduced.

[0086] Specifically, the multiple multivariate time series features are sequentially input into the average pooling layer and the fully connected layer to obtain multiple multivariate time series results; an integrated learning strategy is adopted to obtain data features corresponding to multiple time scales in the multiple multivariate time series results, and the spatiotemporal dynamic graph neural network model is subjected to integrated learning processing according to the data features corresponding to the multiple time scales; the electronic medical record analysis results corresponding to the patient's electronic medical records are output through the spatiotemporal dynamic graph neural network model after the integrated learning processing.

[0087] Among them, the multiple multivariate time series features in the present invention are subjected to an average pooling layer and a fully connected layer to obtain multiple multivariate time series results, and subsequently integrated learning is performed based on the multiple multivariate time series results.

[0088] For the integrated learning set in the present invention: integrated learning is a powerful tool. By adopting the strategy of integrated learning, it is possible to integrate the STDGNN model for different segmentation lengths of time series, and can also significantly improve the prediction performance and generalization ability of the model. Different time series segmentation lengths allow the model to capture data features at multiple time scales. Long time series can reveal trends and long-term dependencies, while short time series can capture more subtle patterns and short-term fluctuations. By integrating these models at different time scales, it is possible to integrate information at each scale and enhance the model's understanding of the complex dynamics of time series. In addition, integrated learning usually reduces the variance of the model, making the model more robust to noise and overfitting under a single parameter setting. Therefore, the integrated learning method set in the present invention can improve the accuracy and reliability of the model without sacrificing explanatory power, wherein the prediction and training errors of integrated learning are shown as follows:

[0089] is the prediction expression of ensemble learning, where represents the prediction output of ensemble learning. This value is usually the average of the prediction results of multiple models, which is used to enhance the accuracy and stability of model prediction. T represents the number of base models in the ensemble model, and f t(X) represents the prediction result of the t-th base model for input X. Each base model in the ensemble model gives a prediction value for input X. Indicates taking the average of the prediction results of all T base models to obtain the final ensemble prediction output is the training error expression of ensemble learning.

[0090] Regarding the novel deep learning model STDGNN set in the present invention: the model integrates a dynamic graph mechanism based on MTS data, which dynamically extracts spatiotemporal information from the electronic medical records of in-hospital patients to enhance the prediction of the severity of HF patients.

[0091] Regarding the design of basic blocks in the present invention: The present invention designs TDGIN and TGPool to establish a dynamic dependency model. TDGIN can effectively extract spatiotemporal features in dynamic graphs. The TGPool method integrates temporal features into the graph structure, significantly improving the representation capability of MTS.

[0092] Regarding the interpretability of deep learning in the present invention: The present invention reveals the complex and diverse dynamic relationships between physiological parameters through the dynamic graph of the STDGNN model. This analysis is crucial for understanding the patient's health status and predicting results, and can provide valuable insights for clinical decision-making. In summary, the spatiotemporal dynamic graph neural network model constructed in the present invention can extract spatiotemporal information from the electronic medical records of patients with myocardial infarction or heart failure, and can accurately obtain the mutual influence between parameters and parameters, variables and variables in the medical records, as well as the dynamic changes of parameters and variables, and then obtain the output electronic medical record analysis results, which can be used as a reference for hospitals or medical staff for the current status of patients with myocardial infarction or heart failure, and is conducive to further analysis of the condition of patients with myocardial infarction or heart failure by hospitals or medical staff.

[0093] Technical effects of the present invention:

[0094] 1. Dynamically capture multivariate time dependencies: STDGNN can adaptively learn the dynamic dependencies between different physiological parameters through dynamic graph mechanisms, especially the changes in time series. Compared with traditional static graphs or time series methods, it can more effectively capture the complex interactions of physiological variables over time and improve the accuracy of predictions.

[0095] 2. Multi-time scale modeling capability: STDGNN combines multiple time scales, allowing the model to handle both short-term and long-term dependencies in different time windows. This makes STDGNN more advantageous than traditional models that rely on a single time scale in capturing the complex and changeable physiological states of in-hospital patients.

[0096] 3. No need for feature extraction, reducing the impact of individual differences: STDGNN learns features directly from the original physiological data, without the need for manual feature extraction in advance. This not only reduces the complexity of data processing, but also effectively reduces the impact of individual differences among patients on prediction performance, ensuring the wide applicability of the model in different patients.

[0097] Experimental verification process:

[0098] In the present invention, two widely recognized and extensive public in-hospital databases: MIMIC-III and MIMIC-IV are used to validate the model of the present invention. These data sets provide a rich information base for analyzing the clinical characteristics and severity prediction of heart failure patients, including comprehensive analysis of dynamic physiological signals and multiple physiological variables. The MIMIC-III data set includes demographic data, bedside vital sign measurements, laboratory test results, and patient severity-related information, laying the foundation for early identification and risk assessment of heart failure patients. As a follow-up product, the MIMIC-IV data set expands the data scope, incorporating additional patient information and detailed medical process records as well as updated clinical information, making it more suitable for medical research related to heart failure. In order to enhance the universality and applicability of the model, commonly used physiological variables were extracted from the two data sets for model input. As Figure 5 As shown, the data types can be divided into static variables (SV), hourly chart variables (HCV) and waveform signals (WAVE). This experiment is divided into two parts: comparative experiment and AHE prediction experiment. In order to ensure the accuracy of the results, the present invention adopts five-fold cross validation to carry out all experiments.

[0099] For the whole process of sample selection and data processing, such as Figure 6As shown, when selecting samples from two large data sets, the first step of the present invention involves identifying patients with heart failure. Subsequently, patients under the age of 18 were excluded. In addition, patients who were transferred to the hospital or admitted to the hospital more than twice during a single hospitalization were also excluded. This exclusion criterion is based on the understanding that patients with an in-hospital stay of less than 24 hours usually receive postoperative monitoring, and their monitoring data may lead to misleading predictions of in-hospital severity. Therefore, considering the purpose of the study and the above reasons, it was decided to exclude patients with an in-hospital stay of less than 48 hours. The main purpose of the present invention is to explore the impact of various physiological parameters on the severity of heart failure patients. In terms of data processing, in order to ensure the universality of the model, a simplified method commonly used in processing in-hospital data sets was adopted. Specifically, SV and HCV data were extracted from the clinical database. It is worth noting that the HCV variable was originally recorded at irregular time points. In order to standardize these data, the present invention uses interpolation techniques to select the closest recorded values ​​around a specific time point to estimate the value of each HCV variable per hour. In addition, for WAVE data, the official hourly data version was used. Subsequently, the three types of data were merged, and the data of each patient in the first 48 hours of hospitalization were extracted for analysis. In addition, it is important to note that multiple admissions of the same patient were considered as separate samples. Figure 7 As shown, the subset extracted from the MIMIC-III dataset consists of 549 positive samples and 7091 negative samples. Meanwhile, the subset derived from the MIMIC-IV dataset includes 723 positive samples and 8078 negative samples.

[0100] The present invention divides the samples into training set, validation set and test set in a ratio of 8:1:1 to ensure that data from the same patient appears only in one of the groups. Regarding the model training parameters, the present invention sets the weight decay to 5, the momentum to 0.9, and the initial learning rate to 0.01. All experiments were performed using six NVIDIA RTX A6000 GPUs. In terms of CPU, the settings include Intel(R)Xeon(R)Gold 6240 CPU 2.60GHz. For optimization, the AdamW optimizer was used with an initial learning rate of 1e-3. Among them, the training epochs were set to 100 and the batch size was set to 64. In order to ensure the fairness of the experiment and the reliability of the design model, all baseline models involved in the experiment optimized their internal parameters through grid search to achieve the best effect. Specifically, the adjustment of the baseline model was carried out strictly in accordance with the official code. In addition, the STDGNN model of the present invention is uniquely designed with three time slots of different lengths: 6, 4, and 3 to more effectively capture the dynamic changes in the patient's status, and each model is subjected to 5-fold cross-validation to alleviate overfitting and optimize hyperparameters, thereby enhancing the robustness of the research results.

[0101] This paper compares the performance of various severity prediction models on MIMIC-III and MIMIC-IV datasets, including rule-based SAPS II, traditional machine learning models, and deep learning models. The purpose of this comparison is to evaluate their effectiveness in processing complex medical data, with a particular focus on the performance of the STDGNN model proposed in this paper.

[0102] The experimental results are as follows: On the MIMIC-III dataset, the linear regression model was observed to have the highest recall rate of 97.01%, surpassing other models in its ability to identify positive cases. In contrast, the XGBoost model leads the field with an accuracy rate of 53.69%, highlighting its efficiency in positive prediction accuracy. Most notably, the STDGNN model developed by the present invention performed well on key performance indicators, with an AUC of 83.93% and an Fbeta score of 64.14%, thereby demonstrating a significant advantage in comprehensive performance. A similar pattern can be seen in the analysis of the MIMIC-IV dataset. The linear regression model once again achieved the highest recall rate, this time at 97.91%, which is lower than its consistent ability to capture true positive results. The XGBoost model continued to maintain its accuracy lead, reaching 63.76%, once again demonstrating its accuracy strength. It is worth noting that the STDGNN model of the present invention maintains its exemplary performance advantage, achieving the highest AUC of 81.74% and the highest Fbeta score of 62.24%. These results demonstrate the robustness and excellent prediction accuracy of the STDGNN model on different medical datasets.

[0103] In addition to the technical content of the embodiments of the present invention, the present invention may also supplement the following contents: 1. The parameters and hyperparameters of each part of the model of the present invention may be changed; 2. The construction of the dynamic graph of the present invention may be designed according to specific tasks; 3. The design of the present invention may be further extended to other medical fields, such as predicting the progression of chronic diseases or the long-term recovery of patients through the STDGNN framework; 4. The present invention may analyze the weights of multiple variables for prediction results, and is therefore also suitable for such tasks.

[0104] Furthermore, if Figure 8 As shown, based on the above-mentioned electronic medical record analysis method for patients with myocardial infarction or heart failure, the present invention also provides an electronic medical record analysis system for patients with myocardial infarction or heart failure, wherein the electronic medical record analysis system for patients with myocardial infarction or heart failure comprises:

[0105] An electronic medical record input module 51 is used to obtain the patient's electronic medical record and input the patient's electronic medical record into the spatiotemporal dynamic graph neural network model to obtain a plurality of time sequence graphs;

[0106] A spatial information learning module 52 is used to obtain a sparse adjacency matrix and node attribute information in a plurality of the time series graphs, and perform spatial information learning processing on the sparse adjacency matrix and the node attribute information to obtain a plurality of multivariate time series graphs;

[0107] A downsampling processing module 53 is used to perform downsampling processing on the multiple multivariate time series graphs to obtain multiple multivariate time series features;

[0108] The medical record analysis result output module 54 is used to perform integrated learning processing on the spatiotemporal dynamic graph neural network model according to the multiple multivariate time series features, and output the electronic medical record analysis results corresponding to the patient's electronic medical record.

[0109] Furthermore, if Fig. 9 As shown, based on the above-mentioned electronic medical record analysis method and system for patients with myocardial infarction or heart failure, the present invention also provides a terminal accordingly, and the terminal includes a processor 10, a memory 20 and a display 30. Fig. 9 Only some components of the terminal are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0110] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Furthermore, the memory 20 may also include both an internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code of the installation terminal, etc. The memory 20 may also be used to temporarily store data that has been output or is to be output. In one embodiment, the memory 20 stores an electronic medical record analysis program 40 for patients with myocardial infarction or heart failure, and the electronic medical record analysis program 40 for patients with myocardial infarction or heart failure can be executed by the processor 10, thereby realizing the electronic medical record analysis method for patients with myocardial infarction or heart failure in this application.

[0111] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 20, such as executing the electronic medical record analysis method for patients with myocardial infarction or heart failure.

[0112] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0113] In one embodiment, when the processor 10 executes the electronic medical record analysis program 40 for patients with myocardial infarction or heart failure in the memory 20, the steps of the electronic medical record analysis method for patients with myocardial infarction or heart failure as described above are implemented.

[0114] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an electronic medical record analysis program for patients with myocardial infarction or heart failure, and when the electronic medical record analysis program for patients with myocardial infarction or heart failure is executed by a processor, the steps of the electronic medical record analysis method for patients with myocardial infarction or heart failure are implemented as described above.

[0115] In summary, the present invention provides an electronic medical record analysis method, system, terminal and computer-readable storage medium for patients with myocardial infarction or heart failure, the method comprising: obtaining the electronic medical record of the patient, and inputting the electronic medical record of the patient into a spatiotemporal dynamic graph neural network model to obtain multiple time series graphs; obtaining sparse adjacency matrices and node attribute information in multiple time series graphs, performing spatial information learning processing on the sparse adjacency matrices and the node attribute information to obtain multiple multivariate time series graphs; performing downsampling processing on multiple multivariate time series graphs to obtain multiple multivariate time series features; performing integrated learning processing on the spatiotemporal dynamic graph neural network model according to multiple multivariate time series features, and outputting the electronic medical record analysis results corresponding to the electronic medical record of the patient. The present invention can dynamically extract the spatiotemporal information corresponding to the physiological parameters in the electronic medical record of the patient by constructing a spatiotemporal dynamic graph neural network model, and can accurately capture the dynamic dependence and interaction between the physiological parameters in the electronic medical record of the patient, significantly improving the accuracy of the output of the electronic medical record analysis results, and providing more reliable support for the hospital's clinical decision-making for patients with myocardial infarction or heart failure.

[0116] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or terminal including the element.

[0117] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware (such as a processor, a controller, etc.) through a computer program, and the program can be stored in a computer-readable storage medium that can be read by a computer, and the program can include the processes of the above-mentioned method embodiments when executed. The computer-readable storage medium can be a memory, a disk, an optical disk, etc.

[0118] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for analyzing electronic medical records of patients with myocardial infarction or heart failure, characterized in that: The electronic medical record analysis method for patients with myocardial infarction or heart failure comprises: Obtaining electronic medical records of patients, and inputting the electronic medical records of patients into a spatiotemporal dynamic graph neural network model to obtain a plurality of time series graphs; Acquire a plurality of sparse adjacency matrices and node attribute information in the time series graphs, perform spatial information learning processing on the sparse adjacency matrices and the node attribute information, and obtain a plurality of multivariate time series graphs; Downsampling the plurality of multivariate time series graphs to obtain a plurality of multivariate time series features; The spatiotemporal dynamic graph neural network model is subjected to integrated learning processing according to the multiple multivariate time series features, and the electronic medical record analysis results corresponding to the patient's electronic medical record are output.

2. The electronic medical record analysis method for patients with myocardial infarction or heart failure according to claim 1, characterized in that: The patient's electronic medical record is obtained and input into the spatiotemporal dynamic graph neural network model to obtain multiple time series graphs, specifically including: Obtaining a patient's electronic medical record, inputting the patient's electronic medical record into a spatiotemporal dynamic graph neural network model, and extracting a multivariate time series in the patient's electronic medical record through the spatiotemporal dynamic graph neural network model; Performing multiple segmentation processing on the multivariate time series to obtain multiple segmented multivariate time series; The plurality of segmented multivariate time series are grouped and processed according to the time slots they are in, to obtain a plurality of time sequence diagrams.

3. The electronic medical record analysis method for patients with myocardial infarction or heart failure according to claim 1, characterized in that: The spatiotemporal dynamic graph neural network model includes a temporal dynamic graph isomorphic network module; The step of acquiring a plurality of sparse adjacency matrices and node attribute information in the time series graphs, and performing spatial information learning processing on the sparse adjacency matrices and the node attribute information to obtain a plurality of multivariate time series graphs specifically includes: Inputting the plurality of time sequence graphs into the time dynamic graph isomorphic network module, and obtaining the sparse adjacency matrix and node attribute information in the plurality of time sequence graphs through the time dynamic graph isomorphic network module; The sparse adjacency matrix and node attribute information are subjected to aggregation processing, dynamic time merging processing, self-circular addition processing and feature connection processing to obtain a plurality of multivariate time series graphs.

4. The electronic medical record analysis method for patients with myocardial infarction or heart failure according to claim 1 or 3, characterized in that: The acquiring of the sparse adjacency matrices in the plurality of time sequence graphs specifically comprises: Acquire connection weight information between different nodes in the plurality of time sequence graphs, and construct a graph adjacency matrix according to the connection weight information; The graph adjacency matrix is ​​initialized to obtain a sparse adjacency matrix.

5. The electronic medical record analysis method for patients with myocardial infarction or heart failure according to claim 4, characterized in that: The initializing process of the graph adjacency matrix to obtain a sparse adjacency matrix specifically includes: Obtaining a plurality of random sampling values ​​in the time sequence graphs, and performing filling processing on the graph adjacency matrix according to the random sampling values ​​to obtain a filled graph adjacency matrix; The fill-in graph adjacency matrix is ​​subjected to sparse conversion processing to obtain a sparse adjacency matrix.

6. The electronic medical record analysis method for patients with myocardial infarction or heart failure according to claim 1, characterized in that: The spatiotemporal dynamic graph neural network model includes a graph pooling module; The downsampling of the plurality of multivariate time series graphs to obtain a plurality of multivariate time series features specifically includes: Inputting the plurality of multivariate time series graphs into the graph pooling module, and obtaining the plurality of node numbers in the plurality of multivariate time series graphs through the graph pooling module; A convolution operation is used to downsample the plurality of node numbers to obtain a plurality of multivariate time series features.

7. The electronic medical record analysis method for patients with myocardial infarction or heart failure according to claim 1, characterized in that: The spatiotemporal dynamic graph neural network model includes an average pooling layer and a fully connected layer; The step of performing integrated learning processing on the spatiotemporal dynamic graph neural network model according to the plurality of multivariate time series features and outputting the electronic medical record analysis result corresponding to the electronic medical record of the patient specifically includes: Inputting the plurality of multivariate time series features into the average pooling layer and the fully connected layer in sequence to obtain a plurality of multivariate time series results; An integrated learning strategy is used to obtain data features corresponding to multiple time scales in multiple multivariate time series results, and an integrated learning process is performed on the spatiotemporal dynamic graph neural network model according to the data features corresponding to the multiple time scales; The spatiotemporal dynamic graph neural network model after ensemble learning processing outputs the electronic medical record analysis results corresponding to the patient's electronic medical record.

8. An electronic medical record analysis system for patients with myocardial infarction or heart failure, characterized in that: The electronic medical record analysis system for patients with myocardial infarction or heart failure comprises: An electronic medical record input module, used to obtain the patient's electronic medical record, and input the patient's electronic medical record into the spatiotemporal dynamic graph neural network model to obtain multiple time sequence graphs; A spatial information learning module, used for acquiring a plurality of sparse adjacency matrices and node attribute information in the time series graphs, performing spatial information learning processing on the sparse adjacency matrices and the node attribute information, and obtaining a plurality of multivariate time series graphs; A downsampling processing module, used for performing downsampling processing on the plurality of multivariate time series graphs to obtain a plurality of multivariate time series features; The medical record analysis result output module is used to perform integrated learning processing on the spatiotemporal dynamic graph neural network model according to the multiple multivariate time series features, and output the electronic medical record analysis results corresponding to the patient's electronic medical record.

9. A terminal, characterized in that: The terminal includes: a memory, a processor, and an electronic medical record analysis program for patients with myocardial infarction or heart failure stored in the memory and executable on the processor. When the electronic medical record analysis program for patients with myocardial infarction or heart failure is executed by the processor, the steps of the electronic medical record analysis method for patients with myocardial infarction or heart failure are implemented as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an electronic medical record analysis program for patients with myocardial infarction or heart failure, and when the electronic medical record analysis program for patients with myocardial infarction or heart failure is executed by a processor, the steps of the electronic medical record analysis method for patients with myocardial infarction or heart failure are implemented as described in any one of claims 1-7.

Images