A stroke recurrence prediction method based on graph neural network and missing feature prediction

Abstract

This application belongs to the field of medical data processing technology and discloses a method for predicting stroke recurrence based on graph neural networks and missing feature prediction. The method first preprocesses collected patient clinical data containing missing values, trains a prediction model using non-missing features to fill in missing feature values, and obtains complete feature data. Then, a feature graph is constructed based on the complete feature data, in which nodes represent features and edges represent correlations between features. Finally, the feature graph is input into a graph neural network model for training, learning high-order interactions between features, outputting stroke recurrence prediction results, and ranking feature importance. By effectively predicting and filling in missing data and using graph neural networks to mine complex topological relationships between features, the present invention improves the accuracy of stroke recurrence prediction and the robustness of the model while maintaining medical interpretability.

Description

Technical Field

[0001] The present invention relates to the field of medical data processing technology, and in particular to a disease prediction method based on artificial intelligence, specifically a stroke recurrence prediction method based on graph neural network and missing feature prediction. Background Art

[0002] Stroke is a common cerebrovascular disease, and the development of an early prediction system has significant cost-effectiveness in reducing the burden of stroke. Predicting stroke recurrence is an important research area in healthcare, aiming to reduce morbidity and mortality by identifying individuals at high risk of recurrence and implementing interventions.

[0003] Traditional recurrence prediction techniques rely primarily on clinical diagnosis, such as imaging studies (CT / MRI, carotid ultrasound, CTA) and molecular marker testing (C-reactive protein, homocysteine, and genetic testing). Physicians use these indicators to create risk scores. However, this manual diagnostic approach has limitations: First, it relies on expensive equipment and specialized physicians, making it difficult to popularize in primary care settings. Second, the criteria for assessing these characteristics vary significantly across medical institutions and physicians, leading to a lack of uniformity in risk stratification and treatment plans.

[0004] With the development of artificial intelligence technology, machine learning and deep learning have been widely used in the risk prediction of cerebrovascular diseases. Although research on stroke recurrence prediction based on traditional machine learning has made initial progress, existing methods still face bottlenecks:

[0005] 1. Shallow feature interaction problem: Traditional algorithms (such as gradient boosting trees) often stay at the shallow feature co-occurrence analysis level and fail to deeply model the nonlinear correlation and medical semantic coupling relationship between features.

[0006] 2. Static weight allocation defects: The fixed feature importance evaluation system is difficult to adapt to dynamic medical scenarios, and the traditional model cannot achieve dynamic adjustment of feature sensitivity.

[0007] 3. Model performance bottleneck: When processing the ultra-high-dimensional features of medical data, the computational complexity of traditional methods increases exponentially.

[0008] 4. Medical interpretability dilemma: Although some models provide a ranking of the importance of basic features, they lack a mapping explanation of pathophysiological mechanisms.

[0009] 5. Parameter optimization and generalization capability: Existing methods rely heavily on manual parameter tuning and are unstable when generalizing cross-institutional data.

[0010] In addition, there are often missing features in clinical data. How to effectively handle missing values ​​and make full use of all available information is also a challenge facing current research. Summary of the Invention

[0011] To solve the problems raised in the above background technology, the present invention aims to provide a stroke recurrence prediction method based on graph neural networks and missing feature prediction. This method can effectively process missing features in clinical data, use graph neural networks to mine complex topological relationships and high-order interactions between features, thereby improving the accuracy of stroke recurrence prediction and providing medically meaningful feature importance ranking.

[0012] In a first aspect, the present invention provides a method for predicting stroke recurrence based on graph neural networks and missing feature prediction, comprising the following steps:

[0013] Step S1: Acquire multimodal medical data including patient basic information, medical history characteristics, test indicators, and clinical score data, preprocess the multimodal medical data to generate a standardized feature vector; construct a missing feature prediction model, use the non-missing features in the standardized feature vector as input, predict and fill in the missing features to generate complete feature data;

[0014] Step S2: Based on the complete feature data, calculate the Pearson correlation coefficient between features, select the features with the highest preset proportion of correlation rankings, and construct a feature graph. In the graph, the nodes represent the features, and the edge weights are the correlation coefficients between the features.

[0015] Step S3: Input the feature graph into the graph neural network model, aggregate node features through the message passing layer to generate an embedding representation of each node; linearly map the node embeddings using a feedforward neural network, and output a binary classification result of the stroke recurrence probability;

[0016] Step S4: Based on the node weights in the graph neural network, calculate the contribution of each feature to the prediction result and generate an interpretable feature importance ranking.

[0017] As an optional implementation of the first aspect of the present application, the preprocessing in step S1 includes: normalizing continuous variables using the Z-score standardization formula; and performing one-hot encoding on discrete variables, wherein the discrete variables specifically include: bubbling test results graded according to the microembolic signal range, OSAHS severity graded according to the apnea-hypopnea index, TOAST stroke subtypes classified according to the pathogenesis of cerebral infarction, and SPI-II scores and ESSEN scores graded according to the total score threshold.

[0018] As an optional implementation of the first aspect of the present application, in step S1, the missing feature prediction model is constructed by selecting the model with the best prediction effect among fully connected neural network, XGBoost and random forest through cross-validation.

[0019] As an optional implementation of the first aspect of the present application, in step S1, when the model with the best prediction effect is a fully connected neural network, the fully connected neural network adopts a three-layer structure including a Dropout layer and a ReLU activation function, projects the input features to 32 dimensions, 64 dimensions and 2 dimensions in turn, and uses CrossEntropyLoss as the loss function for training.

[0020] As an optional implementation of the first aspect of the present application, the calculation process of the message delivery layer in step S3 is defined as: ,in It is a feature In the graph neural network l Hidden representation of layer nodes, is the feature in the adjacency matrix Features The correlation coefficient, represents the parameters that can be learned when mapping features to high-dimensional space, represents a nonlinear transformation, represents the output of the message passing layer, Represents the features in the adjacency matrix With its own characteristics The correlation coefficient, Representation characteristics In the graph neural network l Hidden representation of layer nodes.

[0021] As an optional implementation of the first aspect of the present application, in step S3, after the message passing layer processing of the graph neural network model, it also includes mapping the node embedding representation to a higher dimension through an adaptively weighted dual-path multi-layer perceptron, specifically including: inputting the node features into the dual channels and performing linear mapping respectively to obtain the first path features and the second path features, and weighting the first path features and the second path features through learnable weight parameters, splicing the weighted results and inputting them into a linear layer to generate high-dimensional features.

[0022] As an optional implementation of the first aspect of the present application, in step S3, after being processed by the adaptively weighted dual-path multilayer perceptron, it also includes using another multilayer perceptron without an activation function to remap the high-dimensional features back to a one-dimensional representation, and then inputting them into the feedforward neural network; the last layer of the feedforward neural network uses a Softmax function to output the binary classification result.

[0023] In a second aspect, an embodiment of the present application provides a stroke recurrence prediction system based on graph neural networks and missing feature prediction, comprising:

[0024] A multimodal data integration and preprocessing module is used to obtain multimodal medical data containing basic patient information, medical history characteristics, test indicators, and clinical score data, preprocess the multimodal medical data to generate a standardized feature vector; construct a missing feature prediction model, use the non-missing features in the standardized feature vector as input, predict and fill in the missing features to generate complete feature data;

[0025] The feature association graph construction module is used to calculate the Pearson correlation coefficient between features based on the complete feature data, and filter the features with a preset proportion of correlation rankings to construct a feature graph. The nodes in the graph represent features, and the edge weights are the correlation coefficients between features.

[0026] A graph neural network prediction module is configured to input the feature graph into a graph neural network model, aggregate node features through a message passing layer to generate an embedding representation of each node; linearly map the node embeddings using a feedforward neural network, and output a binary classification result of the stroke recurrence probability;

[0027] The explainability traceability module is used to calculate the contribution of each feature to the prediction result based on the node weights in the graph neural network and generate an explainable feature importance ranking.

[0028] In a third aspect, an embodiment of the present application provides an electronic device comprising a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein when the program or instruction is executed by the processor, the steps of the method described in the first aspect are implemented.

[0029] In a fourth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the method described in the first aspect are implemented.

[0030] Compared with the existing technology, the present invention proposes a stroke recurrence prediction method based on graph neural network and missing feature prediction, which has the following beneficial effects:

[0031] (1) By filling in missing features through a special missing feature prediction model, the existing data is effectively utilized, the data integrity is improved, the information loss is reduced, and a higher quality data foundation is provided for subsequent model training.

[0032] (2) The graph neural network model can explicitly model and learn the complex topological relationships and high-order interactions between clinical features, breaking through the limitations of traditional models in feature relationship mining, thereby improving the accuracy of stroke recurrence risk assessment.

[0033] (3) This method can output the ranking of the contribution of features to risk assessment, which enhances the interpretability of the model, helps clinicians understand the basis of risk assessment results, and provides a reference for formulating targeted intervention measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flowchart of a method for predicting stroke recurrence based on graph neural network and missing feature prediction according to an embodiment of the present invention;

[0035] Figure 2 2 is a diagram illustrating data processing of a missing feature prediction model according to an embodiment of the present invention;

[0036] Figure 3 This is a diagram illustrating data processing of a graph neural network model according to an embodiment of the present invention;

[0037] Figure 4 This is a diagram showing the ranking of feature contributions in an embodiment of the present invention;

[0038] Figure 5 This is a structural diagram of a stroke recurrence prediction system based on graph neural network and missing feature prediction provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0040] The terms "first", "second", etc. in the specification and claims of this application are used to distinguish similar objects, and are not used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the application can be implemented in a sequence other than those illustrated or described here. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally represents that the objects associated before and after are in a kind of "or" relationship. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically limited.

[0041] Example 1

[0042] See also Figure 1 , is a flowchart of a stroke recurrence prediction method based on graph neural network and missing feature prediction provided by an embodiment of the present invention. The method may include the following steps:

[0043] Step S1: Acquire multimodal medical data containing patient basic information, medical history characteristics, test indicators and clinical score data, preprocess the multimodal medical data to generate a standardized feature vector; construct a missing feature prediction model, use the non-missing features in the standardized feature vector as input, predict and fill in the missing features to generate complete feature data.

[0044] For example, data from inpatients in the neurology department of a hospital were collected and desensitized. After review by the hospital's biomedical research ethics committee, the remaining 928 patients with ischemic stroke were included in this study. In terms of patient characteristics, basic information was collected, including age, gender, occupation, place of residence, smoking status, alcohol consumption, BMI, family history of stroke, transient ischemic attack (TIA), previous stroke, hypertension, diabetes, heart disease (atrial fibrillation, myocardial infarction, heart failure, coronary artery disease, and other heart diseases), peripheral vascular disease, cirrhosis, tumors, blood diseases, and hyperlipidemia. In addition, the system also includes relevant data from patient admission tests, including systolic and diastolic blood pressure, white blood cell count, red blood cell count, hemoglobin, platelet count, total cholesterol (CHOL), triglycerides (TG), high-density lipoprotein (HDL), low-density lipoprotein (LDL), apolipoprotein A (APO.A), apolipoprotein B (APO.B), blood glucose, glycosylated hemoglobin, and homocysteine. It also includes data on vascular stenosis, vascular plaques, patent foramen ovale (PFO), and sleep apnea-hypopnea syndrome (OSAHS). It also includes clinical data such as foaming test results, TOAST etiology classification, ESSEN score, and SPI-II score. Next, this data is preprocessed and converted into features. For continuous variables, they are normalized; for discrete variables, one-hot encoding is used to convert them into feature vectors. The five features of the bubble test, OSAHS severity, TOAST stroke subtype, SPI-II score, and ESSEN score were processed as follows: Bubble test result grading: Based on the classification criteria, the range of detected microembolic signals (Grade I: 1-10, Grade II: 11-25, Grade III: >25, no rain curtain formation, Grade IV: >25, rain curtain formation) was used to represent this variable using one-hot encoding, resulting in four distinct categories: Grade I, Grade II, Grade III, and Grade IV. Given that severity is primarily determined by the apnea-hypopnea index (mild: 5-15, moderate: 5-30, severe: >30), one-hot encoding was applied to this variable to create three independent categories: mild, moderate, and severe. TOAST stroke subtype: Because subtypes are categorized based on the pathogenesis of cerebral infarction, this variable was coded into five discrete categories: large artery atherosclerosis (LAA), cardioembolism (CE), small artery occlusion (SAA), other etiology determined (SOE), and stroke of unknown etiology (SUE). SPI-II score and ESSEN score: These risk levels are categorized based on the total score threshold for assessing stroke recurrence risk (SPI-II: low 2, intermediate 3-6, high -6; ESSEN: low <3, intermediate 4-7, high >7). Therefore, one-hot encoding was implemented for these two features to create three separate categories: low risk, intermediate risk, and high risk.

[0045] Furthermore, the continuous data in the samples was normalized, and the discrete data was one-hot encoded. The non-missing values ​​in the samples were then used as training data. The samples with missing features served as the test set, while those without missing features served as the training labels. These samples were trained using a neural network, XGBoost, and random forest algorithms, respectively. The ones with the best prediction performance were retained. The neural network used a three-layer fully connected neural network, projecting features into 32, 64, and 2 dimensions before generating the final predictions. The CrossEntropyLoss loss function was used.

[0046] The normalization formula is as follows:

[0047] ,

[0048] in is the original value of the feature, is the characteristic mean, The standard deviation of the feature is obtained, and the standardized data is obtained .

[0049] The rule of one-hot encoding is: for a discrete variable with K mutually exclusive categories, i The category is represented as the i The encoding format in which the first bit is 1 and the rest of the bits are 0. The one-hot encoding rules are as follows:

[0050] ,

[0051] Discrete data is one-hot encoded, where Indicates the i One-hot encoding of the class positions.

[0052] The process of passing input through a neural network to get a prediction is as follows:

[0053]

[0054] Among them, X represents the input features, represents the trainable parameters of the neural network, Represents the final predicted feature value, It consists of multiple linear layers and includes Dropout layers and ReLU activation layers.

[0055] In this step, to utilize features with missing values ​​for stroke recurrence prediction, the present invention first fills in missing data, using features with missing values ​​as labels and features without missing values ​​as predictive features. Because neural networks can process both continuous and discrete data, this method trains a missing feature prediction model to predict missing values.

[0056] Step S2: Based on the complete feature data, calculate the Pearson correlation coefficient between features, select the features with a preset proportion of correlation rankings, and construct a feature graph. The nodes in the graph represent the features, and the edge weights are the correlation coefficients between features.

[0057] Step S21: In order to eliminate the influence of redundant features on prediction, the present invention uses the Pearson correlation coefficient Feature selection was performed, and the specific calculations are as follows:

[0058] ,

[0059] in and is a feature and its corresponding label, and are the corresponding means, and n is the number of samples.

[0060] Finally, this real-time method retains the features with the top 70% correlation values ​​for the final prediction.

[0061] Step S22: The data features of the sample obtained in step S1 are treated as nodes, and then these nodes are connected to generate an adjacency matrix. The weights of the edges between nodes are the simple correlation coefficients of the features. The adjacency graph is used as data input for training.

[0062] Step S23: Calculate the correlation matrix between features using the Pearson correlation coefficient The specific process is as follows:

[0063] ,

[0064] in and are two features in the sample, and is the mean value corresponding to the feature, and n is the number of samples. The adjacency matrix A obtained by adding the correlation coefficients between each feature is as shown below:

[0065]

[0066] in is the set weight.

[0067] In this step, each processed data instance is represented as a graph by embedding its feature values ​​into a higher dimension. Edges exist between two features, and the edge value is the correlation coefficient between the features, calculated using the Pearson correlation coefficient. Multiple iterations of message passing are then applied. Finally, the learned node representations are projected into a single value, and the entire graph representation is obtained by concatenating the projected values.

[0068] Step S3: Input the feature graph into the graph neural network model, aggregate the node features through the message passing layer to generate an embedding representation of each node; use the feedforward neural network to perform linear mapping on the node embeddings, and output a binary classification result of the stroke recurrence probability.

[0069] Step S31: Message passing in the graph neural network. Each node transmits a message to its neighboring nodes, aggregates the transmitted information from its neighbors, and uses the aggregated information to update the node representation, as shown in the following formula:

[0070] ,

[0071] in It is a feature In the graph neural network l Hidden representation of layer nodes, In the adjacency matrix, the feature Features The correlation coefficient of . represents the parameters that can be learned when mapping features to high-dimensional space, Represents a nonlinear transformation. represents the output of the message passing layer, Represents the features in the adjacency matrix With its own characteristics The correlation coefficient, Representation characteristics In the graph neural network l Hidden representation of layer nodes.

[0072] After the graph neural network performs feature interaction, the present invention also uses an adaptively weighted dual-path MLP to map the features to a higher dimension to achieve feature interaction in the latent space. The feature change dimension is R∈{B,N,1} ->R∈{B,N,D}, where R represents the feature space, B represents the batch size, N represents the number of features, and D represents the feature dimension. The specific process is shown in the following formula:

[0073] ,

[0074] in, and Represents the learnable weights. and It represents the dual path feature. and It represents the result of processing the dual-path features. Represents a linear layer, Concat represents a concatenation operation, gAfter using the graph neural network to map the features to high dimensions, the present invention uses an MLP to remap the high-dimensional features back to one dimension:

[0075] ,

[0076] At this time, the dimension of the feature changes to R∈{B,N,D} ->R∈{B,N,1} ->R∈{B,N}.

[0077] Step S32: Feature Importance Ranking: By mapping each feature, converting it into a separate value, and finally using an FNN to predict its category, the contribution of each node to the prediction result can be obtained and the feature importance can be ranked according to the contribution.

[0078] ,

[0079] in is assigned to The weights of the nodes represented by , and the Softmax function is used for the final classification.

[0080] In this step, after linearly projecting the features into a high-dimensional space, several message passing layers are added to the model, each equipped with a ReLU activation function. Multiple learnable weights are also applied to the node representations, and three batch normalization layers are added. Concatenation is performed via four skip connections to form a GNN layer. Following the GNN layer, a feedforward neural network (FNN) is used to map the multidimensional representation of each node to a single value. FNNs do not use activation functions internally to maintain the linear nature of the mapping. However, a Softmax function is applied after the final layer to perform binary classification predictions, ensuring that the output value for each class is between 0 and 1. After the FNN processing, the final node values ​​are concatenated, and the output weights are used to complete the prediction.

[0081] Step S4: Based on the node weights in the graph neural network, calculate the contribution of each feature to the prediction result and generate an interpretable feature importance ranking.

[0082] The present invention compares the objective evaluation indicators and uses AUC (area under the ROC curve) and F1 evaluation indicators to conduct comparative experiments. The experimental results are shown in Table 1. The calculation formula of F1 is as follows:

[0083] ,

[0084] Among them, TP (1 Positive): the number of predicted positive examples that are actually positive, FP (0 ​​Positive): the number of predicted positive examples but actually negative, FN (0 Negative): the number of predicted negative examples but actually positive.

[0085] In this step, the graph neural network model (GNN) captures the potential relationship between features, maps the features to a high-dimensional space, and builds a graph to allow message transmission between features. It then remaps the features to a low-dimensional space and predicts them through the FNN to obtain the probability of each category. The weight of each feature for the specified category can be obtained, and the feature importance is ranked according to the weight of each feature.

[0086] In summary, if Figure 2 As shown, for a series of samples in the original data X { , ,..., }, the present invention first attempts to use the missing feature prediction model to fill in the missing feature value. Next, if Figure 3 As shown, for the filled data X´ , the present invention builds a map for it, and then uses the graph neural network model to take the feature set { , , , ,..., } Build a map for the node and process it, so that information can be transferred between related features to achieve interaction between features. In the dual-channel weighting module of the present invention, for feature nodes, dual-channel mapping is used to obtain feature and , and then through the learnable parameters and Weight it to get features and , then through the Concat splicing operation and a linear mapping layer Get the final features , and finally, each feature is mapped to the probability of stroke recurrence through a layer of neural network.

[0087] According to the method provided by the present invention, the missing feature prediction model was trained on a data set of 928 patients collected in cooperation with a hospital. The data set contains a total of 47 basic patient information: age, gender, past history: previous stroke, hypertension, diabetes, etc., of which 22 are missing features. After one-hot encoding of discrete variables, the total number of features obtained is 69. The present invention uses 26 features without missing values ​​to train the graph neural network model and predicts the 21 missing features. All experiments were carried out under five-fold cross-validation conditions, and after dividing the training and testing, the training data was used to train the neural network to fill the model to avoid data leakage.

[0088] As shown in Table 1, the proposed method based on graph neural networks and missing feature prediction outperforms all other models and methods, demonstrating significant progress compared to traditional machine learning algorithms. Compared to the KNN algorithm, this method improves the AUC by 17.7% and the F1 by 16.9%. Compared to the random forest model, this method improves the F1 by 17.7% and the AUC by 12.3%. Compared to the graph convolutional network (GCN), this method improves the AUC and F1 by 17.3% and 15.3%, respectively. Even compared to neural networks, this method still improves the AUC by 9.1% and the F1 by 10.6%.

[0089] Table 1 Comparative experimental results of this method and existing methods

[0090]

[0091] like Figure 4 The figure shows the contribution ranking of features. Feature contribution ranking is based on the absolute value of the feature scores calculated during training. Positive values ​​indicate a positive impact on the prediction results, while negative values ​​indicate a negative impact. As can be seen, OSAHS_Moderate (moderate OSAHS) has the greatest positive impact on classification performance.

[0092] Example 2

[0093] See also Figure 5 , shown is a schematic diagram of the structure of a stroke recurrence prediction system based on graph neural network and missing feature prediction proposed in the second embodiment of the present application. The system includes the following key modules:

[0094] The multimodal data integration and preprocessing module 100 is used to obtain multimodal medical data including basic patient information, medical history characteristics, test indicators, and clinical score data, preprocess the multimodal medical data to generate a standardized feature vector, and construct a missing feature prediction model, using the non-missing features in the standardized feature vector as input to predict and fill in the missing features to generate complete feature data;

[0095] The feature association graph construction module 200 is used to calculate the Pearson correlation coefficient between features based on the complete feature data, and filter the features with a preset proportion of correlation rankings to construct a feature graph, in which the nodes represent the features and the edge weights are the correlation coefficients between the features;

[0096] A graph neural network prediction module 300 is configured to input the feature graph into a graph neural network model, aggregate node features through a message passing layer to generate an embedding representation of each node; linearly map the node embeddings using a feedforward neural network, and output a binary classification result of the stroke recurrence probability;

[0097] The explainability traceability module 400 is used to calculate the contribution of each feature to the prediction result based on the node weights in the graph neural network and generate an explainable feature importance ranking.

[0098] In the embodiments of the present application, a stroke recurrence prediction system based on graph neural networks and missing feature prediction can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, the mobile electronic device can be a mobile phone, tablet computer, laptop computer, PDA, in-vehicle electronic device, wearable device, ultra-mobile personal computer (UMPC), netbook or personal digital assistant (PDA), etc. The non-mobile electronic device can be a server, network attached storage (NAS), personal computer (PC), etc., which is not specifically limited in the embodiments of the present application.

[0099] In the embodiments of the present application, a stroke recurrence prediction system based on graph neural networks and missing feature prediction can be a device having an operating system. The operating system can be an Android operating system, an iOS operating system, or other possible operating systems, which are not specifically limited in the embodiments of the present application.

[0100] The stroke recurrence prediction system based on graph neural network and missing feature prediction provided in the embodiment of the present application can achieve Figure 1 In the method embodiment, each process of implementing a stroke recurrence prediction method based on graph neural network and missing feature prediction is not described here to avoid repetition.

[0101] Optionally, an embodiment of the present application also provides an electronic device, including a processor, a memory, and a program or instruction stored in the memory and executable on the processor. When the program or instruction is executed by the processor, each process of the above-mentioned embodiment of a stroke recurrence prediction method based on graph neural network and missing feature prediction is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.

[0102] An embodiment of the present application also provides a readable storage medium having a program or instruction stored thereon. When the program or instruction is executed by a processor, the various processes of the embodiment of the above-mentioned method for predicting stroke recurrence based on graph neural network and missing feature prediction are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0103] The processor is the processor in the electronic device described in the above embodiment. The readable storage medium includes a computer-readable storage medium, such as a computer read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk.

[0104] It should be noted that, in this article, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising 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 device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element. In addition, it should be noted that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the opposite order according to the functions involved. For example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0105] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a more preferred embodiment. Based on this understanding, the technical solution of this application, or the part that contributes to the existing technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of this application.

[0106] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of this application, ordinary technicians in this field can also make many forms without departing from the purpose of this application and the scope of protection of the claims, all of which are within the protection of this application.

Claims

1. A stroke recurrence prediction method based on graph neural network and missing feature prediction, characterized in that: The following steps are involved: Step S1: Acquire multimodal medical data including patient basic information, medical history characteristics, test indicators, and clinical score data, preprocess the multimodal medical data to generate a standardized feature vector; construct a missing feature prediction model, use the non-missing features in the standardized feature vector as input, predict and fill in the missing features to generate complete feature data; Step S2: Based on the complete feature data, calculate the Pearson correlation coefficient between features, select the features with the highest preset proportion of correlation rankings, and construct a feature graph. In the graph, the nodes represent the features, and the edge weights are the correlation coefficients between the features. Step S3: Input the feature graph into the graph neural network model, aggregate node features through the message passing layer to generate an embedding representation of each node; linearly map the node embeddings using a feedforward neural network, and output a binary classification result of the stroke recurrence probability; Step S4: Based on the node weights in the graph neural network, calculate the contribution of each feature to the prediction result and generate an interpretable feature importance ranking.

2. The method according to claim 1, characterized in that The pre-processing in step S1 includes: Continuous variables were normalized using the Z-score standardization formula; And perform one-hot encoding on discrete variables, the discrete variables specifically include: The results of the bubble test were graded according to the range of microembolic signals, the severity of OSAHS was graded according to the apnea-hypopnea index, the TOAST stroke subtypes were classified according to the pathogenesis of cerebral infarction, and the SPI-II score and ESSEN score were graded according to the total score threshold.

3. The method according to claim 1, characterized in that In step S1, the missing feature prediction model is constructed by selecting the model with the best prediction effect among fully connected neural network, XGBoost and random forest through cross validation.

4. The method according to claim 3, characterized in that In step S1, when the model with the best prediction effect is a fully connected neural network, the fully connected neural network adopts a three-layer structure including a Dropout layer and a ReLU activation function, projects the input features to 32 dimensions, 64 dimensions and 2 dimensions in turn, and uses CrossEntropyLoss as the loss function for training.

5. The method according to claim 1, characterized in that The calculation process of the message passing layer in step S3 is defined as: ， in It is a feature In the graph neural network l Hidden representation of layer nodes, is the feature in the adjacency matrix Features The correlation coefficient, represents the parameters that can be learned when mapping features to high-dimensional space, represents a nonlinear transformation, represents the output of the message passing layer, Represents the features in the adjacency matrix With its own characteristics The correlation coefficient, Representation characteristics In the graph neural network l Hidden representation of layer nodes.

6. The method according to claim 5, characterized in that In step S3, after the message passing layer processing of the graph neural network model, the node embedding representation is mapped to a higher dimension through an adaptively weighted dual-path multi-layer perceptron, specifically including: The node features are input into the dual channels and linearly mapped to obtain the first path features and the second path features, and the first path features and the second path features are weighted by learnable weight parameters. The weighted results are spliced ​​and input into a linear layer to generate high-dimensional features.

7. The method according to claim 6, characterized in that In the step S3, after being processed by the adaptively weighted dual-path multilayer perceptron, the high-dimensional features are remapped back to a one-dimensional representation using another multilayer perceptron without an activation function, and then input into the feedforward neural network; the last layer of the feedforward neural network uses a Softmax function to output the binary classification result.

8. A stroke recurrence prediction system based on graph neural network and missing feature prediction, characterized in that: include: A multimodal data integration and preprocessing module is used to obtain multimodal medical data containing basic patient information, medical history characteristics, test indicators, and clinical score data, preprocess the multimodal medical data to generate a standardized feature vector; construct a missing feature prediction model, use the non-missing features in the standardized feature vector as input, predict and fill in the missing features to generate complete feature data; The feature association graph construction module is used to calculate the Pearson correlation coefficient between features based on the complete feature data, and filter the features with a preset proportion of correlation rankings to construct a feature graph. The nodes in the graph represent features, and the edge weights are the correlation coefficients between features. A graph neural network prediction module is configured to input the feature graph into a graph neural network model, aggregate node features through a message passing layer to generate an embedding representation of each node; linearly map the node embeddings using a feedforward neural network, and output a binary classification result of the stroke recurrence probability; The explainability traceability module is used to calculate the contribution of each feature to the prediction result based on the node weights in the graph neural network and generate an explainable feature importance ranking.

9. An electronic device, characterized in that: The invention comprises a processor, a memory, and a program or instruction stored in the memory and executable on the processor. When the program or instruction is executed by the processor, the steps of a method for predicting stroke recurrence based on a graph neural network and missing feature prediction as described in any one of claims 1 to 7 are implemented.

10. A readable storage medium, characterized in that: The readable storage medium stores a program or instruction, and when the program or instruction is executed by the processor, the steps of the stroke recurrence prediction method based on graph neural network and missing feature prediction as described in any one of claims 1 to 7 are implemented.

Images