Attention-Driven Graph Neural Network Framework for Cardiovascular Risk Prediction Using Electronic Health Records

Abstract

The cardiovascular diseases (CVDs) can be listed among the significant causes of mortality in the world and this is why there is a need of a reliable and early answer to the risk. Since Electronic Health Records (EHRs) are integrated in numerous hospitals, the amount of structured patient data to analyze and predict is enormous. However, the historical machine learning techniques typically assume the existence of individual samples of patient records and ignore the already existing interrelations among individuals within the same clinical phenotype. This is a weakness because it narrows their ability to adopt complex relationships, which are present in real medical records. In this paper, the proposal of a model based on Graph Neural Network (GNN) and predicting cardiovascular risk alongside the structured EHR data is provided. The proposed model is a patient is represented as a node of a graph and the graph is formed based on similarity data of both clinical and demographic variables that comprise age, blood pressure, cholesterol levels and heart rate. Such representation using graphs helps the model to learn the characteristics of the patients, and the relation that may be possible between the patients. The architecture uses two modern GNNs, i.e. Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) to train meaningful representations of the constructed graph of patient similarities. The publicly available datasets, including UCI Heart Disease, are tested and trained on the models. Evaluation of the performance utilizes the assistance of such common measurement criteria as accuracy, precision, recall, F1-score, and ROC-AUC. The experimental findings show that the suggested GNN-based method provides better predictive results than the conventional machine learning models due to the effective use of the relational information in the data. The created system will be able to help medical staff detect high-risk people early enough and facilitate informed clinical judgment. This work indicates how graph-based deep learning methods may be effective to increase predictive analytics and emerging individualized healthcare solutions.