Predicting Complications And Mortality In Myocardial Infarction Patients Using A Graph Neural Networks Model

Abstract Myocardial infarction results from sudden coronary artery blockage, leading to myocardial necrosis and severe complications. Early prediction is critical to mitigate these life-threatening outcomes and improve patient survival, but it remains challenging due to data heterogeneity and limited temporal feature utilization. To address this crucial gap, we proposed a GraphTransformer-based model with three key innovations: First, a density-adaptive optimized K-Nearest Neighbor graph to model patient similarities dynamically. Second, a short-term Convolutional Neural Network for fine-grained temporal feature extraction with dynamic gating. Third, a long-term Gated Recurrent Unit with Mamba-inspired dynamic gating to capture robust long-term dependencies. Experiment on the Myocardial Infarction Complications dataset, our model achieved an average AUC of 0.7236, with a standout 0.8777 for mortality prediction. Moreover, SHAP analysis revealed critical features like serum sodium levels and temporal dynamics, and enhanced clinical interpretability. Leveraging automated feature extraction from admission data, 72-hour temporal sequences, and derived temporal features with missing values, our approach empowers a robust DL model for multi-outcome prediction. Our method enables timely interventions, significantly reducing mortality and rehospitalization rates while equipping clinicians with a vital tool to boost patient survival, optimize critical care resources, and support tailored treatment strategies for MI patients' dynamic needs.