Graph-Based Temporal Attention for Coronary Artery Disease Prediction Using Electronic Health Records

Deep learning has significantly enhanced the ability to predict health outcomes using Electronic Health Records (EHR). However, the challenges posed by EHR data, such as sparsity, heterogeneity, and temporal complexity, necessitate innovative representation techniques. This research introduces a novel graph-based approach to effectively represent EHR data for the early prediction of coronary artery disease (CAD). We proposed a dual-graph approach by constructing then fusing two distinct graphs: a diagnosis graph and a lab graph. The diagnosis graph employs diagnosis nodes and visit nodes to capture patient diagnoses over time, while the lab graph discretizes continuous laboratory measurement results into four quantile levels, represented as lab nodes connected to visit nodes. Graph convolutional networks (GCN) are used separately on each graph to manage data sparsity, and a time-aware transformer layer captures the temporal dynamics of patient visits. The resulting visit embeddings from both graphs are concatenated and processed through fully connected layers for dimensionality reduction and final classification. Our model was evaluated on a dataset of 6,272 CAD patients and 47,295 non-CAD patients, using three different observation periods. Ablation studies were also performed to assess the impact of using only diagnosis data or only laboratory measurement data. The results demonstrate that our dual-graph approach effectively integrates diagnosis and laboratory measurement data, capturing both spatial and temporal information to accurately predict CAD, achieving a best AUROC (area under the receiver operation characteristic) of 0.908 ± 0.008. This method simplifies the graph structure, enhances interpretability, and improves predictive performance compared to previous techniques.