CT-HEG Graph Model Predicts ICU Mortality with High Accuracy

Predicting in-hospital mortality in Intensive Care Units (ICUs) accurately is a complex challenge due to the irregular nature of clinical observations and the diverse types of entities involved. Existing sequence models can handle irregular sampling but often overlook the structured relationships between data points, while traditional graph models typically assume fixed-interval inputs, which is not representative of real-world EHR data. Researchers propose the Continuous-Time Heterogeneous EHR Graph (CT-HEG) schema to address these limitations. CT-HEG represents each ICU stay as a typed, timestamped graph, featuring three node types (visit, vital, lab_event) and 2D edge attributes that encode both timing and value without requiring data imputation. They instantiated this schema with CHIRP-Net, a four-layer heterogeneous GATv2Conv network, and evaluated its performance on the MIMIC-IV v3.1 dataset. CHIRP-Net achieved a mean AUROC of 0.8449, outperforming several baseline models. A key finding from the ablation study was the absolute necessity of bidirectional connectivity, as removing reverse edges drastically reduced AUROC by nearly 0.2. Time-attentive edge features also contributed significantly to performance. Interestingly, collapsing heterogeneous edge types into a single relation, which reduced parameters, surprisingly outperformed the full model, suggesting potential for simplification. The results support CT-HEG's utility for irregular EHR data, though external validation and fairness audits are still needed.