Interpretable Adaptive Graph Fusion Network for Mortality and Complication Prediction in ICUs

Background: This study introduces the Adaptive Graph Fusion Network, an interpretable graph-based learning framework developed for large-scale prediction of intensive care outcomes. The proposed model dynamically constructs patient similarity networks through a density-aware kernel that adjusts neighborhood size based on local data distribution, thereby representing both frequent and rare clinical patterns. Methods: To characterize physiological evolution over time, the framework integrates a short-horizon convolutional encoder that captures acute variations in vital signs and laboratory results with a long-horizon recurrent memory unit that models gradual temporal trends. The approach was trained and internally validated on the publicly available eICU Collaborative Research Database, which includes more than 200,000 admissions from 208 hospitals across the United States. Results: The model achieved a mean area under the receiver operating characteristic curve of 0.91 across six critical outcomes, with in-hospital mortality reaching 0.96, outperforming logistic regression, temporal long short-term memory networks, and calibrated Transformer-based architectures. Feature attribution analysis using SHAP and temporal contribution mapping identified lactate trajectories, creatinine fluctuations, and vasopressor administration as dominant determinants of risk, consistent with established clinical understanding while revealing additional temporal dependencies overlooked by existing scoring systems. Conclusions: These findings demonstrate that adaptive graph construction combined with multi-horizon temporal reasoning improves predictive reliability and interpretability in heterogeneous intensive care populations, offering a transparent and reproducible foundation for future research in clinical machine learning.