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

Area of Science:

• Clinical machine learning
• Artificial intelligence in healthcare
• Predictive modeling for critical care

Background:

• Introduces the Adaptive Graph Fusion Network, an interpretable graph-based learning framework.
• Dynamically constructs patient similarity networks using a density-aware kernel.
• Represents both frequent and rare clinical patterns for comprehensive analysis.

Purpose of the Study:

• Develop a large-scale prediction framework for intensive care outcomes.
• Enhance the interpretability and reliability of clinical prediction models.
• Identify key determinants of risk in heterogeneous intensive care populations.

Main Methods:

• Integrates a short-horizon convolutional encoder for acute variations and a long-horizon recurrent memory unit for temporal trends.
• Employs adaptive graph construction with a density-aware kernel for dynamic patient similarity networks.
• Trained and validated on the eICU Collaborative Research Database (>200,000 admissions).

Main Results:

• Achieved a mean AUC of 0.91 across six critical outcomes, with in-hospital mortality at 0.96.
• Outperformed logistic regression, LSTM, and Transformer-based architectures in predictive accuracy.
• Identified lactate, creatinine, and vasopressor administration as key risk determinants via SHAP and temporal mapping.

Conclusions:

• Adaptive graph construction and multi-horizon temporal reasoning enhance predictive reliability and interpretability.
• The framework offers a transparent and reproducible foundation for clinical machine learning research.
• Demonstrates improved prediction for heterogeneous intensive care populations, aligning with clinical understanding.