Abstract

Objective: This framework aimed to combine real-time flexibility with causal reasoning, leading to more reliable, explainable, and clinically compatible AI-informed decisions in the emergency environment.
Methods: The causal dynamic feature importance on triage (CDFI-Triage) framework was designed to support real-time clinical decision-making by continuously adjusting the importance of physiological features based on evolving causal relationships with the onset of critical illness. A time-evolving causal graphs, updates causal strengths were conducted using an adaptive Kalman filter and approximates interventional probabilities with a neural causal estimator employing counterfactual dropout. The system integrated causal features with non-causal contextual data through a modified Temporal Fusion Transformer, producing both predictions and interpretable explanations of the evolving physiological drivers of risk.
Result: Baseline models performed worse than the CDFI-Triage in predictive performance (maximum AUC = 0.80) using the MIMIC-III database (CDFI-Triage 1-6-hour horizons (AUC = 0.89, 0.87, 0.85, 0.83)). The highest physiological plausibility (Causal Consistency Score = 0.86 + ± 0.03, p < 0.01) and adaptation to both acute and gradual physiological changes (mean update time = 8.2-15.7 min) were also obtained with the framework.
Conclusion: The CDFI-Triage framework integrated dynamic causal reasoning with real-time monitoring, improving prediction accuracy and interpretability. It adapted feature importance according to evolving patient states, distinguishing proper physiological drivers from noise. The system provided transparent and actionable insights for clinicians in real-time triage.

Key words: AI, causal dynamic feature importance, emergency department, triage.