Integrating Reinforcement Learning With Explainable Artificial Intelligence for Real‐Time Clinical Decision Support in Dynamic Healthcare Environments

Abstract

In critical care settings such as intensive care units (ICUs) and emergency departments, clinical decision support must adapt rapidly to changing patient states while remaining transparent enough to earn clinician trust. This article presents an integrated Reinforcement Learning and Explainable Artificial Intelligence (RL‐XAI) framework for real‐time critical‐care decision support. A Deep Q‐Network (DQN) is trained offline on 18 142 mechanically ventilated ICU stays from the eICU Collaborative Research Database, learning adaptive policies for ventilator settings and the titration and timing of vasopressors and fluids. An explainability layer embedded within the DQN inference loop generates dual‐modal explanations in real time: global, population‐level feature‐importance summaries and local, patient‐specific rationales. Running on a single Intel Xeon Gold 6154 CPU server (3.0 GHz, 18 cores, no GPU), end‐to‐end latency—comprising the DQN forward pass, TreeSHAP attribution (background n  = 100), and LIME surrogate fitting (50 perturbations)—ranged from 210 to 300 ms across 500 held‐out test cases (median 247 ms; 95th percentile 291 ms), well below the 500 ms threshold for clinical actionability. An initial clinician assessment confirmed the explanations’ usefulness and interpretability. The framework advances real‐time, explainable, adaptive AI for safety‐critical healthcare.