Safety-aware explainable deep reinforcement learning for nephrotoxic medication management in critical care

Antimicrobial agents are essential for sepsis treatment but often pose significant nephrotoxic risks. Current treatment policies rarely account for the trade-off between therapeutic efficacy and adverse outcomes such as acute kidney injury, limiting their clinical applicability. To address this gap, we present a safety-aware decision support system that models both efficacy and toxicity using a Markov Decision Process and state-of-the-art offline reinforcement learning algorithms. Our agent explicitly predicts the risks of eight terminal outcomes, including mortality, renal replacement therapy, and septic shock, for each nephrotoxic medication. It learns decomposed value functions to distinguish between therapeutic benefits and safety concerns, enabling interpretable and risk-calibrated recommendations. The model achieved strong predictive performance with AUROC scores of 0.852 for ICU mortality, 0.923 for dialysis, and 0.843 for septic shock. To guide clinical decisions, the agent issues early warnings when patient states exceed safety thresholds and advises against high-risk medications based on value estimates. By integrating statistical validation and sensitivity analysis, our framework ensures that treatment recommendations are both clinically interpretable and aligned with physician judgment, ultimately advancing the safe application of AI in critical care.