Unmasking Toxic Mimicry in Medical Offline Reinforcement Learning for ICU Sepsis Management via Counterfactual Clinical Audits

Offline reinforcement learning (RL) holds significant potential for optimizing treatment decisions in Intensive Care Units (ICU), particularly for sepsis management. However, current standard evaluation metrics such as Mean Squared Error (MSE) and Fitted Q-Evaluation (FQE) are insufficient as they only assess behavioral imitation. These metrics fail to detect 'Toxic Mimicry,' a critical failure mode where RL agents may replicate harmful patterns, such as inappropriate treatment withdrawal during comfort-care transitions. Researchers propose the Counterfactual Clinical Audit (CCA) framework to stress-test these RL agents against physiological perturbations, anchored in established Surviving Sepsis Campaign (SSC) guidelines, thereby enhancing the reliability and safety of AI-driven medical decision support.

• Offline reinforcement learning (RL) shows promise for optimizing ICU treatment decisions, specifically for sepsis management.
• Standard RL evaluation metrics (MSE, FQE) are limited to assessing behavioral imitation and cannot identify harmful AI behaviors.
• 'Toxic Mimicry' is a critical failure mode where RL agents replicate detrimental patterns, such as inappropriate treatment withdrawal during palliative care.
• The Counterfactual Clinical Audit (CCA) framework is proposed to rigorously evaluate RL agents by stress-testing them with physiological perturbations.
• CCA framework uses Surviving Sepsis Campaign (SSC) guidelines as benchmarks for clinical relevance in audits.

Reported by arXiv — Computers and Society · International. This briefing summarises the publication for executive use; the document itself is not reproduced here.