icu

This paper introduces the Counterfactual Clinical Audit (CCA) framework to evaluate offline reinforcement learning agents for ICU sepsis management, exposing 'toxic mimicry' where agents replicate harmful treatment patterns that standard metrics miss. Using MIMIC-III data, it shows a Medical Decision Transformer fails to escalate vasopressors under rising lactate, while a causal transformer performs safely.

This paper presents VDSS, a human-in-the-loop multi-agent framework for ventilator decision support that uses contextual bandit preference learning to adapt to clinician-specific tuning styles, with retrospective ICU trajectory replays showing improved recommendation acceptability and reduced interaction rounds.

RealICU is a hindsight-annotated benchmark for evaluating LLMs in ICU settings, covering four physician-motivated tasks. Experiments reveal that existing LLMs struggle with recall-safety tradeoffs and anchoring bias, while a new structured-memory agent improves reasoning but not fully eliminate safety failures.

This paper introduces a stochastic causal representation learning framework to resolve the bias-precision paradox in personalized medicine, demonstrating improved accuracy and interpretability in ICU clinical decision support.