AI Can Now Explain ICU Death Risk Clearly

Can AI explain its own ICU mortality predictions? A new feasibility study compared a standalone large language model with a multi-step agentic pipeline. On the eICU Demo dataset, the agentic pipeline produced no explanations with explicit outcome leakage, while the standalone model produced one. The agentic pipeline scored higher on guideline grounding, value specificity, and plausibility, whereas the standalone model showed higher alignment with SHAP attributions and direction consistency. The authors suggest that agentic decomposition may improve safety-relevant grounding and patient-specific detail, but should be paired with attribution-based checks before use in high-stakes risk explanation.