Right Diagnosis, Unsafe Route: Auditing Clinical AI Decision Trajectories and Corrective Oversight for Validity and Potential Harm
This paper proposes a trajectory audit that replays real hospital admissions checkpoint by checkpoint, letting each AI system decide using only the evidence released so far, and rates every test, treatment and disposition for potential patient harm and clinical validity. An LLM judge scores each decision, calibrated against a physician reference standard built from triple annotation and senior adjudication of a stratified sample. Across 300 cases from a hospital cardiology cohort, MIMIC-IV and the ASCENT benchmark, with up to seven AI systems per source, 55.2% of hospital trajectories that reached an acceptable final diagnosis passed through a decision with moderate or severe potential harm, and 24.8% through a clinically unacceptable one; harm rates were 76.2% on MIMIC-IV and 39.6% on ASCENT. An oversight audit found structured corrections resolved 73.6% of first errors versus 24.8% after bare rejection, but new potentially unsafe decisions followed 30.2% of corrections versus 19.7% of rejections.
Why it matters: It shows that correct final diagnoses and successful error corrections can still hide unsafe intermediate clinical AI decisions, giving evaluators a concrete method and benchmark evidence for auditing decision trajectories rather than outcomes alone.