-
1
Clinical decisions in critical care often vary because evidence does not establish one correct course, leaving room for clinician judgment, local practice, and patient preferences.
-
2
AI trained on routine clinical decisions or outcomes can reduce variation while reproducing biases embedded in prevailing practice.
-
3
For ICU triage, models trained on all admission decisions may standardize recommendations at the average rate, even when best-practice clinicians admit fewer patients.
-
4
The authors propose training models on cases managed by clinicians identified a priori as best-practice using problem-specific, stakeholder-defined criteria.
-
5
This approach requires risk-adjusted evaluation, local calibration, subgroup fairness audits, drift monitoring, and prospective assessment of safety, workflow, clinician behavior, and patient outcomes
