Interpretability of Concept Bottleneck Models for Artificial Intelligence Triage Under Emergency Department Overcrowding: A Healthcare Safety Evaluation

Concept bottleneck models (CBMs) promise interpretable triage by routing predictions through clinically meaningful concepts, yet their safety under emergency department overcrowding remains uncharacterised. The analysis characterised the operating characteristics of the estimator rather than observing any clinical deployment. Power reached 80% at 54 clusters for a 5-percentage-point effect; minimum detectable effects at 80% power were derived for each cluster count. The results indicate that the BACI estimator is well calibrated for detecting moderate safety degradation in CBM-mediated triage, but that small cluster counts (8 and 16) yield unacceptably low power for effects below 5 percentage points. Interpretability claims for CBMs cannot be separated from the statistical design used to audit them, and it proposes a minimum reporting standard for CBM safety evaluations in overcrowded emergency settings.