Interpretability of Concept Bottleneck Models for Tuberculosis Triage Under Chest Radiograph Domain Shift: A Healthcare Calibration Fairness Evaluation

Concept bottleneck models (CBMs) promise interpretable tuberculosis triage by routing every prediction through a small set of clinically named concepts, yet their calibration and subgroup fairness under chest-radiograph domain shift remain uncharacterised. The analysis treats the CBM as a two-stage estimator whose concept encoder and label head are separately exposed to shift, and it characterises the resulting calibration gap as a function of concept-conditional prevalence drift and label-head sensitivity. A design-based evaluation is specified to quantify the operating characteristics of the resulting fairness audit under a declared variance structure with an intraclass correlation of 0.64, between-cluster SD of 6.0 and residual SD of 4.5; power was estimated across 2000 design cells for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, and 56 clusters are required for 5 percentage points at 80% power. Concept-level transparency and label-level calibration are not co-extensive, that subgroup fairness audits of CBMs must be powered at the concept stratum rather than the label stratum, and that interpretability claims for clinical triage models should be indexed to the shift regime under which they were established.