Interpretability of Concept Bottleneck Alignment for Artificial Intelligence Triage Under Clinical Guideline Revision: A Healthcare Calibration Stability Estimation

Concept bottleneck models (CBMs) promise interpretable clinical triage by routing predictions through human-readable concepts, but their calibration stability under clinical guideline revision is not well characterised. The framework treats concept alignment as a constrained projection problem and characterises the bias introduced when the concept set is revised without retraining. A replication-based evaluation of the before-after-control-impact estimator was conducted under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), with power estimated across 2000 design cells for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The analysis indicates that 54 clusters are required for 5 percentage point detection at 80% power. The framework proposes that calibration stability is governed by the alignment residual between the deployed concept set and the revised guideline concept set, and that this residual can be estimated prospectively from the design parameters. The contribution is a formal characterisation of when concept bottleneck triage models remain calibrated across guideline revisions and a design-level estimator for planning stability evaluations.