Interpretability Bounds for Computer Vision Sepsis Detection Under Bedside Monitor Alarm Suppression Policies: A Calibration Robustness Evaluation

Computer vision sepsis detection models that ingest bedside monitor waveforms are increasingly proposed as surveillance adjuncts, yet their interpretability guarantees are rarely evaluated under the alarm-suppression policies that govern which monitor events are retained. the present analysis formalises the interpretability bound as the maximum calibration drift a saliency attribution can tolerate before its clinical ranking inverts, and derives the bound for a class of additive attribution maps under threshold-based suppression. The analysis is a design-based evaluation of the before-after-control-impact estimator used to recover suppression effects, with no observational data: the operating characteristics are characterised under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64) across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effects of 3, 5 and 8 percentage points. Power reaches conventional levels at 54 clusters for a 5-percentage-point effect, and the minimum detectable effect at 80% power falls steeply as clusters increase. The derived bound shows that suppression-induced prevalence shifts translate into attribution-rank inversions at calibration drift well below the drift that leaves discrimination intact, so discrimination-preserving recalibration is insufficient to preserve interpretability. The contribution is a formal bound, a design-level power characterisation, and the identification of a calibration-robustness gap that alarm-suppression policy must close before saliency maps can be treated as clinically stable.