Interpretability Audits of Electronic Health Record Mortality Predictors Under Clinician Override Workflow Substitution: A Computer Calibration Fairness Evaluation

Electronic health record mortality predictors are routinely embedded in clinical workflows where clinicians may override their outputs, yet the interpretability of these models is typically audited on the model's own outputs rather than on the substituted workflow that actually reaches the patient. The framework distinguishes three audit objects: the unsubstituted model, the override-conditioned workflow, and the substituted decision rule that results when overrides are applied. For each object, calibration is defined as a property of the conditional distribution of outcomes given the decision actually taken, and fairness is defined as parity of calibration error across protected subgroups. Two theorems are proved: first, that override substitution preserves calibration only when the override probability is conditionally independent of the outcome given the model score, a condition that fails whenever clinicians respond to unobserved severity; second, that subgroup calibration error under substitution is bounded by the product of override-rate disparity and the conditional outcome variance within score strata. The framework is illustrated with a reanalysis of published calibration audits of electronic health record mortality predictors, and the implications for interpretability reporting are stated as a set of audit requirements. The contribution is a formal account of why interpretability audits that ignore override substitution can certify a model that the workflow does not deploy, and a computable criterion for detecting when this occurs.