AI-Assisted Audit of Intensive-Care Risk Scores under Missing Laboratory Data: Evidence, Failure Modes and Deployment Trade-offs

Abstract

Intensive-care risk scores can become unreliable when laboratory results arrive late or are missing at the moment clinicians need a prediction. An evaluation that overlooks these constraints may reach conclusions that cannot be reproduced. This study aims to test how missing-data mechanisms affect the calibration and usefulness of AI-based deterioration alerts. The evaluation would follow a retrospective, patient-level validation protocol using de-identified longitudinal ICU records, with hospital and calendar-period holdouts. The choice of evaluation units and exclusion criteria is informed by the broader methodological record [3–7]. The analysis will contrast complete-case scoring, imputation and missingness-aware prediction under matched alert budgets. The evidence-mapping component examines why a successful result in one technical environment may fail elsewhere. The technical pipeline must remain auditable: prior work on storage operations provides system-level context [1], while a second source identifies infrastructure conditions relevant to data continuity [2], rather than patient-level outcomes. Assessment will focus on calibration error, sensitivity at a fixed alert rate, subgroup performance and time to actionable warning. The supporting evidence base motivates the chosen comparator conditions and robustness criteria [8–12]. Critical sensitivity checks address patient-level leakage, changes in test-ordering practices and clinicians’ review of ambiguous outcomes. Further systems and modelling literature supplies context for the proposed sensitivity analyses [13–17]. Stress tests would distinguish outright failures, reduced confidence and cases where a simpler baseline is preferable. The expected output is a reusable evaluation framework with transparent uncertainty and reporting rules, informed by wider reproducibility literature [18–23].