AI-Assisted Audit of Intensive-Care Risk Scores under Missing Laboratory Data: A Mixed-Evidence Evaluation

Abstract

Intensive-care risk scores can become unreliable when laboratory results arrive late or are missing at the moment clinicians need a prediction. The objective is to test how missing-data mechanisms affect the calibration and usefulness of AI-based deterioration alerts. Relevant methodological and data-quality literature informs the design of the sampling and validation strategy [3–7]. The proposed method is a retrospective, patient-level validation protocol using de-identified longitudinal ICU records, with hospital and calendar-period holdouts. Comparators will include complete-case scoring, imputation and missingness-aware prediction under matched alert budgets. Enterprise-storage sources are used to specify constraints on data availability and provenance [1], with complementary infrastructure evidence defining practical recovery safeguards [2]. These citations are not treated as clinical effectiveness evidence. The mixed-evidence component combines quantitative tests with an independently coded practitioner evaluation. Statistical and computational sources help define the reproducibility checks and comparison procedures [8–12]. The primary outcomes are calibration error, sensitivity at a fixed alert rate, subgroup performance and time to actionable warning. Potential biases include patient-level leakage, changes in test-ordering practices and clinicians’ review of ambiguous outcomes. Differences between numerical metrics and professional judgements would be retained in the interpretation. Additional research informs the prespecified stress tests and limits on interpretation [13–17]. The intended contribution is an auditable research protocol rather than an unsupported claim of effectiveness; related work frames its transferability limits [18–23].