AI-Assisted Audit of Intensive-Care Risk Scores under Missing Laboratory Data: A Cross-Site Validation Agenda

Abstract

Intensive-care risk scores can become unreliable when laboratory results arrive late or are missing at the moment clinicians need a prediction. The difficulty is particularly visible when a system is transferred to a new setting. 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 cross-site extension holds out an entire institution or deployment location rather than relying only on random test splits. 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. Between-site heterogeneity would be examined through prespecified subgroup estimates and transparent domain-shift diagnostics. 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].