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

Abstract

Intensive-care risk scores can become unreliable when laboratory results arrive late or are missing at the moment clinicians need a prediction. This gap has consequences for interpreting performance outside carefully curated evaluation settings. The planned investigation seeks 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 protocol specifies 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. 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. The reproducibility extension records dataset versions, model settings, exclusion rules and provenance before any held-out 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. An independent replication exercise would repeat the main analysis from the documented artefacts and quantify disagreements. 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].