Version 2 (June 2026) — Correction: The gradual-drift detection ratios for the Isolation Forest on the FHGD cohort have been recomputed following a pipeline error (flat drift applied instead of the intended linear ramp). The FHGD IF gradual multivariate DR is 1.89× (previously 2.91×); this advantage is therefore cohort-dependent. All Pima results, all OCSVM results, and all abrupt-drift results including the principal result (OCSVM abrupt multivariate, FHGD, DR = 3.18×) ,are unchanged. The hybrid monitoring recommendation is unchanged. AI models in clinical practice are susceptible to silent failure, where unmonitored population drift degrades diagnostic performance without triggering system alerts. To evaluate methods for detecting such shifts, we developed a chronologically partitioned, leak-free experimental pipeline using two independent cohorts: the Pima Indians Diabetes dataset (n=768) and the Frankfurt Hospital Glucose Dataset (FHGD, n=2,000). We compared the sensitivity of unsupervised anomaly detection algorithms against simulated clinical failure modes, including gradual physiological deterioration and sudden affine shocks. The analysis revealed a clear performance trade-off dependent on the morphology of the data shift. Distance-based One-Class Support Vector Machines (OCSVM) were highly sensitive to abrupt systemic shocks, achieving a Detection Ratio (DR) of 3.18× in the FHGD cohort. Conversely, partitioning-based Isolation Forests showed stronger detection of gradual multivariate erosion in the Pima cohort (DR = 4.40×), although this advantage was cohort-dependent. This study is reported in accordance with the STARD-AI 2025 guideline for AI-centred diagnostic accuracy. These findings indicate that relying on a single unsupervised algorithm leaves diagnostic systems vulnerable to specific drift profiles. Ensuring the clinical safety of deployed medical AI will therefore require a hybrid monitoring framework that integrates both distance and partitioning logic to detect the full spectrum of population drift.
Malik Adeel Anjum (Wed,) studied this question.