Adaptive Thresholding in AI-CDSS: Dynamic Clinical Decision Boundaries for Evolving Patient Populations — A Narrative Review

Abstract

Background. Artificial intelligence-based clinical decision support systems (AI-CDSS) typically convert a model's predicted risk into an actionable recommendation by comparing the output against a fixed decision threshold. Such static boundaries implicitly assume a stable patient population and data distribution, an assumption that fails when case-mix, disease prevalence, and clinical workflows drift over time. Objective. This narrative review synthesises the emerging literature on adaptive thresholding — the set of mechanisms by which AI-CDSS dynamically adjust their clinical decision boundaries in response to evolving patient populations — mapping architectural patterns, theoretical foundations, empirical evidence, and implementation and regulatory implications. Key Contributions. Across 28 sources published between 2025 and 2026, we identify three adaptive-thresholding architectural patterns — post-hoc calibration and recalibration, dynamic temporal risk modelling, and adaptive multi-agent architectures — and formalise threshold adaptation as Bayesian decision-making under distribution shift. The review finds that calibration and recalibration methods are well developed at the framework level, that several retrospective studies demonstrate improved decision utility from threshold optimisation, and that no prospective randomised controlled trial has used missed-finding reduction as a primary endpoint. Conclusions. Adaptive thresholding is a conceptually coherent paradigm with growing methodological support, but its clinical benefit remains unproven in prospective designs. Priority research directions include drift-monitoring benchmarks, subgroup-calibrated decision boundaries, and regulatory frameworks for continuously updating models.