Answers are grounded in this paper and cite their sources.
Predictive reliability of artificial intelligence in healthcare extends beyond discrimination, particularly when models are deployed in populations whose epidemiology changes over time. We study how predictive invariance, distributional shift, and asymmetric decision costs jointly shape the robustness of respiratory mortality risk models. A Monte Carlo experiment with known data-generating mechanisms separates prior-probability shift, covariate shift, concept drift, and joint shift, evaluating discrimination, calibration, and decision regret against a Bayes oracle. A longitudinal retrospective analysis of Brazilian SIVEP-Gripe surveillance data (2019–2025) examines temporal and geographic transportability, out-of-time recalibration, and sensitivity to case mix, outcome ascertainment, and predictor timing. In both settings, discrimination remained stable or recovered, while probability accuracy and decision utility deteriorated; covariate shift and concept drift acted differently; and simple recalibration helped in some but not all settings. Invariance, discrimination, calibration, and decision utility are distinct dimensions of robustness that reliable evaluation of clinical artificial intelligence must examine separately under explicit assumptions about distributional change and asymmetric decision consequences.
No takes yet. Share an insight, caveat, or question.
Guzzi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: