Cuomo’s Paradox and the Design of Artificial Intelligence in Medicine: A Stage-Specific Framework for Learning, Evaluation, and Deployment

Abstract

Artificial intelligence is rapidly diffusing across clinical settings where it informs screening, triage, prognosis, and treatment selection. Yet many clinical prediction tasks sit at the intersection of disease prevention and disease management, where the direction and magnitude of associations between risk factors and outcomes can reverse across the disease trajectory. Cuomo’s Paradox refers to this stage specific reversal, in which factors that correlate with lower incidence may correlate with improved survival after diagnosis for reasons that include treatment tolerance, physiologic reserve, and selection processes. Artificial intelligence systems that learn from prevention oriented labels and features, then are deployed to guide care for patients who already have the disease, are vulnerable to systematic error, inequity, and unintended harm. In this manuscript, we formalize the implications of Cuomo’s Paradox for artificial intelligence across the model lifecycle. We conduct a narrative synthesis of empirical literatures on reverse epidemiology, heterogeneity of treatment effect, distribution shift, and clinical artificial intelligence, and we derive a causal framework that explains how conditioning on diagnosis and treatment can flip associations in ways that mislead models. We then present worked examples and thought experiments that show how objective mismatch, label leakage, stage misclassification, and treatment policy confounding can produce clinically significant misranking and calibration errors. We map these failure modes to practical remedies that include stage specific objectives and datasets, explicit modeling of treatment and physiologic reserve, counterfactual evaluation strategies aligned with target trial principles, and reporting practices that communicate the intended stage of use. The discussion connects these proposals to existing reporting and evaluation guidance for artificial intelligence and clinical prediction, and outlines a research agenda for paradox aware learning and deployment. We argue that Cuomo’s Paradox should be treated as a core design constraint for clinical artificial intelligence, not a peripheral curiosity. Systems that respect stage specific biology and clinical context can improve validity, safety, and equity, while systems that ignore it risk consistent error where patients are most vulnerable. The result is a blueprint for developing artificial intelligence that is aligned with the realities of disease trajectories and the goals of precision medicine (Obermeyer et al., 2019; Kent et al., 2018; Cuomo 2025; Collins et al., 2015; Wolff et al., 2019).