The early warning paradox.

Logan Ellis, Hugh; Palmer, Edward; Teo, James T; Whyte, Martin; Rockwood, Kenneth; Ibrahim, Zina · NPJ Digit Med · 2025

retrospective_cohort · Level III

Where this comes from

Abstract

Machine learning models in healthcare aim to predict critical outcomes but often overlook existing Early Warning Systems’ impact. Using data from King’s College Hospital, we demonstrate how current evaluation methods can lead to paradoxical results. We discuss challenges in developing ML models from retrospective data and propose a novel approach focused on identifying when patients enter a ‘risk state’ through latent health representations, potentially transforming clinical decision-making.