The crystal ball problem: why prediction models are not guaranteed to improve health outcomes

Clinical prediction rests on a seductive premise: accurate forecasts of deterioration will prompt earlier intervention and better outcomes. We argue that this assumption is not merely unproven but structurally flawed. The distinction between forecasting and decision evaluation is established yet routinely neglected. Using anticoagulation during extracorporeal membrane oxygenation as a case study, we show why predictive accuracy alone cannot establish clinical benefit, and how causal analysis of existing data can bridge validation and prospective impact trials.