From Prediction to Action: Why Bleeding Risk in Cardiac Surgery Still Eludes Us in the Era of Artificial Intelligence

Perioperative bleeding in cardiac surgery remains common, consequential, and difficult to predict despite decades of research and expanding data availability. Established risk scores and newer analytical approaches have improved risk stratification, yet few tools have translated into real-time bedside decision support or demonstrated reductions in bleeding-related complications. This gap reflects not simply a failure of technology, but a failure of alignment between prediction and clinical need, data and context, and innovation and implementation. Current approaches often emphasize statistical performance over clinical utility, rely on surrogate endpoints, and apply static models to a dynamic clinical problem. This article argues that the key challenge is no longer whether bleeding can be predicted but whether prediction can meaningfully improve care. Future progress will require clinically relevant outcomes, time-varying models, rigorous external validation, interpretability appropriate to the use case, and seamless workflow integration. Prediction should ultimately be judged by its ability to support timely decisions and improve outcomes.