Machine learning can create better clinical severity scores by learning from patient trajectories and mortality outcomes rather than relying on decades-old fixed formulas—and you don't need expensive per-timestep labels to do it.
Researchers built a machine-learned sepsis severity score from patient data instead of using outdated fixed formulas. Using 43 routine vital signs and lab values over 72 hours from nearly 37,000 patients across two hospitals, they trained a model that ranks patients by mortality risk without requiring hour-by-hour labels.