One model predicts both sepsis and heart failure, despite opposite treatments.

In intensive care, sepsissepsisA life-threatening condition where the body's response to an infection damages its own tissues. and heart failureheart failureA chronic condition where the heart cannot pump enough blood to meet the body's needs. look similar early on, but treating them requires opposite actions. Today, computer models only watch for one of these conditions at a time. Researchers built a single system to predict both conditions at once. The model uses two neural network streams to process forty-eight hours of hourly vital signs alongside seven days of daily lab tests. It connects these streams using an attention mechanism to weigh how different clues relate, blends in tree-based data, and balances both tasks together. An upgraded version adds summary statistics, patient traits, extra pretraining, and multiple runs combined. Forty-eight hours after admission, this upgraded model achieved strong internal prediction scores of 0.867 for sepsis and 0.899 for heart failure. However, accuracy dropped when predicting illness four to twelve hours before it began, and scores fell when tested on patients from different hospitals.

Why it matters

In intensive care, sepsis and heart failure can look alike early on, but treating one when the patient has the other can be dangerous. Existing models only predict one condition at a time, forcing clinicians to choose which to monitor. This work shows a single model can predict both simultaneously, potentially reducing diagnostic errors and improving survival rates.

Questions answered