AI in Sepsis Prediction

Sepsis is the body's dysregulated response to infection, leading to organ dysfunction. Because early antibiotics and fluids improve survival, hospitals have long used rule-based screens such as SIRS criteria or qSOFA. AI early-warning systems go further: they combine many variables, such as heart rate, respiratory rate, temperature, white blood cell count, lactate, creatinine, medications and trends over time, into a single risk score that updates as new data arrive in the electronic health record (EHR). The most instructive case is the Epic Sepsis Model, a proprietary tool deployed in many US hospitals. A 2021 external validation at Michigan Medicine, published in JAMA Internal Medicine, found it performed substantially worse than its developer had reported: it missed about two thirds of sepsis cases and generated alerts on roughly one in five hospitalized patients, most of whom did not have sepsis. The study became a reference point for a broader lesson: a model validated on its developer's data may not hold up at a different hospital, and accuracy metrics alone say little about whether an alert actually helps. More encouraging evidence came from Johns Hopkins' TREWS system, where studies published in Nature Medicine in 2022 reported better outcomes for patients whose alerts were confirmed promptly by a clinician. Those studies were observational rather than randomized, so how much the tool itself caused the improvement is still debated. In 2024 the FDA authorized the Prenosis Sepsis ImmunoScore, one of the first AI sepsis tools to go through FDA review. A common misconception is that a sepsis model diagnoses sepsis. It flags risk; clinicians still assess the patient. Another is that more alerts mean more safety. Alert fatigue, where staff become desensitized after frequent false alarms, can cause real warnings to be ignored. Success depends on threshold choice, who receives the alert, and what action is expected afterward.