ICU sepsis alerts stop at yes-or-no: SepsisFlow puts warning, onset context and a 72-hour course on one hourly clock

ICU sepsis warning becomes three linked judgements along one hourly clock: hourly 6-, 12- and 24-hour risk before onset, a patient state and onset context at onset, and a description of the 72-hour organ-support course after onset. Earlier warning models predicted only whether sepsis would occur, leaving clinicians uncertain of the onset context and early course once an alarm sounded; published sepsis phenotypes anchor at hospital presentation or triage, which does not coincide with ICU-acquired onset. On the MIMIC-IV internal test set of 4,161 ICU stays, 997 cases and 182,415 hourly scores, the 24-hour AUROC was 0.822 (95% CI 0.807–0.836) and AUPRC 0.292 (0.270–0.317); 16 published algorithms retrained on the same cohort, Sepsis-3 label and hourly protocol reached AUROC 0.747–0.808.

At equal alarm burden, hourly re-scoring by SepsisFlow identifies more patients about to develop sepsis than the bedside scores NEWS2, qSOFA and SIRS. Bedside scores discriminate poorly in the ICU and are usually computed once in evaluations, whereas deployed models are updated continuously. At the NEWS2≥5 burden of 46.0 first alarms per 100 patient-days, SepsisFlow identified 76.4% (73.7–79.1) of cases against 56.8% (53.7–59.8) for NEWS2≥5, 19.7 percentage points more, and 3.5 points (1.9–5.1) more than a context-only reference using age, sex, first ICU unit and time since ICU admission.

The patient state at onset clusters into three onset contexts reproducible by two bedside questions: did onset occur within 12 hours of ICU admission, and if so was the first ICU the cardiovascular ICU. Published sepsis phenotypes such as the Seymour α–δ classes and the SepsisFormer coagulation–inflammation subtypes require a full laboratory profile that is not always available at onset. The two questions agreed with the model contexts in 87.9% (85.8–89.8, Cohen's κ=0.81) of test-set cases; with live discharge as a competing event, cumulative 28-day in-hospital death was 17.1% (13.4–21.2), 15.1% (11.8–18.7) and 2.2% (0.8–4.8) across the three contexts (Gray's test P=2.4×10⁻⁷).

Organ-support trajectories actually recorded in the 72 hours after onset fall into three courses, and within the same context persistent support carries about fivefold the 28-day in-hospital mortality of de-escalation. Earlier course studies described SOFA or temperature trajectories starting at admission, while warning studies stopped at onset. From the 72-hour landmark (891 patients still in hospital), 28-day in-hospital mortality was 21.5% (17.3–26.6) for persistent organ support, 4.4% (2.7–7.2) for de-escalation and 2.0% (0.8–4.5) for early ICU discharge; within Late ICU onset it was 24.3% against 4.7%, and within Early onset, non-CVICU 20.4% against 4.3%.

SepsisFlow was developed in a single health system with a random split by patient, and external validation covered only the pre-onset warning, using antimicrobials to define infection and concentrating cases in one eICU hospital; the onset contexts and 72-hour courses were evaluated on the MIMIC-IV test set.

The organ-level models cover only the next 24 hours, the three early courses classify trajectories already recorded in the 72 hours after onset rather than predicting them, and pre-onset risk did not predict the later course beyond a bedside onset model; the event-stream encoder processes numeric events of 31 concepts without treatments, antimicrobials, cultures or time gaps, and the model gives no estimate of its own uncertainty.