Learning a Continuous Sepsis Severity Score Without Hour-by-Hour Supervision: A Two-Site Retrospective Study

TL;DR - This two-site retrospective study learns a continuous, hourly sepsis severity score from patient trajectories using mortality as a treatment-level ranking signal rather than requiring per-hour labels. The resulting index tracks outcomes and clinically relevant changes, suggesting potential as a decision-support complement to clinician judgment.

• The model uses 43 routinely charted variables over a 72-hour treatment window across cohorts of 29,116 and 7,691 adults meeting Sepsis-3 criteria.
• Within every baseline SOFA-2 stratum, non-survivors scored 1.19–1.64 points higher than survivors on the 0–10 scale; stratification by lactate, mean arterial pressure, and creatinine yielded similar separation.
• Within-patient score changes correlated most strongly with lactate changes (Spearman ρ = 0.39; n = 1,854), with weaker correlations for mean arterial pressure and creatinine.
• Models trained at different institutions achieved 70–77% of same-site cohort-level correlation, while external within-patient correlations reached 0.54 and 0.59 against estimated ceilings of 0.92 and 0.90.