An Explainable Leakage-Audited Framework for Multi-Horizon Sepsis-3 Onset Forecasting from Complex ICU Dynamics
Chaos and Fractals · 30.07.2026 · 3 (2) · 76-91
DOI: 10.69882/adba.chf.2026072 ↗Kurz gefasst
Fachartikel in der Zeitschrift Chaos and Fractals, online erschienen am 30.07.2026. Verfasst von Mohammed Mansour und 1 weiteren Autor:innen. Die Arbeit nennt 21 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Themen: Sepsis
Abstract (Originalsprache)
Sepsis-3 onset forecasting remains challenging because event prevalence is low at each patient-time landmark and retrospective electronic health record (EHR) models may inadvertently exploit workflow signals that reflect clinician suspicion rather than underlying patient physiology. We present an explainable, leakage-audited, multi-horizon LightGBM framework for physiology-based Sepsis-3 onset forecasting using MIMIC-IV v3.1. Adult first ICU stays with at least 8 hours of observation were transformed into hourly physiologic trajectories, Sepsis-3 onset labels, and repeated landmark prediction examples. Separate models estimated the probability of Sepsis-3 onset within the next 4, 6, 12, and 24 hours. The final cohort comprised 64,363 ICU stays, including 42,291 forecastable stays and 6,431 forecastable Sepsis-3 cases. The deployment-oriented no-admission-year CLEAN model achieved held-out AUROCs of 0.782, 0.772, 0.750, and 0.720 for the 4-, 6-, 12-, and 24-hour prediction horizons, respectively, with patient-level bootstrap 95% confidence intervals of 0.763–0.801, 0.752–0.792, 0.729–0.769, and 0.699–0.742. A paired leakage audit demonstrated that removing vasopressor exposure, measurement-age channels, and observation-mask channels resulted in AUROC changes of less than 0.01, indicating that predictive performance was primarily driven by physiologic information rather than workflow-related artifacts. Temporal validation on the latest deidentified admission-year group yielded a 6-hour AUROC of 0.762. At the 6-hour horizon, a top-2% no-year alarm policy with a 6-hour cooldown detected 17.8% of sepsis cases within 0–6 hours, 29.3% within 0–24 hours, and 40.3% before onset overall while generating 15.5 alerts per 100 patient-days. SHAP-based explainability analyses consistently identified neurologic status, temperature, urine-output dynamics, renal markers, inflammatory variables, vital signs, and lactate-related dynamics as the dominant predictors across forecasting horizons. These findings provide a robust and explainable physiology-based framework for future studies on Sepsis-3 onset forecasting in complex critical-care systems.
Zitieren
APA
Mansour, M., & Donmez, T. B. (2026). An Explainable Leakage-Audited Framework for Multi-Horizon Sepsis-3 Onset Forecasting from Complex ICU Dynamics. Chaos and Fractals, 3(2), 76-91. https://doi.org/10.69882/adba.chf.2026072
Vancouver
Mansour M, Donmez TB. An Explainable Leakage-Audited Framework for Multi-Horizon Sepsis-3 Onset Forecasting from Complex ICU Dynamics. Chaos and Fractals. 2026;3(2):76-91. doi:10.69882/adba.chf.2026072
BibTeX
@article{Mansour2026,
author = {Mansour, Mohammed and Donmez, Turker Berk},
title = {An Explainable Leakage-Audited Framework for Multi-Horizon Sepsis-3 Onset Forecasting from Complex ICU Dynamics},
journal = {Chaos and Fractals},
year = {2026},
volume = {3},
number = {2},
pages = {76--91},
doi = {10.69882/adba.chf.2026072}
}
EUVIMED