1553: Development and Implementation of a Trajectory-Based Machine Learning Model for Sepsis

Introduction: Sepsis management remains challenging due to heterogeneous clinical trajectories and delayed recognition of deterioration. Current tools lack dynamic risk assessment capabilities, limiting timely interventions. Methods: This multicenter retrospective cohort study aimed to develop and validate a machine learning model for predicting sepsis recovery trajectories and enabling early deterioration alerts. Using data from 47,936 patients across three cohorts (institutional development cohort: n=2,843; internal validation: n=1,213; external validation: MIMIC-III n=25,633, eICU n=18,247), we integrated group-based multi-trajectory modeling (GBTM) with ensemble learning algorithms to classify patients into three trajectories: rapid recovery (41.5%), slow recovery (36.4%), and clinical deterioration (22.1%). Temporal features, including physiological variability metrics and SOFA score dynamics, were analyzed. Results: The model demonstrated robust discrimination, achieving AUROCs of 0.84 (95% CI: 0.82–0.86) in the development cohort and 0.80–0.82 in external cohorts, with a median deterioration warning time of 17.6 hours (IQR: 11.7–23.5). Was associated with reduced heart rate variability (HR SD < 10 bpm) emerged as a key predictor of deterioration, associated with a 2.17-fold mortality risk increase (adjusted HR: 2.17, 95% CI: 1.68–2.79). Implementation of trajectory-guided interventions was associated with reduced ICU length of stay (mean difference 1.8 days, 95% CI: 1.2-2.4), mechanical ventilation duration (2.3 days, 95% CI: 1.5-3.1), and 28-day mortality (absolute reduction 5.7%, 95% CI: 2.3-9.1). Conclusions: This study establishes a validated framework for dynamic sepsis prognosis, leveraging temporal physiological patterns to enable personalized interventions. The model’s generalizability across diverse healthcare settings and its clinical impact highlights its potential as a decision-support tool for precision critical care.