C54-04 Machine Learning-driven Sepsis Prediction and Endotype Discovery: A Systematic Review and Meta-analysis

Background Artificial intelligence (AI) and machine-learning (ML) frameworks are redefining the understanding of sepsis by identifying distinct biological endotypes and improving early risk stratification. However, heterogeneity in modeling approaches and outcome definitions has impeded direct clinical translation. This systematic review and meta-analysis synthesized quantitative evidence on diagnostic performance of ML models and qualitatively examined data-driven sepsis phenotyping frameworks across critical-care populations. Methods Following PRISMA-MA 2020 methodology, eight studies published between 2015 and 2025 were included: six ML prediction studies reporting model discrimination metrics and two phenotyping investigations. Random-effects inverse-variance meta-analyses (DerSimonian-Laird estimator) were performed on logit-transformed AUROC, sensitivity, and specificity. Heterogeneity was quantified with I² and τ². Phenotype-only studies were narratively summarized to contextualize mechanistic insights beyond predictive performance. Results Six ML studies collectively analyzed >12 000 patients, with average 58 % male predominance, from multiple international ICU databases. The pooled sensitivity 0.86 [0.75-0.93] and specificity 0.86 [0.71-0.94]; heterogeneity was substantial (I² > 95 %). Despite methodological variation, ensemble gradient boosting, recurrent neural networks, and hybrid rule-based models consistently outperformed traditional severity indices such as SOFA and SAPS II. Model interpretability analyses frequently highlighted lactate kinetics, MAP trends, and inflammatory biomarkers as dominant features. Unsupervised and semi-supervised frameworks—including k-means, Two-Step Cluster, and finite-mixture modeling—consistently delineated 2-5 reproducible sepsis phenotypes differentiated by metabolic signatures, inflammatory profiles, and organ-dysfunction trajectories. Low-lysophospholipid metabolic states and NK-cell-exhaustion transcriptomic patterns were repeatedly associated with excess mortality. Predictive tools such as SORP (AUC 0.95) and SepxFindeR provided accurate early-warning capability and biological interpretability. Individualized treatment simulations, including Dynamic Treatment Regimens, suggested phenotype-specific responses to fluid and vasopressor therapy. Both phenotype studies confirmed internal or partial external validation, supporting emerging translational robustness. Conclusions AI-driven clustering and predictive modeling collectively redefine septic shock as a continuum of biologically and clinically distinct endotypes. Integrative ML pipelines that combine clinical, metabolic, and transcriptomic data enable precise early warning, mechanistic biomarker discovery, and personalized resuscitation strategies. While pooled model discrimination remains high, substantial heterogeneity underscores the need for standardized pipelines, transparent calibration reporting, and multicenter real-time validation to transition ML-based sepsis phenotyping from computational insight to precision critical-care practice. This abstract is funded by: None