Machine Learning in the Prediction and Management of Sepsis: A Meta-Analysis

Abstract

Introduction:

Sepsis remains a leading cause of morbidity and mortality worldwide, requiring early identification for effective management. In recent years, machine learning (ML) has emerged as a tool for improving diagnostic and prognostic accuracy in sepsis care. This meta-analysis aimed to evaluate the performance of ML models in predicting sepsis diagnosis and related clinical outcomes.

Methods:

A systematic literature search was conducted using PubMed, ScienceDirect, and Google Scholar for studies published from 2013 to 2024. Original research articles employing machine learning algorithms to predict sepsis diagnosis, mortality, ICU admission, or length of stay were included. Exclusion criteria were non-human studies, reviews, and non-ML models. PRISMA guidelines were followed. Risk of bias was assessed using the PROBAST tool. Model performance was evaluated using metrics such as area under the curve (AUC), sensitivity, specificity, and accuracy.

Results

Out of approximately 350 screened articles, 20 met the inclusion criteria. The most frequently used models were tree-based algorithms such as Random Forest and XGBoost. Across studies, the pooled AUC values commonly exceeded 0.85, suggesting high diagnostic accuracy. ML models also showed consistent performance in predicting sepsis-related mortality and ICU admissions. Variability was observed due to differences in sepsis definitions, sample sizes, and model validation strategies.

Conclusion

Machine learning models demonstrate promising potential for early sepsis prediction and outcome forecasting, with high diagnostic accuracy across multiple algorithms. However, significant heterogeneity in study designs, model interpretability, and lack of real-world validation limits their current clinical utility.

Limitations: Variability in sepsis definitions, model heterogeneity, potential publication bias, and absence of standardized external validation protocols.

Ethical Statement:

As this meta-analysis used only publicly available data, institutional review board (IRB)approval was not required.