Artificial intelligence tools in sepsis prediction: a systematic review and meta-analysis

Abstract

Abstract Background Machine learning (ML) models have been used to improve diagnosis of sepsis, but their effectiveness is uncertain. We aimed to systematically evaluate the diagnostic accuracy of ML models for sepsis prediction through a systematic review and network meta-analysis (NMA) and identify key determinants of their accuracy. Methods We searched five major databases for studies published up to June 2025 that compared a ML model against a comparator for sepsis prediction in adults. Data were synthesized using a random effects model for the primary meta-analysis, a network meta-analysis to compare model architectures and a multi-variable meta-regression to explore heterogeneity. The PROBAST-AI tool was used to assess risk of bias. Results Fifty-three studies encompassing over 7 million patient encounters were included. The pooled area under the receiver operator curve for the best-performing ML models was 0.88 (95% CI [0.86, 0.90]), significantly outperforming traditional comparators (mean difference 0.119, p < 0.001). The pooled sensitivity and specificity were 77.2% and 84.7%, respectively. Decision tree and ensemble models showed highest performance on network meta-analysis. Model type was the only covariate to significantly influence performance in the meta-regression. However, heterogeneity was substantial (I²>95%), the 95% prediction interval was wide (-0.06 to 0.30), and most studies were judged to have a high risk of bias. Conclusion ML models demonstrate statistically superior performance for sepsis prediction, with decision trees showing highest performance. Extreme heterogeneity, wide prediction intervals, and a high risk of bias in existing studies are critical barriers to translation. Future research must prioritise standardised validation and prospective trials to establish the real-world impact of these algorithms.