JMIR Publications

ABSTRACT

Background:

Early detection of sepsis is critical for reducing mortality and healthcare burden. Traditional tools have limited sensitivity, particularly in non-ICU settings. ML and DL offer data-driven alternatives for timely prediction across inpatient populations.

Objective:

This study aimed to assess the predictive accuracy of machine learning and deep learning models for sepsis onset in hospitalized patients and to investigate the real-world clinical burden of early sepsis onset using national claims data.

Methods:

This systematic review and meta-analysis included 34 studies evaluating ML/DL models show strong predictive performance; however, clinical utility remains uncertain due to limited prospective validation. Their integration may support earlier detection and improved outcomes, pending real-world implementation. We also conducted real-world validation using Korean NIS claims data to assess clinical burden based on onset timing.

Results:

The pooled AUROC was 0.895 (95% CI: 0.873–0.916), suggesting high predictive accuracy. Subgroup analyses showed better performance in studies using hospital-specific data and larger sample sizes. Early-onset sepsis (within 1 month of admission) was associated with significantly longer hospital stays and higher medical costs. ML/DL models demonstrate promising performance in retrospective settings; however, their clinical utility remains uncertain due to limited prospective and external validation.

Conclusions:

ML/DL models offer strong performance in predicting sepsis onset across diverse hospital settings. Their early integration into clinical systems may support timely intervention and reduce healthcare burden. Further validation in prospective settings is warranted. Clinical Trial: PROSPERO CRD420251005274.