JMIR Publications

ABSTRACT

Background:

Sepsis is a major global cause of morbidity and mortality, and early, accurate severity assessment is essential to improve outcomes among older adults. Traditional diagnostic tools often fail to capture the syndrome’s complexity, while machine learning (ML) can exploit routinely collected clinical and laboratory data for more precise classification.

Objective:

This study aims to develop and validate an explainable machine learning pipeline for early detection and severity stratification of sepsis in geriatric emergency care using routinely available admission laboratory and clinical data.

Methods:

The analysis was conducted on a large Italian hospital dataset including patients with and without sepsis and six ML algorithms using nested cross-validation were compared. Recursive Feature Elimination was applied for feature selection, and model interpretability was examined through SHAP values.

Results:

Ensemble methods, particularly XGBoost and Random Forest, showed the best performance in distinguishing sepsis from non-sepsis cases and provided solid severity stratification using only 10–20 features with an accuracy of 76.7%. Key predictors involved hepatic, renal, and immune markers, with Random Forest uniquely identifying absolute neutrophil count and total bilirubin as strong indicators of severe sepsis

Conclusions:

These findings highlight the potential of tree-based ML models for clinically interpretable, real-time sepsis risk stratification. Clinical Trial: Data for this study were collected as part of the ReportAGE project at IRCCS INRCA (Ancona, Italy), approved by the IRCCS INRCA Ethics Committee (reference CEINRCA-20008) and registered on ClinicalTrials.gov (reference NCT04348396).