# Machine Learning Models for Sepsis: From Early Detection to Short- and Long-Term Prognosis

## TL;DR

This study uses machine learning to improve sepsis detection and predict outcomes better than traditional methods.

## Contribution

The study introduces ML models that integrate clinical and biochemical data for improved sepsis prognosis.

## Key findings

- ML models achieved high AUC scores for predicting sepsis and septic shock.
- SHAP analysis confirmed the clinical relevance of key predictors like biomarkers and clinical scores.
- ML models outperformed conventional methods in predicting mortality and ICU admission.

## Abstract

Sepsis is a leading cause of morbidity and mortality worldwide, and its outcomes depend on early recognition and timely intervention. Conventional clinical scores and biomarkers provide prognostic value but often lack accuracy for individualized prediction. Machine learning (ML) offers the ability to integrate multidimensional data to improve risk stratification. We analyzed 477 patients admitted to our hospital, including 251 with sepsis, 100 with septic shock, and 126 controls. Demographic, clinical, and laboratory data were collected. Univariate correlation analyses explored associations with sepsis severity and mortality (in-hospital, 30-day, and 90-day). Several ML models were tested, with performance assessed by area under the receiver operating characteristic curve (AUC-ROC) and Matthews’s correlation coefficient (MCC). Model interpretability was evaluated using SHAP (SHapley Additive exPlanations). Sepsis severity and mortality correlated with biomarkers (procalcitonin, mid-regional pro-adrenomedullin, lactate) and clinical scores (SOFA, qSOFA). In-hospital mortality was associated with ADM, catecholamine use, and SOFA, while 90-day mortality involved smoking and Gram-negative or polymicrobial infections. Different machine learning models were evaluated, and the model achieving the highest performance on the validation set was selected. The selected model either outperformed or demonstrated comparable performance to logistic regression, depending on the specific prediction task (AUC 0.99 for sepsis, 0.96 for septic shock, 0.70 for ICU admission; 0.90, 0.72, and 0.87 for in-hospital, 30-day, and 90-day mortality). SHAP confirmed the clinical relevance of these predictors. ML models integrating clinical and biochemical data outperform conventional methods in predicting sepsis progression and mortality, while maintaining interpretability. These findings support the use of ML-based tools for early diagnosis and personalized risk stratification in sepsis, though external validation is required before clinical application.