Personalized sepsis mortality prediction: An interpretable machine learning nomogram

TL;DR

A machine learning tool predicts sepsis mortality using seven clinical factors and offers interpretable results to help doctors identify high-risk patients early.

Contribution

A novel interpretable machine learning nomogram for sepsis mortality prediction with SHAP-based interpretability.

Findings

The nomogram achieved an AUC of 0.900 in the training cohort and 0.796 in the validation cohort.

Seven clinical parameters were identified as independent mortality predictors.

SHAP analysis confirmed the model's interpretability and clinical utility.

Abstract

•A machine learning nomogram predicts in-hospital mortality in sepsis.•Seven clinical parameters were identified as independent mortality predictors.•The nomogram showed excellent discrimination with an AUC of 0.900.•SHAP analysis offers interpretability to the machine learning model.•The tool facilitates early risk stratification of high-risk patients.

A machine learning nomogram predicts in-hospital mortality in sepsis.

Seven clinical parameters were identified as independent mortality predictors.

The nomogram showed excellent discrimination with an AUC of 0.900.

SHAP analysis offers interpretability to the machine learning model.

The tool facilitates early risk stratification of high-risk patients.

Sepsis remains a major cause of mortality in ICU patients, requiring accurate prognostic tools for optimal management. This study aimed to develop and validate an interpretable machine learning-based nomogram for predicting in-hospital mortality in sepsis patients to guide clinical decision-making.

This retrospective cohort study included 407 adult sepsis patients from ICU admissions between January 2019 and December 2024. Patients were randomly divided into training (n = 284) and validation (n = 123) cohorts. A LASSO-based multivariate logistic regression was applied to construct a predictive nomogram.

Seven independent predictors emerged for in-hospital mortality: age, abdominal infection, vasopressor requirement, WBC, BNP, APACHE II score, and mechanical ventilation support. The nomogram showed strong discrimination with an AUC of 0.900 in the training cohort and 0.796 in the validation cohort. Calibration curves, decision curve analysis, and SHAP interpretation confirmed good model performance and clinical utility.

The novel machine learning-based nomogram provides a practical, interpretable risk assessment tool for early mortality prediction, potentially improving patient outcomes through enhanced clinical understanding and timely interventions in critically ill sepsis patients.

Figures