Machine learning prediction of sepsis in paralytic ileus using interpretable clinical models

Area of Science:

• Critical Care Medicine
• Infectious Diseases
• Data Science in Healthcare

Background:

• Paralytic ileus (PI) is a frequent complication in critically ill patients.
• Sepsis development in PI patients significantly worsens outcomes.
• Early sepsis risk identification in PI patients is a clinical challenge.

Purpose of the Study:

• To develop and validate an early predictive model for new-onset sepsis in adult PI patients.
• To identify key clinical predictors for sepsis risk in PI.
• To support timely risk stratification and preventive interventions.

Main Methods:

• Retrospective analysis of adult PI patients from MIMIC-IV and Guangxi Medical University Cancer Hospital databases.
• Feature selection using LASSO regression and Boruta algorithm.
• Development and evaluation of seven machine learning models, with logistic regression selected.
• Performance assessment using AUC, calibration, and DCA; interpretability via SHAP.

Main Results:

• A total of 579 PI patients were analyzed; 29.78% developed new-onset sepsis.
• Seven predictors identified: pneumonia, RDW, heart failure, BUN, atrial fibrillation, serum chloride, and WBC count.
• The logistic regression model achieved AUCs of 0.687 (internal) and 0.715 (external) with good calibration and clinical utility.
• Pneumonia and RDW were identified as the most influential predictors by SHAP analysis.

Conclusions:

• An interpretable logistic regression model using accessible clinical variables can effectively predict sepsis in PI patients.
• The model facilitates early risk stratification for prompt preventive measures.
• This approach aids in improving prognosis for critically ill patients with PI.