Interpretable mortality prediction model for ICU patients with pneumonia: using shapley additive explanation method

Area of Science:

• Medical Informatics
• Critical Care Medicine
• Machine Learning in Healthcare

Background:

• Pneumonia is a major global cause of illness and death, often requiring Intensive Care Unit (ICU) admission.
• Predicting pneumonia mortality is vital for personalized care, but current models lack clinical interpretability.
• This limits the adoption and utility of existing predictive tools in practice.

Purpose of the Study:

• To develop an interpretable model for predicting pneumonia mortality in ICU patients.
• To utilize Shapley Additive Explanation (SHAP) to understand an Extreme Gradient Boosting (XGBoost) model.
• To identify key prognostic factors influencing pneumonia mortality.

Main Methods:

• Retrospective cohort study using eICU-CRD electronic health records (2014-2015).
• Analysis focused on the first 24 hours of adult pneumonia ICU admissions.
• An XGBoost model was trained (70%) and validated (30%), with performance measured by AUC; SHAP explained model predictions.

Main Results:

• The study included 10,962 pneumonia patients, with an in-hospital mortality rate of 16.33%.
• The XGBoost model achieved a superior AUC of 0.778 ± 0.016, outperforming traditional scores by 10%.
• SHAP analysis revealed Aspartate Aminotransferase (AST) as the most significant predictor of mortality.

Conclusions:

• Interpretable models improve transparency and accuracy in assessing pneumonia mortality risk in ICUs.
• Aspartate Aminotransferase (AST) is identified as the primary prognostic factor, followed by age and albumin.
• These findings support enhanced clinical decision-making and optimized resource allocation for pneumonia patients.