Interpretive machine learning predicts short-term mortality risk in elderly sepsis patients

Area of Science:

• Geriatric Medicine
• Critical Care Medicine
• Data Science in Healthcare

Background:

• Sepsis is a significant cause of mortality in hospitalized patients, with a rising incidence in the elderly.
• Early identification and mortality risk prediction in elderly sepsis patients are critical for improving outcomes.

Purpose of the Study:

• To develop a machine learning model for predicting short-term mortality risk in elderly patients with severe sepsis.
• To create a clear, concise, and interpretable prediction tool for clinical use.

Main Methods:

• Utilized the MIMIC-IV database, randomly splitting data into training and validation sets (7:3 ratio).
• Employed Recursive Feature Elimination (RFE) to identify key mortality predictors from 49 variables.
• Built and evaluated six machine learning algorithms, including Extreme Gradient Boosting (XGBoost), and used SHAP and LIME for model interpretability.

Main Results:

• Analyzed 4,056 elderly sepsis patients, identifying eight key predictive variables using RFE.
• The XGBoost model demonstrated superior performance with an AUC of 0.88 and accuracy of 0.84 on the validation set.
• SHAP and LIME analyses provided global and detailed interpretations of the model's key variables and their predictive weights.

Conclusions:

• The developed machine learning model serves as a reliable tool for predicting the prognosis of elderly patients with severe sepsis.
• The model's interpretability enhances its potential for clinical adoption and decision support.