Interpretable Machine Learning to Optimize Early In-Hospital Mortality Prediction for Elderly Patients with Sepsis: A...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Geriatric Medicine

Background:

• Sepsis poses a significant mortality risk for elderly patients, particularly in intensive care units (ICUs).
• Early prognosis prediction is crucial for timely and effective sepsis treatment.
• Existing machine learning models for sepsis prediction have limited focus on the elderly population.

Purpose of the Study:

• To develop and validate a machine learning model for early prediction of in-hospital mortality in elderly patients with sepsis.
• To identify key clinical variables influencing mortality risk in this demographic.
• To leverage explainable AI (SHAP) for model interpretability.

Main Methods:

• Utilized the Medical Information Mart for Intensive Care IV (MIMIC-IV) database for patient data extraction.
• Developed and compared multiple machine learning models, including XGBoost, LGBM, LR, RF, DT, and KNN.
• Evaluated model performance using Area Under the Receiver Operating Characteristic Curve (AUROC) and F1 score, with SHAP for feature analysis.

Main Results:

• The study included 18,522 elderly patients, with an in-hospital mortality rate of 15.4%.
• The extreme gradient boosting (XGBoost) model demonstrated superior performance with an AUROC of 0.871 and an F1 score of 0.547.
• Key predictors of mortality included age, PO2, RDW, SPO2, WBC, and urine output, as identified by feature importance analysis and SHAP.

Conclusions:

• The XGBoost model provides accurate early prediction of in-hospital mortality for elderly sepsis patients.
• The model's interpretability through SHAP analysis enhances understanding of individual risk factors.
• This predictive tool can assist clinicians in identifying high-risk patients, potentially leading to improved clinical decision support systems and patient outcomes.