Dự báo nhập đơn vị hồi sức tích cực ở bệnh nhân viêm phổi cộng đồng nhập viện: Mô hình eXtreme gradient boosting

Abstract

Objective: To evaluate the performance of the machine learning model eXtreme Gradient Boosting (XGBoost) in predicting the need for Intensive Care Unit admission in hospitalized patients with Community-Acquired Pneumonia. Subject and method: A retrospective, descriptive study conducted on Community-Acquired Pneumonia patients ≥ 18 years old in the Emergency Department, 108 Military Central Hospital, from 2021 to 2022. Clinical and paraclinical data were collected within the first 24 hours of admission. The XGBoost model was trained on 70% of the data and tested on the remaining 30%. The final model utilized 6 predictors with the highest Gain scores. Result: A total of the 350 patients studied, 75 (21.4%) required ICU admission. The six most important preditors (in descending order of importance) were: impaired mental status, elevated blood urea, thrombocytopenia, tachypnea, multilobar infiltrates on X-ray, and neutrophilia. The XGBoost model had a higher predictive performance compared to CURB-65 score, with AUCs of 0.876 and 0.826, respectively. Decision Curve Analysis confirmed that the XGBoost model provided maximum and stable net benefit over a wide range of threshold probabilities (20% to 80%). Conclusion: The XGBoost model, using 6 key clinical and paraclinical variables, significantly outperformed the CURB-65 score in predicting the need for ICU admission in hospitalized CAP patients. This simple model is a good tool to stratify the risk among patients with community-acquired pneumonia presenting to the Emergency Department.