Predicting Mortality in Intensive Care Unit Patients With Allergic Bronchopulmonary Aspergillosis (ABPA) Using an Interpretable Machine Learning Model: A Retrospective Cohort Study.

Abstract

Background: Allergic bronchopulmonary aspergillosis (ABPA) is a hypersensitivity lung disease caused by Aspergillus infection, with severe cases often requiring admission to the intensive care unit (ICU). Early prediction of in-hospital mortality in ICU ABPA patients is crucial for optimizing clinical decision-making and resource allocation.

Methods: This retrospective study collected clinical data from ICU patients diagnosed with ABPA at Yuebei People's Hospital between January 2020 and July 2024. An in-hospital mortality prediction model was developed using an explainable XGBoost machine learning algorithm. SHapley Additive Explanations (SHAP) was employed to interpret key predictive factors, and internal validation was conducted to assess model performance.

Results: A total of 82 ICU ABPA patients were included, with mortality rates of 46.3% (26/57) in the training set and 48% (12/25) in the validation set. The XGBoost model demonstrated excellent predictive performance, achieving areas under the receiver operating characteristic (ROC) curve (AUC) of 0.995 (95% CI: 0.903-1.000) in the training set and 0.881 (95% CI: 0.846-0.909) in the validation set. SHAP analysis identified key predictors of mortality, including BMI, peak procalcitonin level, peak eosinophil count, age, asthma history, peak leukocyte count, and lowest platelet count.

Conclusion: The XGBoost model effectively predicts in-hospital mortality in ICU ABPA patients and provides interpretable results using SHAP analysis. Although the model performed well in internal validation, external validation is needed to enhance its generalizability. Future multicenter studies and integration of dynamic biomarkers are recommended to optimize predictive accuracy and support individualized clinical decision-making.