A Joint Ensemble Framework for the Detection of Acute Exacerbations in Chronic Obstructive Pulmonary Disease

Abstract

Abstract Objective The purpose of this study was to create a joint ensemble framework for identifying AECOPD and providing a plausible explanation of model predictions. Methods From MIMIC-III, we extracted and organized records for COPD and AECOPD patients. Furthermore, we integrated missing value imputation, joint feature selection, advanced ML algorithms, Bayesian optimization techniques, and the SHAP interpretable method to construct a joint optimized ensemble framework, serving as the predictive model for AECOPD risk identification. The efficacy of the model's prediction was evaluated using a composite score of six evaluation measures. Results CAD and 19 other variables significantly impacted AECOPD. Various resampling methods and classifiers yielded diverse prediction accuracies. LightGBM and LR models with NC processing showcased optimal combined performance pre-heterogeneous combination. The Voting ensemble with MWMOTE achieved superior balanced classification. Conclusion The joint ensemble framework improved AECOPD risk identification performance in clinically relevant data of COPD patients admitted in the ICU.