An interpretable ensemble learning framework for COPD classification using population-based clinical and laboratory...

Chronic obstructive pulmonary disease (COPD) represents the third leading cause of death globally, affecting 213.39 million people, yet approximately 70% of cases remain undiagnosed due to insufficient spirometry utilization in primary care settings.

We developed machine learning-based COPD classification models using nationally representative data from 8061 participants aged ≥40 years in the National Health and Nutrition Examination Survey 2007-2012. Least Absolute Shrinkage and Selection Operator (LASSO) regression with repeated ten-fold cross-validation systematically selected sixteen features from 33 candidate variables. We compared twelve machine learning algorithms. SHapley Additive exPlanations (SHAP) analysis provided model interpretability.

CatBoost demonstrated optimal performance on the validation set with an area under the receiver operating characteristic curve (AUC) of 0.788, a sensitivity of 0.706, and a specificity of 0.728. SHAP analysis identified age and smoking duration as dominant risk factors, revealing non-linear relationships, including body mass index's U-shaped association with COPD risk and synergistic age-smoking interactions. Decision Curve Analysis confirmed net clinical benefit across risk thresholds of 0.1 to 0.6.

This study establishes a machine learning framework combining systematic feature selection, comprehensive algorithm evaluation, and explainable artificial intelligence for COPD risk classification using readily available clinical data. The model achieved a clinically appropriate sensitivity-specificity balance through threshold optimization, demonstrating potential for risk-stratified spirometry referral in primary care settings. External validation in independent populations represents the essential next step before clinical implementation.