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

Area of Science:

• Medical Informatics
• Public Health
• Computational Biology

Background:

• Chronic obstructive pulmonary disease (COPD) is a leading global cause of death, impacting over 213 million people.
• A significant gap exists in COPD diagnosis, with an estimated 70% of cases remaining undiagnosed, primarily due to underutilization of spirometry in primary care.
• This highlights a critical need for improved diagnostic strategies in primary care settings.

Purpose of the Study:

• To develop and evaluate machine learning models for accurate COPD risk classification.
• To identify key risk factors and their relationships with COPD using explainable AI.
• To assess the clinical utility of a machine learning-based approach for risk-stratified spirometry referral in primary care.

Main Methods:

• Utilized nationally representative data (N=8061) from the National Health and Nutrition Examination Survey (2007-2012) for participants aged 40 and above.
• Employed Least Absolute Shrinkage and Selection Operator (LASSO) regression with cross-validation to select 16 predictive features from 33 candidates.
• Compared twelve machine learning algorithms, with CatBoost showing optimal performance, and used SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• The CatBoost model achieved an Area Under the Curve (AUC) of 0.788, with a sensitivity of 0.706 and specificity of 0.728 on the validation set.
• SHAP analysis identified age and smoking duration as primary risk factors, revealing complex non-linear associations, including a U-shaped relationship for body mass index and synergistic interactions between age and smoking.
• Decision Curve Analysis confirmed the model's net clinical benefit across a wide range of risk thresholds (0.1-0.6).

Conclusions:

• A robust machine learning framework was established for COPD risk classification, integrating feature selection, algorithm comparison, and explainable AI.
• The developed model demonstrates potential for primary care by achieving a clinically relevant balance between sensitivity and specificity, facilitating risk-stratified spirometry referrals.
• External validation in diverse populations is the crucial next step for clinical implementation and broader applicability.