An explainable artificial intelligence framework for risk prediction of COPD in smokers

Area of Science:

• Pulmonary Medicine
• Artificial Intelligence
• Data Science

Background:

• Early Chronic Obstructive Pulmonary Disease (COPD) signs are often missed, delaying crucial prevention and treatment.
• This limits opportunities for timely intervention in at-risk individuals.

Purpose of the Study:

• To develop an explainable artificial intelligence (AI) framework for identifying high-risk COPD individuals within the smoking population.
• To provide interpretable insights into the factors driving COPD risk prediction.

Main Methods:

• Data preprocessing techniques (FAMD, Boruta, NRSBoundary-SMOTE) addressed missing data, high dimensionality, and class imbalance.
• Seven machine learning models were evaluated, with CatBoost demonstrating superior performance on balanced datasets.
• Model interpretability was achieved using Shapley Additive Explanations (SHAP) and Partial Dependence Plots (PDP).

Main Results:

• Age and 14 other variables, including CAT score, BMI, and environmental factors, were significant COPD predictors in smokers.
• CatBoost, random forest, and logistic regression models showed good performance on unbalanced data.
• CatBoost combined with NRSBoundary-SMOTE achieved the best classification performance based on AUC, F1-score, and G-mean.

Conclusions:

• The study successfully integrated feature screening, unbalanced data handling, and advanced ML for early COPD risk identification in smokers.
• Identified COPD risk factors provide a basis for targeted screening and self-management strategies in the smoking population.