Machine learning-enabled risk prediction of chronic obstructive pulmonary disease with unbalanced data

Area of Science:

• Medical informatics
• Public health surveillance
• Computational epidemiology

Background:

• Early symptoms of chronic obstructive pulmonary disease (COPD) are often subtle, hindering timely diagnosis and intervention.
• Effective early identification of at-risk individuals is crucial for improving COPD prevention and treatment outcomes.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting COPD risk.
• To enhance the efficiency of early COPD identification through data-driven approaches.

Main Methods:

• Utilized data from 5807 individuals in the 2019 Shanxi Province COPD Surveillance Program.
• Applied feature selection techniques (Generalized elastic net, Lasso, Adaptive lasso) to identify key predictive variables.
• Employed supervised classifiers with cost-sensitive learning and SMOTE for imbalanced data, including Logistic Regression, SVM, Random Forest, XGBoost, LightGBM, NGBoost, and Stacking.

Main Results:

• Ten significant variables were identified, including frequent cough before age 14.
• The Stacking ensemble model demonstrated good performance on unbalanced datasets.
• Logistic Regression with class weighting achieved the best classification performance on balanced data, with F1-Scores and G-means of 0.290 and 0.660, respectively.

Conclusions:

• Machine learning models, particularly those utilizing survey questionnaire data, can automate the identification of individuals at risk of COPD.
• These models offer a simple, scientific tool to aid in the early detection of COPD.
• Integrating feature selection and unbalanced data processing methods enhances the predictive power of ML models for COPD surveillance.