Developing an interpretable machine learning predictive model of chronic obstructive pulmonary disease by serum PFAS...

Area of Science:

• Environmental Health
• Toxicology
• Data Science

Background:

• Chronic obstructive pulmonary disease (COPD) poses a significant global health burden with limited early detection methods.
• Research on the link between per- and polyfluoroalkyl substances (PFAS) and COPD exists, but few studies utilize interpretable machine learning (ML).

Purpose of the Study:

• To investigate the association between exposure to various PFAS and the risk of developing COPD.
• To develop and validate an interpretable ML model for predicting COPD risk based on PFAS exposure levels.

Main Methods:

• Utilized data from 4,450 National Health and Nutrition Examination Survey (NHANES) participants (2013-2018).
• Applied data cleaning, K-nearest neighbors (KNN) imputation, and developed nine ML models, including CatBoost.
• Evaluated models using accuracy, AUC, sensitivity, and specificity; further analyzed the best model with PDP and SHAP.
• Deployed the final model as a public web-based risk calculator.

Main Results:

• The CatBoost model achieved 84% accuracy, 0.89 AUC, 81% sensitivity, and 84% specificity.
• Higher levels of perfluorooctane sulfonic acid (PFOS) and perfluoroundecanoic acid (PFUA) were linked to decreased COPD risk.
• Perfluorooctanoic acid (PFOA) and 2-(N-Methyl-perfluorooctane sulfonamido) acetic acid (MPAH) showed positive associations with increased COPD risk.
• Other PFAS like PFNA, PFDE, and PFHxS exhibited mixed or non-linear effects.

Conclusions:

• PFOS and PFUA act as protective factors against COPD, whereas PFOA and MPAH elevate COPD risk.
• Interpretable ML, particularly CatBoost, effectively clarifies complex PFAS-COPD relationships.
• Findings support the need for stricter PFAS regulations and underscore ML's role in public health prevention strategies.