Exploring the relationship between per- and polyfluoroalkyl substances exposure and rheumatoid arthritis risk using...

Area of Science:

• Environmental Health
• Toxicology
• Computational Biology

Background:

• Rheumatoid arthritis (RA) is a chronic autoimmune disease.
• Environmental exposures, such as per- and polyfluoroalkyl substances (PFAS), influence RA risk.
• Previous research suggests a link between PFAS and RA, but lacks interpretable predictive models.

Purpose of the Study:

• To develop and validate an interpretable machine learning model for predicting rheumatoid arthritis risk based on PFAS exposure.
• To identify key PFAS predictors and understand their non-linear associations with RA risk.
• To create a publicly accessible tool for assessing individual RA risk.

Main Methods:

• Analysis of 11,705 participants from the National Health and Nutrition Examination Survey (2003-2018).
• Evaluation of twelve machine learning algorithms using AUC, accuracy, sensitivity, specificity, and F1 score.
• Identification of key predictors using SHapley Additive exPlanations (SHAP) and examination of non-linear relationships with partial dependence plots and LOWESS curves.

Main Results:

• The CatBoost model demonstrated superior performance (AUC: 0.82, Accuracy: 74%, F1 score: 0.62).
• Perfluorooctane sulfonic acid (PFOS) and MPAH were identified as significant predictors.
• PFOS showed increased RA risk above 15.10 ng/ml (U-shaped), and MPAH at 0.22 ng/ml, with non-linear exposure-response relationships.

Conclusions:

• Machine learning effectively predicts rheumatoid arthritis risk associated with PFAS exposure.
• Identified non-linear exposure-response patterns offer insights into environmental factors contributing to RA.
• The developed web-based risk calculator provides a practical tool for public health and clinical applications.