# Uncovering Nonlinear Predictors of Serum Biomarker Uric Acid Using Interpretable Machine Learning in Healthy Men

## TL;DR

This study uses interpretable machine learning to uncover nonlinear relationships between uric acid and health factors in healthy men, revealing new insights into metabolic thresholds.

## Contribution

The study introduces interpretable machine learning to identify nonlinear and threshold-based predictors of uric acid levels that traditional methods miss.

## Key findings

- Waist-to-hip ratio influences uric acid only below a threshold of 0.969.
- Creatinine's effect on uric acid becomes significant above 0.97 mg/dL, indicating a renal threshold.
- Betel nut exposure shows a complex, non-binary association with uric acid metabolism.

## Abstract

Background: Uric acid (UA) is linked to gout, renal dysfunction, and cardiovascular disease. Prior studies often assume linear relationships, potentially oversimplifying physiological complexity. Methods: We analyzed data from 5200 healthy Taiwanese men. Demographic, biochemical, lifestyle, and inflammatory variables were assessed using Pearson correlation, multiple linear regression (MLR), and multivariate adaptive regression splines (MARS), an interpretable machine learning method for detecting nonlinear, threshold-based effects. Results: Pearson correlation showed broad linear associations, whereas MARS identified fewer but more physiologically meaningful predictors. Waist-to-hip ratio (WHR) had a strong threshold effect, influencing UA only below 0.969. Creatinine showed a nonlinear impact, becoming substantial above 0.97 mg/dL, suggesting a renal threshold within the “normal” range. Calcium and high-sensitivity C-reactive protein (hs-CRP) each displayed inflection points (9.5 mg/dL and 3.38 mg/L, respectively), indicating range-specific effects. Notably, betel nut exposure, nonsignificant in linear models, emerged in MARS as a predictor with a complex, non-binary association with UA metabolism. Predictive performance was comparable (RMSE: 1.6694 for MARS vs. 1.6666 for MLR), but MARS offered superior interpretability by highlighting localized nonlinear effects. Conclusions: MARS modeling revealed critical nonlinear, threshold-dependent associations between UA and WHR, creatinine, calcium, hs-CRP, and betel nut exposure, which were not captured by conventional methods. These findings underscore the value of interpretable machine learning in metabolic research and suggest precise thresholds for clinical risk stratification.