Machine learning-based prediction of gout using polygenic risk scores and clinical variables: A Korean cohort study

Abstract

The prevalence of gout, a chronic metabolic disease, has recently increased. Polygenic risk scores (PRS) represent a useful tool for predicting patient outcomes of this condition. However, the clinical utility of PRS in disease prediction remains controversial. Using data from the Korean Genome and Epidemiology Study, machine learning (ML) models were developed to predict gout based on PRS and clinical variables such as uric acid, lifestyle habits, and metabolic syndrome (MetS) profiles. Five supervised learning algorithms were applied: logistic regression (a traditional statistical model often used in machine learning contexts), random forest (RF), decision tree (DT), extreme gradient boosting, and light gradient boosting. Among the models, the RF model incorporating PRS, age, sex, MetS, and uric acid levels achieved the highest area under the curve (0.7204, 95% CI = 0.7124-0.7284). Feature importance analysis highlighted uric acid levels as the most important predictor of gout, followed by PRS and age. Although PRS enhanced the predictive power of the ML models, its effect was modest, suggesting that traditional risk factors remain important for gout prediction. This study demonstrated that integrating genetic data with clinical variables improves gout prediction. Further research is necessary to optimize the utility of PRS in diverse populations.