Abstract
PURPOSE: We assessed the risk of gout in the Taiwan Biobank population by applying various machine learning algorithms. The study aimed to identify crucial risk factors and evaluate the performance of different models in gout prediction. PATIENTS AND METHODS: This study analyzed data from 88,210 individuals in the Taiwan Biobank, identifying 19,338 cases of gout and 68,872 controls. After data cleaning and propensity score matching for gender and age, the final analytical sample comprised 38,676 individuals (19,338 gout cases and 19,338 controls). Five machine learning models were used: Bayesian Network (BN), Random Forest (RF), Gradient Boosting (GB), Logistic Regression (LR), and Neural Network (NN). The predictive performance was evaluated using a split dataset (80% training set and 20% test set). RESULTS: Variable importance analysis was performed to identify key variables, with uric acid and gender emerging as the most influential risk factors. Descriptive data highlighted significant differences between the control group and gout patients, with a higher prevalence of gout in men (51.36% vs 48.64%). Both the RF and GB demonstrated high performance across multiple metrics, with RF consistently achieving a high area under the curve (AUC) of 0.986 to 0.987, alongside excellent sensitivity (0.945-0.947) and specificity (0.998-0.999). GB also performed robustly, with AUC values around 0.987-0.988 and maintaining high sensitivity (0.944-0.950) and specificity (0.995-0.999) across different model variations. The F1 scores for both models (GB and RF) indicate strong predictive capabilities, with values around 0.971-0.972. CONCLUSION: The RF and GB demonstrated exceptional accuracy in predicting gout status, particularly when incorporating genetic data alongside clinical factors. These findings underscore the potential for integrating machine learning models with genetic information to enhance gout prediction accuracy in clinical practice.