Background/Purpose: Long-term hyperuricemia is essential for the development of gout. However, only 9.9% of individuals with asymptomatic hyperuricemia (AHU) develop gout within 10 years. Identifying high-risk individuals with AHU is crucial for implementing timely and targeted interventions to prevent or delay the onset of gout. We aimed to create a machine learning-based risk stratification model to predict the onset of gout in individuals with AHU.

Methods: We conducted a retrospective cohort study utilizing a Common Data Model (CDM) database from a single academic medical center. Subjects (age ≥ 18 years) were identified if they had at least one serum urate level ≥ 7 mg/dL or a diagnosis of AHU (concept ID 4147761) between 2010 and 2022. The index date was defined as the earliest report of a high serum urate level or the first diagnosis of AHU. Subjects were excluded if they had a prior diagnosis of gout or received a prescription for colchicine, allopurinol, febuxostat, or benzbromarone within 365 days before the index date. The outcome measured was the development of gout between 90 and 3650 days after the index date. Machine learning models, including LightGBM, XGBoost, Random Forest, AdaBoost, Decision Tree, and Logistic Regression, were employed to predict the development of gout. Model performance was assessed using the area under the receiver operating characteristic curve (ROC-AUC) and the precision-recall AUC (PR-AUC). We evaluated the contribution of each variable to the model’s predictions using feature importance and SHAP values. The Cox proportional hazard model was applied to construct a risk-scoring model with key predictors.

Results: Among 8,139 subjects with AHU, 223 (2.7%) developed gout during the 10-year follow-up period. Eighty-three percent of the patients were male, most between 30 and 40 years of age. LightGBM demonstrated the best performance among the models, achieving an ROC-AUC of 0.812 and a PR-AUC of 0.112. Among the top 15 features identified by feature importance and SHAP values (figure), serum urate level, LDL cholesterol, total carbon dioxide (tCO2), ESR, age, chronic kidney disease, and hypertension were identified as key predictors. In the final risk-scoring model, chronic kidney disease was assigned 24 points, hypertension 4, age 4 to 5, low tCO2 7, high LDL cholesterol 5, high ESR 3, and serum urate level 1 to 5 points (table). The C-index of the model was 0.730.

Conclusion: We developed a machine learning-based risk-scoring model using real-world data to predict the onset of gout in individuals with AHU. This model showed promising results based on routinely collected clinical parameters that may help identify high-risk individuals who would benefit from rigorous lifestyle monitoring or early intervention.

Figure. Top 15 features ranked based on feature importance in the machine learning model

Table. Final risk-scoring model for predicting the onset of gout in individuals with asymptomatic hyperuricemia

« Back to ACR Convergence 2025

ACR Meeting Abstracts - https://acrabstracts.org/abstract/machine-learning-based-prediction-of-gout-onset-in-individuals-with-asymptomatic-hyperuricemia-using-a-common-data-model/