JMIR Publications

ABSTRACT

Background:

Hyperuricemia affects 20%–25% of adult men and 5%–10% of adult women globally, is a major risk factor for gout, cardiovascular disease, and chronic kidney disease, and remains substantially underdiagnosed. Serum uric acid testing requires venipuncture and is not universally included in routine health examinations, while urine dipstick analysis is performed broadly in primary care and population screening programs. Whether the multivariate pattern of routine urine dipstick parameters can substitute blood-based testing for opportunistic hyperuricemia detection has not been established.

Objective:

This study aimed to develop and externally validate calibrated machine learning models that predict hyperuricemia at three clinically relevant cut-offs using routinely collected urine dipstick parameters with age, sex, and body mass index (BMI).

Methods:

Electronic medical records from a tertiary hospital in the Seoul metropolitan area, Korea (2010–2025; 145,678 paired urine–blood records within ±14 days having complete urine dipstick, demographic, and anthropometric data) were used to train calibrated soft-voting ensembles (LightGBM, XGBoost, CatBoost) for three clinically relevant targets: hyperuricemia (unified ≥7.0 mg/dL, the institutional clinical reporting standard), markedly elevated hyperuricemia (≥9.0 mg/dL, associated with substantially increased gout incidence), and an institutional urate-lowering therapy (ULT) initiation threshold (≥10.0 mg/dL, corresponding to the development institution’s prescription practice). Thirteen input features comprised 10 urine dipstick parameters, age, sex, and BMI. Hyperparameters were tuned by Bayesian optimization; probabilities were calibrated by isotonic regression with 5-fold cross-validation. External validation used the Korea National Health and Nutrition Examination Survey (KNHANES) 2016–2024 (N=54,660).

Results:

On external validation, the unified ≥7.0 mg/dL model achieved an area under the receiver operating characteristic curve (AUROC) of 0.850 (95% CI 0.845–0.854), the markedly elevated ≥9.0 mg/dL model achieved 0.866 (95% CI 0.853–0.878), and the institutional ULT threshold ≥10.0 mg/dL model achieved 0.865 (95% CI 0.835–0.889). Discriminative performance was highest among younger adults and attenuated with advancing age. Shapley additive explanations (SHAP) analysis identified sex, age, BMI, urine pH, and urine protein as the most influential predictors.

Conclusions:

Calibrated machine learning models using routinely collected urine dipstick parameters and basic clinical information achieved good discriminative performance for hyperuricemia detection in a nationally representative cohort, supporting use as a low-cost screening triage instrument for opportunistic identification of high-risk individuals across three clinically anchored thresholds without requiring blood-based inputs.