An XGBoost-SHAP classifier to distinguish asymptomatic hyperuricemia from acute primary gouty arthritis

Asymptomatic hyperuricemia (AH) is an important precursor of acute primary gouty arthritis (AGA), but not all patients with AH will progress to acute primary gouty arthritis. Given the limitations of relying solely on uric acid levels to accurately identify AH and AGA, there is an urgent clinical need for an auxiliary identification tool that integrates routine laboratory indicators, possesses discriminative ability, and is easy to interpret. Therefore, this study aims to use machine learning techniques, combined with routine clinical data and laboratory test indicators, to construct and validate an auxiliary decision-making model for identifying patients with AH and AGA and providing clinical support. We will compare the overall performance of different algorithms and elucidate the key features influencing model decisions through SHAP interpretability analysis. This is a single-center retrospective observational study involving a total of 11,763 AH and AGA patients treated at the Third Affiliated Hospital of Guangzhou University of Traditional Chinese Medicine from 2015 to 2025. Cases were divided into a development cohort (2015–2023, n = 9,584) and a temporal validation cohort (2024–2025, n = 2,179), with the development cohort further split into a training set (80%) and a test set (20%). Baseline analysis, collinearity analysis, univariate analysis, multivariate analysis, and univariate ROC analysis were conducted in the training set. Feature selection was performed using LASSO regression and XGBoost-RFECV, followed by multivariate analysis to streamline variables, ultimately incorporating 16 core variables into the model. A total of nine machine learning models were constructed, and their performance was evaluated based on the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, Kappa value, calibration curves, and decision curve analysis. SHAP methodology was utilized for interpretability analysis of the models. After screening 34 indicators, 16 variables were incorporated into the model, including monocyte count (MONO), uric acid (UA), age (Age), coefficient of variation of red cell distribution width (RDWCV), lymphocyte/monocyte ratio (LYMPH/MONO), neutrophil count (NEUT), basophil count (BASO), creatinine (Crea), hematocrit (HCT), neutrophil/lymphocyte ratio (NLR), red blood cell count (RBC), sex, percentage of eosinophils (EOrate), standard deviation of red cell distribution width (RDWSD), lymphocyte count (LYMPH), and percentage of monocytes (MONOrate). Comparative results indicated that the XGBoost model exhibited the best performance. The AUC of the validation set was 0.867, with an accuracy of 0.787, sensitivity of 0.775, specificity of 0.807, and average precision (AP) of 0.918. The AUC of the internal test set was 0.876, with an accuracy of 0.797, sensitivity of 0.763, and specificity of 0.851. In the temporal validation series, the XGBoost model achieved an AUC of 0.823, an accuracy of 0.720, a sensitivity of 0.634, and a specificity of 0.847, demonstrating a degree of generalizability. The XGBoost machine learning model constructed based on routine laboratory indicators (such as MONO, UA, age, RDW-CV, lymphocyte-to-monocyte ratio, NEUT, BASO, creatinine, HCT, NLR, red blood cell count, sex, EO rate, RDW-SD, lymphocyte count, and MONO rate) demonstrated moderate discriminative ability and stability in identifying patients with acute primary gouty arthritis versus asymptomatic hyperuricemia, providing an important basis for clinical differential diagnosis and decision-making, and holding significant clinical value.