Determination of factors to predict hyperuricemia in admitted patients using machine learning-based statistical model

Abstract

While hyperuricemia is known to indicate a poor prognosis in hospitalized patients, factors determining its onset during hospitalization remain unclear. In this study, we used a machine learning-based statistical model to identify and combine factors predicting hyperuricemia. We enrolled 595 patients hospitalized at NHO Yonago Medical Center. Hyperuricemia (serum uric acid level >7.0 mg/dL) was the dependent variable, while age, sex, emaciation, medical history, medication, and biochemical data were independent variables. We used the machine learning-based statistical model to classify predictive factors and combined them using a decision tree based on the XGBoost model. Compared with normouricemic patients, hyperuricemic patients had significantly higher BMI and higher incidences of obesity, hypertension, diabetes, and treatment with ST-combo, diuretics, beta-blockers, and ACE inhibitors. They also had significantly lower serum K levels and eGFR. The XGBoost model showed a higher AUC value compared with logistic analysis in predicting hyperuricemia. SHAP values indicated that the main predictive factors were: creatinine, diuretics, eGFR, BUN, serum K level, age, AST, and ALT. The decision tree revealed that in patients not on diuretics, hyperuricemia could be predicted by either a combination of higher serum K levels and younger age or lower serum K levels and higher BUN levels. In patients on diuretics, a combination of higher creatinine levels and lower ALT levels was predictive of hyperuricemia. In conclusion, the XGBoost model effectively identified and combined factors to predict hyperuricemia in hospitalized patients.