Risk factor analysis and predictive model construction and evaluation of fractional uric acid excretion in Chinese inpatients with type 2 diabetes Authors: Lin Zhuo, Xiantong Zou, Wenjia Yang, Ling Chen, Linong Ji, Lingli Zhou, Xueyao Han Published: 2026 ABSTRACT AbstractFractional excretion of uric acid (FEUA) is currently acknowledged as a new risk factor of nephropathy. Precise measurement of FEUA requires withdraw medication during the washout time and 24 h urine sample collecting, not always convenient, fostering construction of predictive models. No existing model could portrait FEUA. Our aim is to develop and evaluate a model which could predict FEUA based on several easily available factors. 2301 patients hospitalized in the Department of Endocrinology, Peking University People’s Hospital were chose. 1510 cases enrolled during 2018–2024 were used to construct predictive models using linear regression and to perform internal validation. 791 cases enrolled in 2017 and 2025 were used as an external validation. Sex, SUA, SCr, and FPG closely related with FEUA in internal dataset. The predictive formula was: Ln (FEUA) = 1.9408 + 0.1462*gender (male=1female = 2) − 0.0031*SUA(µmol/L) + 0.0112*SCr(µmol/L) + 0.0068*FPG (mmol/L); Predicted FEUA = Exp(Ln(FEUA))*1.0265; 1.0265 was the Duan’s smearing factor. Root mean squared errors (RMSE) between the observed and predicted FEUA was 1.690%. Spearman correlation coefficient was 0.7870 (P < 0.001). For predicting elevated FEUA (≥ 11%), the overall accuracy was 94.70%; ROC was 0.9552(95%CI: 0.9413 ~ 0.9691). Youden index (0.7814) at the cut-point (8.0257%) of FEUA was identified with a high sensitivity (0.9541) and specificity (0.8273). RMSE for 1-fold, 5-fold, 10-fold and leave-one-out cross-validation were 1.6898, 1.7087 ± 0.1716, 1.6964 ± 0.2761, and 1.7081, respectively. ROC values in subgroups of internal dataset were all higher than 0.949. In external dataset validation, Spearman correlation coefficient between observed and predicted FEUA was 0.7451(P < 0.001), ROC for elevated FEUA was 0.8946 (95% CI: 0.8545 ~ 0.9347). In validation on those excluded patients who were under medication, Spearman correlation coefficient between observed and predicted FEUA was 0.7899 (P < 0.001), ROC for elevated FEUA was 0.9029 (95% CI: 0.8817 ~ 0.9241). A stability of all indices was identified for subgroups of external dataset and those under medication. Our study identified an effective and practical FEUA predictive model in T2DM inpatients using linear regression with four common indexes: sex, SUA, SCr, FPG.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-68183-5.Keywords: Fractional excretion of uric acid, Predictive model, Risk factor, Regression, Type 2 diabetesSubject terms: Diseases, Endocrinology, Medical research, Nephrology, Risk factors Introduction A recent large multi-regional study in China reported diabetic kidney disease (DKD) prevalence is about 31% among patients with type 2 diabetes mellitus (T2DM)1. As the prevalence of diabetes worldwide reported to have reached epidemic proportions, up to 40% of people living with diabetes develop diabetic nephropathy (DN)2. DN is expected to continue to be the leading cause of end-stage kidney disease (ESKD) in both developed and developing countries. To reduce the incidence and progression of DN, current guidelines are now focused on developing strategies for early identification and comprehensive management of risk factors of DN. Besides achieving strict glycemic control, managing risk factors and find out more specific risk factors also play a vital part3. Recently, besides the traditional risk factors of kidney damage such as hypertension, dysglycemia and dyslipidemia4, clinical and epidemiological studies confirmed the positive relationship between fraction excretion of uric acid (FEUA) and chronic kidney disease (CKD) and estimated glomerular filtration rate (eGFR) decrease in different populations5–9. Two thirds of uric acid excretion is carried out by kidneys10. FEUA is currently acknowledged as a precise measurement of UAE, with the normal FEUA range of 5.5–11.0%11. Investigating closely associated variables of FEUA and developing a predictive model is driven by two urgent needs. On the one hand, accurate assessment of FEUA requires assessment of 24-h urine sample under a 2-week washout period and 5-day purine-restricted diet12 given the fact that many drugs or diet may interfere with renal urate excretion. However, withdrawal of medications during the washout period can be problematic. In addition, the inconvenience of 24-h urine collection and strict lifestyle control also limits its application in daily practice. On the other hand, there is a pressing need to extend the scope of research utilizing large-scale public databases. A simple and reliable method to calculate FEUA is needed, both for clinical practice and for research purposes, particularly when assessing patients with T2DM, who have dysregulated uric acid (UA) metabolism and are already prone to suffer from kidney damage because of hyperglycemia, hyperinsulinemia and unhealthy diet/lifestyle13. So far, no existing model could portrait FEUA. This study was designed to develop and evaluate a prediction model for FEUA in T2DM population based on several easily available factors. Predictive model for FEUA level was constructed using linear regression model14. Besides, we categorized FEUA based on predicted FEUA level and validated the accuracy of both predicted FEUA level and the categorization based on the prediction. Traditional regression models are proved to be as well as optimized machine algorithms when using a small number of well understood, strong predictor variables15. In the big data era, clinical predictive modeling based on data mining improves the diagnosis and monitoring of diseases16. Our study contributes to identifying individuals with elevated FEUA (≥ 11%) and would further benefit the control of comorbidities related to high FEUA16. Our study could also help extend the scope of research utilizing large-scale public databases. Methods Subjects This study involved 6 061 patients hospitalized in the Department of Endocrinology, Peking University People’s Hospital, between 1 st January, 2017 and 31 st December, 2025, who had a diagnosis T2DM according to the diagnostic criteria of T2DM by American Diabetes Association (ADA)17. Owing to the retrospective nature of the study and the use of de-identified data with no personal identifiers, the requirement for written informed consent was waived. Patients or the public WERE NOT involved in the design, or conduct, or reporting, or dissemination plans of our research. The exclusion criteria were (1) patients with malignant tumors, infection, hyper- or hypothyroidism, immune system diseases, blood system diseases, polycystic kidney diseases, acute or chronic kidney diseases and ureteral stenosis or after (partial) nephrectomy or renal transplant surgery or with solitary kidney; (2) patients with pregnancy; (3) people using medicine including: NSAIDs, female hormone, diuretics, immunosuppressors, uric acid lowering therapy, Fenofibrate, Angiotensin receptor blocker (ARB), Angiotensin-converting enzyme inhibitors (ACEI), Sodium-glucose contransporter-2 inhibitors (SGLT2 inhibitors); (4) patients with extreme values: SCr>260µmol/L or FPG>30mmol/L or SUA>700µmol/L; (5) age below 18 years old. Ultimately, 2 301 patients were included into the study. Participants enrolled during the year of 2018–2024 were used for training and internal validation and subsets validation. To test if our model is applicable in a wide range of time, and also for the purpose of facilitating FEUA investigation using some existing public or local dataset which didn’t come with FEUA-related urinary variables originally, we used cases enrolled during 2017 and 2025, periods outside of the coverage of model training dataset, as independent (external) validation18. We reported independent (external) validation using cases enrolled in 2025-only and 2017-only, as prospective-style temporal validation and retrospective-style validation, respectively. To validate the prediction of FEUA elevation based on predicted FEUA values in complementary to Evaluation of Predictive Accuracy and Agreement of linear prediction model, we combined 2017 retrospective validation dataset with the 2025 temporal validation dataset to ensure a robust evaluation with an adequate number of events. This combining strategy is also due to a decreasing trend in the prevalence of elevated FEUA in recent years (2024–2025). Besides, the participants under medication that were initially excluded were kept as another subgroup for sensitivity validation of our predictive model. The flowchart of this cross-sectional study was presented in Fig. 1. The study adhered to the principles of the Declaration of Helsinki and was approved by the Institutional Review Board of Peking University People’s Hospital (approval No. 2025PHB080-001). Owing to the retrospective nature of the study and the use of de-identified data with no personal identifiers, the requirement for written informed consent was waived by the Institutional Review Board of Peking University People’s Hospital. This study was reported in accordance with the TRIPOD statement by each item18. Flow chart of the selection and exclusion of the participant. Data collection Detailed data were extracted from the hospital information system (HIS). Demographic and clinical characteristics: age, sex, duration of T2DM, comorbidities (e.g., hypertension, coronary artery disease, diabetic retinopathy, and diabetic peripheral neuropathy), and in-hospital treatments were recorded by the attending physicians. Height and weight were measured to calculate body mass index (BMI). Systolic blood pressure (SBP) and diastolic blood pressure (DBP) were measured by a physician using a standard mercury sphygmomanometer after the subject had seated for at least 10 min. 24-hour urine sample was collected from the participants to assess 24 h-urinary uric acid excretion (UUAE), 24 h- urinary creatinine (uCr). Spot urine was collected to examine β2-microglobulin (β2MG), Urine Albumin-to-Creatinine Ratio (UACR). Venous blood samples were drawn after an overnight fasting, laboratory evaluations were all obtained, including: serum creatinine (SCr), serum uric acid (SUA), fasting plasma glucose (FPG), total cholesterol (TC), low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C), Triglyceride (TG), alanine aminotransferase (ALT), aspartate aminotransferase (AST), hemoglobin A1c (HbA1c). Generating and grouping new variables FEUA calculation: FEUA (%) =[urinary urate (µmol/L) × serum creatinine (µmol/L)]/[serum urate (µmol/L) × urinary creatinine (µmol/L)] × 10019. FEUA elevation was defined as FEUA > = 11%11. eGFR was calculated by using the new glomerular filtration rate estimating equation20: eGFR(mL/min/1.73m2) = 141×(SCr(mg/dL)/a)b×0.993age×1.018 (if female) (in male, if SCr ≤ 0.9 mg/dL: a = 0.9, b=−0.411, else a = 0.9, b=−1.209; in female, if SCr ≤ 0.7 mg/dL: a = 0.7, b=−0.329, else a = 0.7, b=−1.209). 88.4µmol/L=1 mg/dL for SCr21. Kidney damage was defined as eGFR < 60 mL/min/1.73 m2 and/or urine albumin-to-creatinine ratio (UACR) > = 30 mg/g2. Statistical analysis Statistical analysis was performed using Stata version 16.0 (StataCorp, College Station, TX, USA). Normally distributed variables were expressed as mean ± standard deviation (± sd) and compared between groups using t-tests. Continuous variables with a skewed distribution are reported as median (P25-P75) and compared after normalization transformation. Categorical data were presented as n (%) and analyzed by χ2 tests. Internal dataset was used to identify influencing risk factors and construct predictive models by multivariate linear regression analysis. Collinearity and residual diagnostics were performed. For back-transformation of the log-linear predictions, Duan’s Smearing Factor was applied to correct for re-transformation bias22. To evaluate the predictive accuracy of the linear regression model, we assessed the agreement between the observed and predicted FEUA values. The Spearman correlation coefficient was calculated to examine the strength of the linear relationship. The overall goodness-of-fit was quantified using the coefficient of determination (R2). Additionally, the root mean squared errors (RMSE) and mean absolute errors (MAE) were used to measure the average magnitude of the errors between predicted and observed values. A scatter plot of predicted versus observed FEUA was generated, with an identity line (y = x) superimposed as a reference to visually inspect the degree of agreement and heteroscedasticity across the measurement range. Furthermore, the discriminative ability of the continuous predicted FEUA for identifying real-world FEUA elevation outcome was evaluated using receiver operating characteristic curve (ROC) analysis. To further assess the calibration and clinical utility of the model, the continuous predicted FEUA values were used as a predictor in a logistic regression model to determine a predicted probability of real-world elevated FEUA. Based on these predicted probabilities, a calibration curve was generated to assess the agreement between the predicted and observed risks. Finally, decision curve analysis (DCA) was performed to estimate the clinical net benefit of the model across a range of threshold probabilities, ensuring its practical utility for clinical decision-making. To assess the stability and internal validity of the final predictive model, internal validation using 5-fold, 10-fold, and leave-one-out cross-validation were performed. These resampling techniques were used to evaluate the consistency of the model coefficients and standard errors, ensuring that the model was not over fitted to the development dataset. To explore the clinical applicability of the model across different patient spectrums and to rule out potential spectrum bias, subgroup analyses in training dataset were performed. Specifically, the dataset was stratified by: (1) renal function (with vs. without kidney damage, defined as UACR ≥ 30 mg/g or eGFR < 60 mL/min/1.73 m²); and (2) glycemic status (FPG < 10 mmol/L vs. ≥ 10 mmol/L, based on the physiological renal threshold for glucose). These analyses aimed to determine whether the model maintains consistent predictive accuracy and clinical utility across varied clinical phenotypes. To evaluate the robustness of the prediction model within the context of contemporary clinical practice and for the exploration potential of existing public or local dataset which didn’t come with FEUA-related urinary variables originally, the robustness of the model was further validated using independent (cross diverse temporal) datasets (data from 2017 to 2025), which were non-overlapping with the development dataset in terms of time range. Besides, the participants under medication that were initially excluded were kept as another subgroup for sensitivity validation of our predictive model to see the generalizability of our predictive model in subgroups of participants under medication which is quite common in clinical settings. The significance level was set at α = 0.05 (two-tailed). Sample size In sample size calculation, we used the “Pmsampsize” module in Stata16 with type of continuous dependent variable. Minimum threshold of sample size required for model development based on the parameters (R-square = 0.5, five predictors, intercept = 7, standard deviation = 3) was 239. In this research, sample sizes were 1 510, for training and internal validation and 791 for (temporal) external validation, of which 406 cases were enrolled during 2017 and 385 cases were enrolled during 2025, all exceeding the minimum threshold needed to fulfill modeling requirements. Results Univariate analysis for internal dataset A total of 1 510 participants was enrolled. The mean age was 55.45 ± 13.24 years (range: 18–88 years). Among them, 543 were female (35.96%). The mean fractional excretion of uric acid (FEUA) was 6.89% ± 2.81%, and the prevalence of elevated FEUA was 109(7.22%, 95%CI:5.96–8.64%). Overall characteristics of participants stratified by elevated FEUA were presented in Table 1. Table 1 showed that the final participants consisted of type 2 diabetes patients with little health inequalities between socio-demographic groups, relatively homogeneous disease profiles, and a low burden of comorbidities. Table 1Overall characteristics of participants and those stratified by FEUA ≥ 11% ([xˉ ±s; n(%)]; n = 1 510).FactorsOverall(n = 1510)FEUA < 11(n = 1401)FEUA > = 11(n = 109)t/z P OR(95CI)Female (n, %)543 (35.96%)492(35.12%)51(46.79%)2.430.0151.625(1.098–2.404)Age (years)55.45 ± 13.2454.92 ± 13.2262.24 ± 11.595.45< 0.0011.051(1.032–1.070)Diabetes duration (years)10.00 (2.00, 16.00)10.00 (2.00, 16.00)12.00 (3.00, 20.00)1.480.1391.142(0.958–1.362)BMI(kg/m2)26.53 ± 3.7726.63 ± 3.7625.23 ± 3.62−3.73< 0.0010.897(0.847–0.950)SBP (mmHg)135.06 ± 17.84135.06 ± 17.62135.12 ± 20.500.040.9711(0.989–1.011)DBP (mmHg)78.18 ± 12.0378.48 ± 12.0274.34 ± 11.58−3.470.0010.97(0.953–0.987)UACR (mg/g)6.82 (3.51, 20.37)6.66 (3.38, 19.36)9.88 (4.86, 62.17)5.36< 0.0011.303(1.183–1.435)eGFR(ml/min/1.73m2)100.80 ± 18.43102.15 ± 17.0683.39 ± 25.38−9.46< 0.0010.954(0.945–0.963)SUA (µmol/L)346.90 ± 89.12354.10 ± 85.29254.30 ± 85.67−10.44< 0.0010.982(0.979–0.985)Male364.04 ± 87.31369.82 ± 83.38273.52 ± 97.88−7.6< 0.0010.983(0.979–0.987)Female316.37 ± 84.06325.07 ± 81.14232.45 ± 63.36−6.92< 0.0010.979(0.973–0.985)UUA(µmol/L)3.63 ± 1.243.61 ± 1.253.78 ± 1.121.340.181.107(0.954–1.283)SCr (µmol/L)65.10 ± 19.8363.75 ± 16.8382.56 ± 38.637.69< 0.0011.032(1.023–1.040)uCr (µmol/L)10.64 ± 3.9010.77 ± 3.918.92 ± 3.21−4.79< 0.0010.856(0.803–0.912)FEUA(%)6.89 ± 2.816.35 ± 1.9013.84 ± 3.35---ALT (U/L)19.00 (14.00, 30.00)19.00 (14.00, 30.00)17.00 (12.00, 29.00)−1.31a0.1910.804(0.579–1.115)AST (U/L)18.00 (14.00, 24.00)18.00 (14.00, 24.00)19.00 (15.00, 26.00)1.17a0.241.279(0.848–1.928)FPG (mmol/l)8.08 ± 3.538.03 ± 3.468.69 ± 4.261.850.0641.047(0.997–1.098)HbA1c(%)9.30 ± 1.899.29 ± 1.909.37 ± 1.670.420.6741.022(0.923–1.132)HDL-C (mmol/L)1.05 ± 0.251.04 ± 0.251.08 ± 0.301.580.1141.793(0.869–3.702)Male0.99 ± 0.230.99 ± 0.231.04 ± 0.281.60.112.3(0.827–6.393)Female1.14 ± 0.261.14 ± 0.251.13 ± 0.31−0.290.7730.846(0.272–2.630)LDL-C (mmol/L)2.69 ± 0.862.70 ± 0.852.58 ± 0.92−1.390.1640.846(0.669–1.071)TC (mmol/L)4.36 ± 1.144.37 ± 1.134.31 ± 1.21−0.540.5920.953(0.800–1.136)TG (mmol/L)1.51 (1.11, 2.15)1.51 (1.12, 2.14)1.54 (1.01, 2.40)−0.15a0.8780.973(0.685–1.381)hsCRP1.41 (0.67, 3.29)1.38 (0.67, 3.19)1.91 (0.77, 4.27)2.3a0.0221.21(1.028–1.424)Urinary β2-MG(µg/mL)242.53 (114.50, 612.80)227.50 (107.40, 567.21)621.56 (237.80, 1478.73)6.43a< 0.0011.778(1.492–2.118)Gout (n, %)519 (34.37%)494(35.26%)25(22.94%)−2.580.010.546(0.345–0.865)DR (n, %)159 (10.53%)148(10.56%)11(10.09%)−0.150.8770.95(0.498–1.813)DPN (n, %)156 (10.33%)144(10.28%)12(11.01%)0.240.8091.08(0.579–2.016)Hypertension (n, %)1,017 (67.35%)935(66.74%)82(75.23%)1.810.071.514(0.966–2.371)CHD (n, %)726 (48.08%)679(48.47%)47(43.12%)−1.070.2830.806(0.544–1.194)Fatty liver (n, %)552 (36.56%)516(36.83%)36(33.03%)−0.790.4280.846(0.559–1.279)a: The probability and odd ratio obtained by taking the natural logarithm of the independent variable.Abbreviations: FEUA: fractional excretion of uric acid; BMI: body mass index; SBP: systolic blood pressure; DBP: diastolic blood pressure; UACR: urinary uric acid to creatinine ratio; eGFR: estimated glomerular filtration rate; SUA: serum uric acid; UUA: urinary uric acid; sCr: serum Creatinine; uCr: urinary creatinine; ALT: alanine aminotransferase; AST: aspartate aminotransferase; FPG: fasting plasma glucose; HbA1c: Hemoglobin A1c; HDL-C: high density lipid protein cholesterol; LDL-C: low density lipid protein cholesterol; TC: total cholesterol; TG: triglyceride; hsCRP: high-sensitive C reactive protein; β2MG: β2-microglobulin; DR: diabetic retinopathy; DPN: diabetic peripheral nerve disease; CHD: coronary heart disease. Overall characteristics of participants and those stratified by FEUA ≥ 11% ([xˉ ±s; n(%)]; n = 1 510). a: The probability and odd ratio obtained by taking the natural logarithm of the independent variable. Abbreviations: FEUA: fractional excretion of uric acid; BMI: body mass index; SBP: systolic blood pressure; DBP: diastolic blood pressure; UACR: urinary uric acid to creatinine ratio; eGFR: estimated glomerular filtration rate; SUA: serum uric acid; UUA: urinary uric acid; sCr: serum Creatinine; uCr: urinary creatinine; ALT: alanine aminotransferase; AST: aspartate aminotransferase; FPG: fasting plasma glucose; HbA1c: Hemoglobin A1c; HDL-C: high density lipid protein cholesterol; LDL-C: low density lipid protein cholesterol; TC: total cholesterol; TG: triglyceride; hsCRP: high-sensitive C reactive protein; β2MG: β2-microglobulin; DR: diabetic retinopathy; DPN: diabetic peripheral nerve disease; CHD: coronary heart disease. Multivariate linear regression models based on internal dataset 1 510 cases were used for constructing multivariate linear regression models. Three steps were adopted while choosing variables for predicting FEUA levels. First, a total of 10 variables which were significant in univariate analysis were enrolled. Second, FPG was added which had long been proved a risk factor of FEUA23. Third, 24-hour urinary creatinine was dropped, which was an index involved in FEUA calculation and those who need to be predicted usually do not include 24-hour urine biochemical indicators. Then, ten variables were kept as predictors for constructing FEUA prediction models. FEUA, β2MG, and UACR underwent logarithmic transformation with base e to correct for skewness before modeling construction. Ln_FEUA was adopted as dependent variable (Y). The procedure for constructing FEUA prediction models by univariate and backward-stepwise multivariate analysis were presented in Table 2. Table 2Predictor coefficients in univariate and multivariate backward stepwise regression analysis for constructing FEUA prediction models (n = 1510).UnadjustedAdjustedPredictorsβ(95%CI)P valueβSE95%CIP valueDeleteOrderR2 beforeDeletedAdj. R2before DeletedAICBICBMI−0.0205(−0.0254~−0.0157)< 0.001−0.00030.0016−0.0035 ~ 0.00290.86510.63810.6356−238.5310−180.0120age0.0080(0.0067 ~ 0.0094)< 0.0010.00040.0010−0.0016 ~ 0.00240.68020.63800.6359−240.5000−187.3027DBP−0.0034(−0.0049~−0.0019)< 0.0010.00030.0005−0.0007 ~ 0.00130.53730.63800.6361−242.330−194.4520Ln_β2MG0.0277(0.0194 ~ 0.0360)< 0.0010.00400.0029−0.0016 ~ 0.00970.16240.63790.6362−243.9469−201.3880eGFR−0.0079(−0.0088~−0.0070)< 0.001−0.00090.0005−0.0017 ~ 0.00080.08350.63740.6360−243.9807−206.7417Ln_UACR0.0330(0.0218 ~ 0.0442)< 0.0010.00680.0037−0.0005 ~ 0.01400.06660.63670.6355−242.9683−211.0491Sex0.1118(0.0734 ~ 0.1502)< 0.0010.14620.01320.1203 ~ 0.1721< 0.0001-0.63590.6349−241.58−214.9807SUA−0.0025(−0.0026~−0.0023)< 0.001−0.00310.0001−0.0032~−0.0029< 0.0001-SCr0.0051(0.0042 ~ 0.0060)< 0.0010.01120.00030.0104 ~ 0.0118< 0.0001-FPG0.0067(0.0014 ~ 0.0120)0.0130.00680.00160.0036 ~ 0.0100< 0.0001-_cons--1.94080.04151.8594 ~ 2.0222< 0.0001-Abbreviations: AIC: Akaike Information Criterion; BIC: Bayesian Information Criterion; BMI: body mass index; DBP: diastolic blood pressure; Ln_β2MG: natural log-transformed β2-microglobulin; eGFR: estimated glomerular filtration rate; Ln_UACR: natural log-transformed urinary uric acid to creatinine ratio; SUA: serum uric acid; sCr: serum Creatinine; FPG: fasting plasma glucose. Predictor coefficients in univariate and multivariate backward stepwise regression analysis for constructing FEUA prediction models (n = 1510). Abbreviations: AIC: Akaike Information Criterion; BIC: Bayesian Information Criterion; BMI: body mass index; DBP: diastolic blood pressure; Ln_β2MG: natural log-transformed β2-microglobulin; eGFR: estimated glomerular filtration rate; Ln_UACR: natural log-transformed urinary uric acid to creatinine ratio; SUA: serum uric acid; sCr: serum Creatinine; FPG: fasting plasma glucose. Table 2 showed that BMI, age, DBP, Ln_β2MG, eGFR, Ln_UACR were dropped successively. Sex, SUA, SCr and FPG were kept in the final model. R-square decreased from 0.6381 to 0.6359 (99.66%), adjusted R-square decreased from 0.6349 to 0.6356 (99.89%). Variance Inflation Factor for all model and each factor were all less than 1.28. The mean of the residuals is 9.39E-11 ± 0.2227. Plots describing the normality of residuals were presented in Figure S1 in supplemental document, in which the histogram and the normal quantile plot indicated a distribution of normality. Besides, other plots in Figure S1 showed that the majority of the scatter points were around the mean of the residuals within the range of −0.5 to + 0.5. No outliers, which deviate from the nearest quartile three times of inter quartile range, were identified. All the diagnostic indices showed a good performance of our modeling construction. Final formula of the prediction model The final predictive modeling for natural logarithmic value of FEUA, Ln(FEUA), was listed below: The predicted FEUA is the product of the value of e raised to the power of the predicted value and the Duan’s Smearing Factor. The value range of the current predictors are SUA (76~683µmol/L), SCr (16–254µmol/L), and FPG (2.36–29.60mmol/L), remind model users to treat samples outside this range with caution. Predictive ability of FEUA levels We used the formula above to calculate predicted FUEA. The comparison of observed FEUA and predicted FEUA was presented in Table 3. The mean difference between observed and predicted FEUA was minimal (−0.005), representing only 0.1% deviation from the observed mean, demonstrating excellent overall agreement. RMSE and MAE were 1.6841% and 1.1856%, respectively. The Spearman correlation coefficient between the observed FEUA and predicted FEUA was 0.7870 (P < 0.001). The violin plot (Fig. 2a) and scatter plot (Fig. 2b) of the observed FEUA and predicted FEUA were presented in Fig. 2. Table 3Difference between the observed FEUA and predicted FEUA.Variable N MeanStd. Dev.MinimumMaximumObserved FEUA15106.8892.8151.79131.069Predicted FEUA15106.8942.3772.07351.352Absolute Difference-0.0050.4380.28220.283Percent(%)-0.0010.1560.1570.653Abbreviation: FEUA: fractional excretion of uric acid. Difference between the observed FEUA and predicted FEUA. Abbreviation: FEUA: fractional excretion of uric acid. Predictive performance between observed FEUA and elevated FEUA and predicted FEUA----Violin plot (a), scatter plot (b), Spine plot (c), ROC curve (d), calibration curve (e), and decision curve (f). Figure 2a showed that the distribution of predicted FEUA was similar to the observed FEUA. Figure 2b showed that R-square between the observed FEUA and predicted FEUA was 0.644. Predictive ability of elevated FEUA We divided all cases into normal FEUA group (FEUA < 11%) and elevated FEUA group based on both observed FEUA and predicted FEUA. The comparison of the classification of elevated FEUA group based on the observed and predicted elevated FEUA was presented in Table 4. Table 4 showed that the overall accuracy was (1390 + 40)/1510 = 94.70%. This model also yielded a sensitivity of 36.70%, specificity of 99.22%, positive predictive value of 78.43%, and negative predictive value of 95.27%. The sensitivity was not as good as overall accuracy due to the low prevalence of elevated FEUA (7.22%). Table 4The classification of the observed and predicted elevated FEUA.ObservedPredicted FEUAFEUAFEUA < 11FEUA > = 11TotalFEUA < 111 390(95.27)11(21.57)1,401FEUA > = 1169(4.738)40(78.43)109Total1 459(100)51(100)1 510Abbreviation: FEUA: fractional excretion of uric acid. The classification of the observed and predicted elevated FEUA. Abbreviation: FEUA: fractional excretion of uric acid. Spine plot of the comparison of the classification of elevated FEUA based on observed and predicted FEUA (Fig. 2c), ROC curve (Fig. 2d), calibration plot (Fig. 2e), and decision curve (Fig. 2f) were presented in Fig. 2. The Spine plot (Fig. 2c) visually demonstrates the high agreement between observed and predicted classifications. Most cases were correctly classified. Figure 2d showed that the area under the curve was 0.9552 (95% confidence intervals (C.I.) 0.9413 ~ 0.9691). We achieved a Youden index (0.7814) at the cut-point (8.0257%) of FEUA with a high sensitivity (0.9541), specificity (0.8273) and ROC area (0.8907). Using an Youden index-derived cut-point was meant to demonstrated the best differentiating ability of our predictive model and further assist subgroup analysis by providing the most balanced classification for comparing relative performance across strata24. This approach is standard practice for evaluating model stability in subgroup analyses25. While the 11% threshold remains the primary reference for clinical interpretation. Calibration curve (Fig. 2e) assessed the agreement between the predicted probability and actual outcomes, and decision curve (Fig. 2f) estimated the clinical net benefit of the model across a range of threshold probabilities, ensuring its practical utility for clinical decision-making. Figure 2 revealed the effectiveness and agreement of the prediction of observed FEUA and elevated FEUA using predicted FEUA values. Cross validation showing the stability of the final model We performed 5-fold and 10-fold cross-validation for the internal dataset and compared the results with 1-fold indices. The good-of-fitness indices for one-fold, five-fold, ten-fold validation analysis was presented in Table S1. Table S1 showed that the average root mean squared errors (RMSEs) for 1-fold, 5-fold, 10-fold, and leave-one-out cross-validation are 1.6841, 1.7038 ± 0.1899, 1.6900 ± 0.2950 and 1.7081, respectively. The mean ROC in 5-fold and 10-fold validation analysis was 0.9536 ± 0.0162 and 0.9535 ± 0.0316, respectively. These indices showed a stability for internal dataset cross validation. Predictive ability in subgroups of training dataset Comparison of indices between observed and predicted FEUA and predictive indices on elevated FEUA for full dataset and four subgroups were presented in Table S2 in supplemental document. Table S2 showed that the ROC values were all higher than 0.9459. Predicted FEUA for Youden index cut-point varied from 7.9605 to 9.3142. Although this range can assist model users in making optimal predictions for the elevated FEUA, model users can define the cut-point that best suits their own study. Other good-of-fitness indices also remained stable. External dataset analysis and validation We selected 791 inpatients admitted from January 1 st, 2017, to December 31 st, 2017 and January 1 st, 2025, to December 31 st, 2025. The value range of the current predictors are SUA (99~704µmol/L), SCr (26–176µmol/L), and FPG (2.4–20.40mmol/L), which overlap with the range of predictors of the model in training set to a large extent. The mean age was 54.91 ± 15.38 years (range: 13–90 years). Among them, 367(46.40%) were female. The mean FEUA was 6.93 ± 3.20%, and the prevalence of elevated FEUA was 59 (7.46%, 95%CI:5.73–9.52%). The overall characteristics of the temporal external validation dataset stratified by whether FEUA ≥ 11% were presented in Table S3 of the supplementary document. We split the full external dataset into two groups by year, the 2017-year-group and 2025-year-group, and into three groups randomly. Resemblance between the observed FEUA and predicted FEUA in full external dataset and in its five subgroups, respectively, was presented in Table S4 and Table S5, among which the predicted FEUA in full external dataset was 6.97 ± 2.17%, very similar with the observed FEUA (6.93 ± 3.20%) in this dataset. The performance metrics (RMSE, Spearman correlation, AUC, calibration) are similar between the 2017-year-group and 2025-year-group, which substantially strengthened the case for the model’s robustness. Table S5 also showed good quality of our predictive model in the subgroup of participants under medication. Comparison of good-of-fitness indices calculated based on predicted FEUA between full external dataset and five subgroups was presented in Table S5. Table S5 showed that the Spearman correlation coefficient between the observed FEUA and predicted FEUA in external dataset was 0.7451 (P < 0.001). Coefficient of Determination was 0.5439 (P < 0.001). The overall accuracy rate of elevated FEUA binary outcome based on the observed and predicted elevated FEUA was 94.18%. Table S5 also showed a stability of all indices for subgroups not only after the time of training dataset, but also before it. Good-of-fitness indices calculated based on predicted FEUA and elevated FEUA for subgroups under medication of diuretics, fenofibrate, ACEI inhibitors, ARB, SGLT2 inhibitors were presented in Table S6. Table S6 showed that the Spearman correlation coefficient between the observed FEUA and predicted FEUA in the special dataset was still 0.7899 (P < 0.001). Coefficient of Determination was 0.4186 (P < 0.001). The overall accuracy rate of elevated FEUA binary outcome for both cut-point were 90.12%. ROC for elevated FEUA was 0.9029 (95% CI: 0.8817 ~ 0.9241). Table S6 showed a stability of all indices except Coefficient of Determination comparing with those subgroups in Table S4 and Table S5. The violin plot (Figure S2a), scatter plot (Figure S2b), Spine plot (Figure S2c), ROC curve (Figure S2d), calibration curve (Figure S2e) and decision curve (Figure S2f) of the observed FEUA and elevated FEUA and predicted FEUA in external dataset were presented in Figure S2 in supplemental document. Figure S2d showed that the area under the ROC was 0.8946 (95% CI: 0.8545 ~ 0.9347), which indicated that the prediction model has good performance in an external dataset, especially when predicting elevated FEUA binary outcome. Calibration curve (Figure S2e) assessed the agreement between the predicted probability and actual outcomes, and decision curve (Figure S2f) estimated the clinical net benefit of the model across a range of threshold probabilities, ensuring its practical utility for clinical decision-making. Discussion In this study, we have developed and validated a linear regression model for predicting the FEUA in hospitalized patients with T2DM, based on serum laboratory tests. This model performed well in terms of their discriminatory ability and practical value in clinical applications. The model provided functions as a quantitative supporting instrument for predicting FEUA in hospitalized patients with T2DM based on serum laboratory tests, aiding physicians in making preventive and therapeutic measures. Even though our model was constructed based on a group of participants not under medications which would alter uric acid metabolism or affect kidney function, as demonstrated in our sensitive analysis, our predictive model also functions well in the subgroup of participants under those medications, featuring good generalizability and satisfying clinical application requirements. The predictors enrolled in our model are easily available from various laboratory centers or public data resources, demonstrating the practicality of this model. If applied wisely and widely, it has the potential to enhance the consistency of evidence-based medical studies associated with FEUA and its related diseases, such as hyperuricemia, gout, metabolic syndrome, T2DM and kidney damages26. Two aspects of contents will be further discussed below. One is why FEUA is becoming increasingly important and is worth and requires predicting. The other is the potential mechanisms of the relationship between the predictors and FEUA. The relationship between SUA and CKD has long been controversial. A great number of studies reported the positive association between HUA and CKD in both general population and T2DM population27–41. Many studies also demonstrated lowering SUA levels could slow CKD progression and prevent kidney failure42–54. Meta-analyses regarding the effect of lowering UA with xanthine oxidase (XORs) inhibitors, mostly allopurinol and febuxostat, on kidney protection, published in recent 5 years, concluded there exists a positive protective effect, including up-regulate eGFR55–57, slow eGFR declination58,59, alleviate albuminuria60, decrease SCr60 and prevent or slow CKD progression61–63. However, still some RCTs reported lowering SUA didn’t exert any reno-protective effect64–67. Mendelian Randomization (MR) studies also reported inconsistent results regarding if SUA is causally associated with kidney damage68–75. In clinical practice, some patients still develop or progress to DKD despite controlling established risk factors, suggesting the presence of incompletely recognized “residual risks”76. Considering both hyperuricemia (HUA)77–79 and genetic hypouricemia80,81 are associated with kidney damage, and both HUA and genetic hypouricemia are related to higher uric acid excretion19,80,81, and non-synonymous SNPs of ABCG2 (rs2231142) are associated with renal overload (ROL) type gout82, nephrolithiasis83, CKD83 and decreased estimated glomerular filtration rate (eGFR)84. Taken together, elevation in urinary uric acid excretion seems more like the fundamental drive behind the seemingly contradictive high-/low- SUA both causing kidney damage. This means much for patients with T2DM, who has dysregulated UA metabolism and are already prone to suffer from kidney damage because of hyperglycemia, hyperinsulinemia and unhealthy diet/lifestyle13. Regarding urinary uric acid excretion (UUAE), two thirds of uric acid excretion is carried out by kidneys10. FEUA is currently acknowledged as a precise measurement of UAE, with the normal FEUA range of 5.5–11.0%11. However, clinically estimating FEUA is arduous, some hospitals still don’t conduct this analysis. Besides, many public databases do not include this variable or lack necessary variables to calculate it, making larger scale investigation regarding FEUA and possible related disease, such as kidney damage, hard to implement. A simple and reliable method to predict FEUA is needed, both for research purposes and clinical practice convenience. Predictors suggested in our model included SCr, SUA, and FPG and sex. Potential mechanisms of the relationship between the predictors and FEUA were discussed below. Five studies showed FEUA was negatively associated with eGFR, which was calculated based on SCr5–9. This could be the underlying reason of the positive relationship between FEUA and SCr, as shown in our prediction formula. By comparing overall renal safety events of Lesinurad, Verinurad, Dotinurad, SHR4640, and Benzbromarone in patients with hyperuricemia, another study used a semi-mechanistic pharmacokinetic/pharmacodynamic model to establish the dose-exposure-effect relationship of URAT1 inhibitors with or without the combination of xanthine oxidase inhibitors (XOIs), and concluded that the adverse renal effects of URAT1 inhibitors are dose-dependent, and the combination with high doses of XOIs can significantly reduce the renal safety risk by reducing uric acid excretion by the kidneys85. Basic experimental studies also reported various underlying mechanisms regarding urinary uric acid causes kidney damage, among which innate immune response, inflammation, oxidative stress mediated apoptosis, endothelial dysfunction, interstitial fibrosis and ER stress are involved86–95. The negative relationship between SUA and FEUA was reported before23 and could be explained by the usefulness of UA in scavenging free radicals in T2DM population. T2DM is related to oxidative stress96–100, and uric acid is a strong free radical scavenger101, about 60% of the free radicals in the body are scavenged by uric acid102. Therefore, up-regulation of UA reabsorption transporters to reabsorb higher percentage of UA in T2DM patients is, perhaps in itself, a protective mechanism, through which the increased uric acid helps cleansing the body of free radicals. Another reason might lie upon the usefulness of UA in evolution. Owing to renal reabsorption of UA and the loss of uricase activity, UA is maintained at a higher physiological level in humans than in other mammals, which provides a survival advantage during evolution10, such as maintaining a higher and more appropriate blood pressure level for primates who walks up-right and during times with very low salt ingestion and is associated with increased intelligence103, allowing human to survive environmental stressors, evolution bottlenecks, life-threatening pathogens, and contributing to lower risk of neurodegenerative diseases104. So, higher SUA accompanying lower FEUA, to accumulate more and more UA in blood, is along the direction of evolution. The relationship between glucose and uricosuria has been extensively investigated. Previous studies have reported that SUA levels were the highest in subjects with impaired glucose regulation, followed by subjects with normoglycemia and finally by subjects with T2DM, more accurately, SUA levels initially increased with increasing fasting blood glucose (FBG) and postprandial blood glucose (PPBG) levels, up to 7 mmol/L and 10 mmol/L, respectively, and thereafter decreased at higher FBG and PPBG levels105–108. Possible underlying mechanisms includes that glucose and UA are absorbed in renal proximal tubule through co-transporter, so when FPG is higher than renal glucose reabsorption threshold, too much glucose in the urine would compete UA for their co-transporter to get reabsorbed, causing UA excretion elevation23. Many studies provide strong evidence for the role of female hormones in regulating FEUA. As early study reported, compared with women’s group, SUA levels were higher, and FEUA levels were lower in men’s group23. Additional studies demonstrated changes in SUA during ovulation in pre-menopausal women, with peak SUA levels occurring during the follicular phase when estrogen levels are lowest109. Furthermore, SUA was positively correlated with follicle stimulating hormone and inversely associated with estradiol and progesterone109. Female sex hormone levels decrease after menopause, SUA increases, but this increase in SUA can be mitigated by hormone replacement therapy110. Another study examined the effects of oral administration of synthetic sex hormones on both women and men, and demonstrated a negative relationship between SUA and administered exogenous estrogen, as well as an increase in urinary UA after the administration of exogenous estrogen111. Similar effects were observed when post-menopausal women received progesterone112. Another study examined the effects of hormone therapy on transsexual participants, demonstrating a decrease in SUA in male-to-female (MTF) participants, and an increase in SUA in female-to-male (FTM) participants after one year of cross-sex hormone administration. Furthermore, baseline FEUA was higher in FTM subjects and significantly decreased after the loss of estrogen, while MTF subjects demonstrated an increase in FEUA with the addition of estrogen113. Evidences further reinforce that the female sex hormones regulate UA handling machinery at either the transcriptional or post-transcriptional levels. Estrogen has been shown to down-regulate SLC2A9 at the post-transcriptional level through estrogen receptor beta (ER-β) induced proteasomal activation114 resulting in a decrease in the reabsorption of UA, leading to higher FEUA. Transcription factor encoded by transcription factor gene, HNF4A, has been reported to interact with estrogen receptor alpha (ER-α) to facilitate the transcription of target genes, such as SLC2A9 (encoding the second-primary renal UA re-absorptive transporter GLUT9), and PDZK1 (encoding the epithelial-specific PDZ domain-containing multivalent scaffolding protein, PDZK1, which interacts with the C-terminal of URAT1, the primary renal UA re-absorptive transporter, and increases the transportation speed of URAT1 by 1.4-fold115116. Since, sex hormone plays such a vital part in UA excretion, sex was adopted in our predictive formula. Strengths and limitations Our study has several strengths. First, it was conducted among a large modeling sample, rendering it statistically efficient. Second, our study processed a successful external validation, which guaranteed practicality of our formula. Thirdly, our formula captures a clear description of the relationship between the predictors and the outcome and has good performance. Last but not least, despite being constructed based on a group of participants not under medications which would alter uric acid metabolism or affect kidney function, our predictive model also functions well in the subgroup of participants under those medications, featuring good generalizability and satisfying clinical application requirements. This study has limitation as well. It is retrospective, single-centered and cross-sectional, which does not accurately represent the wider population and future multicenter external validations are still needed. Conclusion In conclusion, we developed a linear regression model to predict the FEUA level in hospitalized patients with T2DM, based on four variables from serum laboratory tests. It has good predictive performance and is easily applicable. The model we developed would benefit the clinical management of patients with hyperuricemia and its related comorbidities, and it also significantly expands the scope of application of public databases. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements We extend our sincere gratitude to the information center of Peking University People’s Hospital for setting up the hospital information system (HIS) and generously allowing us access to the information of targeted population in our research. Abbreviations 95% confidence intervals β2-microglobulin Angiotensin-converting enzyme inhibitors American Diabetes Association Angiotensin receptor blocker area under the curve body mass index confidence intervals chronic kidney disease diastolic blood pressure decision curve analysis diabetic kidney disease diabetic nephropathy estimated glomerular filtration rate endoplasmic reticulum estrogen receptor beta end-stage kidney disease fractional excretion of uric acid fasting plasma glucose fasting blood glucose female-to-male glucose transporter type 9 hyperuricemia and mean of absolute error Mendelian Randomization male-to-female odds ratio postprandial blood glucose receiver operating characteristic renal overload root mean squared errors Systolic blood pressure serum creatinine standard error single nucleotide polymorphism serum uric acid type 2 diabetes mellitus uric acid uric acid excretion urine albumin-to-creatinine ratio urinary creatinine urinary uric acid urinary uric acid excretion urate transporter 1 xanthine oxidase inhibitors Author contributions X. H., Lingli Z. and L. J.:Conceptualization, supervision and funding acquisition; Lin Z., and X. Z: Methodology; X. Z, W.Y., and L.C.: Investigation and data curation; Lin Z: Software analysis and original draft. All authors reviewed the manuscript. Funding This study was funded by National Science and Technology Major Project (Grant No.2024ZD0530204). This work is also funded by the 2024 National Clinical Key Specialty Construction Program of China (Department of Endocrinology, Peking University People’s Hospital) with support from the central government budget. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data availability The data and the Stata code (calculating Duan’s smearing factor) that support the findings of this study are openly available in the “Figshare” at https://doi.org/10.6084/m9.figshare.32963489, reference number 32963489. Declarations Competing interests The authors declare no competing interests. Ethics approval and Consent to participant The study adhered to the principles outlined in the World Medical Association Declaration of Helsinki, with study protocol reviewed and approved by the Institutional Review Board of Peking University People’s Hospital (approval No. 2025PHB080-001). Owing to the retrospective nature of the study and the use of de-identified data with no personal identifiers, the requirement for written informed consent was waived by the Institutional Review Board of Peking University People’s Hospital. Consent for publication This article has not been submitted before and there exists no previous report that might be considered redundant publications of the same or very similar work. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Linong Ji, Email: Jiln@bjmu.edu.cn. Lingli Zhou, Email: zhoulingli@pkuph.edu.cn. Xueyao Han, Email: xueyaohan@sina.com. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The data and the Stata code (calculating Duan’s smearing factor) that support the findings of this study are openly available in the “Figshare” at https://doi.org/10.6084/m9.figshare.32963489, reference number 32963489.