Evaluating the risk of developing hyperuricemia in patients with type 2 diabetes mellitus using least absolute shrinkage and selection operator regression and machine learning algorithm

Abstract

Background

Hyperuricemia is a common complication of type 2 diabetes mellitus and can lead to serious consequences such as gout and kidney disease.

Methods

Patients with type 2 diabetes mellitus from six different communities in Fuzhou were recruited from June to December 2022. Questionnaires, physical examinations, and laboratory tests were conducted to collect data on various variables. Variable screening steps were performed using univariate and multivariate stepwise regression, least absolute shrinkage and selection operator (LASSO) regression, and Boruta feature selection. The dataset was divided into a training-testing set (80%) and an independent validation set (20%). Six machine learning models were built and validated.

Results

A total of 8243 patients with type 2 diabetes mellitus were included in this study. According to Occam's razor method, the LASSO regression algorithm was determined to be the optimal risk factors selection method, and nine variables were identified as parameters for the risk assessment model. The absence of diabetes medication and elevated fasting blood glucose levels exhibited a negative correlation with the risk of hyperuricemia. Conversely, seven other variables demonstrated a positive association with the risk of hyperuricemia among patients diagnosed with type 2 diabetes mellitus. Among the six machine learning models, the artificial neural network (ANN) model demonstrated the highest performance. It achieved an areas under curve of 0.736, accuracy of 68.3%, sensitivity of 65.0%, specificity of 72.2%, precision of 73.6% and F1-score of 69.0%.

Conclusions

We developed an ANN model to better evaluate the risk of hyperuricemia in the type 2 diabetes population. In the type 2 diabetes population, women should pay particular attention to their uric acid levels, and type 2 diabetics should not neglect their obesity level, blood pressure, kidney function and lipid profile during their regular medical check-ups, in order to do their best to avoid the risks associated with the combination of type 2 diabetes and hyperuricemia.

Keywords

Machine learning type 2 diabetes mellitus hyperuricemia risk assessment model LASSO regression

Introduction

The diabetes epidemic has become a global public health problem, posing serious health, social, economic and medical problems to mankind. According to the International Diabetes Federation (IDF), 537 million people will have diabetes in 2021, and the number of people with diabetes is expected to reach 643 million by 2030 and 783 million by 2045, with about 90 per cent of the world's diabetic population suffering from type 2 diabetes.¹ Data from 2019 showed that the standardized incidence, standardized mortality and disability-adjusted life-years rates of type 2 diabetes mellitus in China increased significantly from 1990, by 15.37%, 6.26% and 9.34%, respectively.²

Hyperuricemia is mainly caused by excessive production or poor excretion of uric acid (UA), and with socio-economic development, hyperuricemia is now very common in the population. The prevalence of hyperuricemia has been reported to be 11.3%–47% in the United States, 11.9%–25.0% in Europe and 13.1%–13.3% in China.³ However, the rates of patient awareness, treatment and control of hyperuricemia are low.^4,5 Previous studies have shown that hyperuricemia is closely associated with type 2 diabetes mellitus. Firstly, the prevalence of hyperuricemia is higher in patients with type 2 diabetes mellitus than in the general population.^6,7 This is due to the metabolic disorders in patients with type 2 diabetes mellitus, resulting in reduced excretion of UA and increased production of UA and other factors. At the same time, hyperuricemia also affects the development and prognosis of type 2 diabetes mellitus⁶ and may aggravate the diabetes condition by increasing insulin resistance^7,8 and impaired β-cell function.^9,10 In addition, elevated UA in patients with type 2 diabetes mellitus may lead to health problems such as gout, renal impairment and cardiovascular disease.¹¹ Finally, studies have also shown that the coexistence of type 2 diabetes mellitus and hyperuricemia can increase the risk of all-cause mortality and end-stage renal disease in one step.^12,13 Therefore, evaluating the risk of hyperuricemia in patients with type 2 diabetes mellitus not only improves patients’ quality of life, but also effectively reduces the associated healthcare costs.

To the best of our knowledge, few studies have been reported on the use of machine learning for evaluating the risk of hyperuricemia in people with type 2 diabetes, as previous studies have only used machine learning to evaluating the risk of diabetes^14,15 or hyperuricemia.^16,17 Even the modeling studies for hyperuricemia were conducted on healthy populations. We used machine learning algorithms to evaluate the risk of hyperuricemia in patients with type 2 diabetes mellitus based on patient demographic data, questionnaire data and physical examination data. And by constructing appropriate risk assessment models, we can analyze and learn patterns and associations between various relevant features and evaluate the risk of patients based on these features. This personalized risk assessment approach helps to identify high-risk patient groups in advance so that targeted interventions and treatments can be taken to reduce the incidence of hyperuricemia.

In this study, we systematically analyzed the risk factors for hyperuricemia in patients with type 2 diabetes mellitus based on cross-sectional studies. In addition, we used six machine learning algorithms to build and validate a risk assessment model for hyperuricemia in patients with type 2 diabetes mellitus and compared the model performance of the different algorithms to ultimately develop effective risk assessment models for clinical applications.

Materials and methods

General information

From June to December 2022, we conducted a cross-sectional study survey of patients with type 2 diabetes mellitus. We recruited 8243 patients with type 2 diabetes mellitus to complete questionnaires and physical examination, Figure 1. The patients were drawn from community health centers (these communities were randomly selected) in six different urban areas of Fuzhou City, Fujian Province, China. Men with UA <420 μmol/L and women with UA <360 μmol/L were considered the normal UA group, with a total of 4477 cases; men with UA >420 μmol/L and women with UA >360 μmol/L were considered the high UA group, with a total of 3766 cases.

Figure 1.

The flow process diagram for selecting patients based on inclusion and exclusion criteria.

The inclusion criteria for this study were patients with type 2 diabetes mellitus. Patients younger than 18 years of age with type 2 diabetes mellitus were required to sign an informed consent form with the permission of their guardian. Additionally, all patients were required to have been residing in Fuzhou City for at least 180 days. The exclusion criteria included: (a) patients with a history of gout or currently diagnosed with gout, malignancy, hyperuricemia occurring before type 2 diabetes mellitus, type 1 diabetes mellitus, gestational diabetes mellitus or other specific types of diabetes mellitus; (b) patients who did not sign the informed consent form; (c) patients who did not actively participate in the investigation or had incomplete examination results. A detailed flow process diagram representation of our rigorous screening procedure is provided by Figure 1.

Questionnaire survey and physical examination

A homemade uniform questionnaire was used, and a face-to-face survey was conducted by trained primary care professionals. All respondents filled out an informed consent form to investigate basic personal information and health status, including gender, age, history of smoking and alcohol consumption, duration of diabetes, and medication history. Physical examination and laboratory tests are performed by professional and technical personnel in primary care institutions. Physical examination items included height, weight, waist circumference, blood pressure, etc. Laboratory tests included fasting blood glucose (FPG), UA, lipids, liver function, kidney function, etc.

Laboratory testing

Patients fasted overnight for more than 10 h and water fasting for more than 8 h. Fasting venous blood was drawn from 7:00 to 9:00 a.m. the next day, and FPG, UA, alanine aminotransferase (ALT), total bilirubin (TBil), serum creatinine (SCr), γ-glutamyl transpeptidase (GGT), blood urea nitrogen (BUN), total cholesterol (TC), triglycerides (TGs), low-density lipoprotein cholesterol (LDL-C), and high-density lipoprotein cholesterol (HDL-C).

Description of classified variable

According to the Chinese Comprehensive Diabetes Control Objectives (2019),¹⁸ the normal reference range: FPG:4.4 to 7.0 mmol/L; blood pressure: <130/80 mmHg; TC: <4.5 mmol/L; TGs: <1.7 mmol/L; LDL-C: <2.6 mmol/L (uncomplicated atherosclerotic cardiovascular disease) or <1.8 mmol/L (complicated atherosclerotic cardiovascular disease); HDL-C: >1.0 mmol/L (men) or >1.3 mmol/L (women); UA: upper limit <420 µmol/L for men and <360 µmol/L for women; serum creatinine (SCr): 55–133 µmol/L for men and 44–97 µmol/L for women; BUN: 2.9–7.5 mmol/L; ALT: 5 to 40 U/L; GGT: ＜40 U/L; TBil: 1.71 to 17.10 μmol/L; waist circumference ≥90 cm for men and ≥85 cm for women as central obesity; body mass index (BMI) <18.5 kg/m² is considered underweight, normal reference range 18.5 kg/m² ≤ BMI <24.0 kg/m², BMI ≥24.0 kg/m² is considered overweight, BMI ≥28.0 kg/m² is considered obese. BMI = weight/height². Estimation of glomerular filtration rate (eGFR): eGFR was calculated using the Chronic Kidney Disease Epidemiology group (CKD-EPI) formula¹⁹: for men with SCr ≤ 0.9 mg/dl, eGFR = 141 × (SCr/0.9)^−0.411 Cooperative Study G × 0.993^age; female SCr ≤ 0.7 mg/dl, eGFR = 144 × (SCr/0.7)^−0.329×0.993^age; male Scr > 0.9 ml/dl: eGFR = 144 × (SCr/0.9)^−1.209 × (0.993)^age; female SCr > 0.7 mg/dl, eGFR = 141 × (SCr/ 0.7)^−1.209×0.993^age. The Chinese guidelines for the prevention and treatment of type 2 diabetes mellitus define eGFR<60 mL/min×1.73 m² as a decrease in the GFR.²⁰ WHO defines smokers as “those who have smoked continuously or cumulatively for ≥6 months in their lifetime”; alcohol drinkers are defined as those who have consumed alcohol at least once a week for ≥6 months; and adequate exercise is defined as achieving moderate intensity exercise, with a duration of ≥30 min per exercise session and frequency ≥three times per week.

Statistical methods

The measurement data conforming to a normal distribution are expressed as ( $\bar{x}$ ±s), t tests were used for comparisons between groups, and x² tests were used for comparisons of count data. Variable screening steps: First, risk factors with P < 0.05 from the univariate analysis were entered into a multivariate logistic regression (LR). Second, 10-fold cross-validation was performed using the least absolute shrinkage and selection operator (LASSO) regression algorithm to select potential risk factors with nonzero coefficients. Finally, Boruta feature selection was used to identify key categorical variables. The optimal risk factor selection algorithm was determined based on Occam's razor method.

The data were stratified into training-testing set (80%) and independent validation set (20%) using stratified sampling. We developed and validated six machine learning algorithms: LR, artificial neural network (ANN), naïve Bayes (NB), K-nearest neighbor (K-nn), random forest (RF) and decision tree (DT). We employed GridSearch to systematically identify the optimal combination of hyperparameters for the six machine learning algorithms. Meanwhile, to avoid overfitting the accuracy of the model, we used 10-fold cross-validation to evaluate the training and test sets and applied the best model to the independent validation set. By using these six algorithms, we were able to compare their performance in evaluating the risk of developing hyperuricemia in patients with type 2 diabetes mellitus. The areas under curves (AUCs) of the six algorithms in the training-testing set were evaluated to assess model performance. Each algorithm has its own unique characteristics and applicable scenarios, and during training and testing, we can evaluate their accuracy, generalization ability and the ability to interpret the features.

EpiData 3.1 was used for double entry of data, and RStudio (version 4.2.3) was applied for statistical analysis.

Results

Comparison of demographic characteristics and biochemical indices between the normal uric acid and hyperuricemia groups

Statistically significant differences were found between the two groups in terms of age, gender, height, weight, BMI, smoking, drinking, diabetes medication use, lack of exercise, waist circumference, blood pressure, and GGT, FPG, TBil, TGs, LDL-C, HDL-C, SCr, BUN, TGs, and eGFR levels in both groups (P < 0.05; Table 1).

Table 1.

Comparison of demographic characteristics and biochemical indices between the two groups [case (%), $\bar{x}$ ±s].

Variables	Normal group	Hyperuricemia group	t/χ²	p
Variables	(n = 4477)	(n = 3766)	t/χ²	p
Age(years)	67.20±7.71	67.76±7.85	3.235	0.001
Female	1864(41.64%)	2791(74.11%)	981.914	＜0.001
Height(cm)	161.87±8.41	157.92±7.83	21.907	＜0.001
Weight(kg)	64.63±10.34	63.23±10.26	6.119	＜0.001
BMI (kg/m²)	24.63±3.43	25.31±3.36	9.013	＜0.001
Disease duration (years)	9.69±6.49	9.59±6.53	0.687	0.492
Waist circumference (cm)	86.58±8.71	87.42±9.03	4.292	＜0.001
SBP (mm Hg)	132.87±15.87	134.27±16.38	3.926	＜0.001
DBP (mm Hg)	77.77±9.67	78.77±9.53	4.724	＜0.001
ALT (U/L)	26.25±20.53	26.82±18.39	1.312	0.189
GGT (U/L)	23.66±12.11	24.81±13.42	4.072	＜0.001
FPG(mmol/L)	7.53±3.14	7.25±2.68	4.21	＜0.001
TBil (μmol/L)	13.19±8.55	11.87±5.21	8.305	＜0.001
TGs(mmol/L)	1.70±1.53	2.01±1.57	9.118	＜0.001
TC(mmol/L)	5.10±2.36	5.29±2.69	3.343	＜0.001
LDL - C(mmol/L)	2.89±0.89	2.95±0.89	3.019	＜0.001
HDL - C (mmol/L)	1.38±0.85	1.30±0.44	5.585	＜0.001
SCr (μmol/L)	71.08±34.96	72.85±31.90	2.383	0.017
BUN(mmol/L)	5.21±2.37	5.45±2.73	4.274	＜0.001
eGFR (mL/min×1.73 m²)	88.15±16.74	81.60±18.51	16.864	＜0.001
Smoking	745(16.64%)	287(7.62%)	154.755	＜0.001
Drinking	698(15.59%)	311(8.26%)	103.641	＜0.001
Lack of sport	1124(25.11%)	1102(29.26%)	17.955	＜0.001
No diabetes medication	613 (13.71%)	416 (11.05%)	13.126	＜0.001

Notes: BMI: body mass index; SBP: systolic blood pressure; DBP: diastolic blood pressure; ALT: alanine aminotransferase; GGT: gamma-glutamyl transpeptidase; TBil: total bilirubin; BUN: blood urea nitrogen; TGs: triglycerides; TC: total cholesterol; LDL-C: low-density lipoprotein cholesterol; HDL-C: high-density lipoprotein cholesterol; FPG: fasting glucose; eGFR: estimated glomerular filtration rate.

Screening factors influencing hyperuricemia in type 2 diabetes mellitus combined with hyperuricemia

Logistic regression analysis

Multifactorial LR analysis revealed that characteristics associated with hyperuricemia in the type 2 diabetes mellitus population included gender, central obesity status, diastolic blood pressure (DBP), GGT, BUN, TGs, HDL-C, eGFR, FPG and diabetes medication use (P < 0.05; Table 2).

Table 2.

Univariate and multivariate stepwise regression analysis of influencing factors of type 2 diabetes mellitus combined with hyperuricemia.

Variables	N = 8243	Uric acid status		Univariate analysis		Multivariate analysis
Variables	N = 8243	Normal group	Hyperuricemia group	χ²	p	OR (95%CI)	p
Age
<65^a	2226	1195(26.7)	1031(27.4)	0.452	0.501
≥65	6017	3282(73.3)	2735(72.6)	0.452	0.501
Gender
Male^a	3588	2613(58.4)	975(25.9)	876.27	<0.001	4.15[3.70,4.66]	<0.001
Female	4655	1864(41.6)	2791(74.1)	876.27	<0.001	4.15[3.70,4.66]	<0.001
Tobacco use
No^a	7211	3732(83.4)	3479(92.4)	151.12	<0.001	0.94[0.78,1.12]	0.48
Yes	1032	745(16.6)	287(7.6)	151.12	<0.001	0.94[0.78,1.12]	0.48
Drinking alcohol
No^a	7234	3779(84.4)	3455(91.7)	101.7	<0.001	1.10[0.93,1.31]	0.261
Yes	1009	698(15.6)	311(8.3)	101.7	<0.001	1.10[0.93,1.31]	0.261
Sport
Adequate^a	6017	3353(74.9)	2664(70.7)	17.71	<0.001	1.08[0.97,1.20]	0.165
Lacking	2226	1124(25.1)	1102(29.3)	17.71	<0.001	1.08[0.97,1.20]	0.165
Waist circumference (cm)
Male<90 / Female<85^a	4326	2602(58.1)	1724(45.8)	124.43	<0.001	1.38[1.23,1.55]	<0.001
Male ≥90 / Female ≥85	3917	1875(41.9)	2042(54.2)	124.43	<0.001	1.38[1.23,1.55]	<0.001
BMI (kg/m²)
<24.0^a	3408	1996(44.6)	1412(37.5)	42.107	<0.001	1.04[0.93,1.17]	0.464
≥24.0	4835	2481(55.4)	2354(62.5)	42.107	<0.001	1.04[0.93,1.17]	0.464
Disease duration (years)
<10^a	4710	2544(56.8)	2166(57.5)	0.371	0.543
≥10	3533	1933(43.2)	1600(42.5)	0.371	0.543
Diabetes medication use
Use^a	7214	3864(86.3)	3350(89.0)	12.867	<0.001	0.75[0.65,0.86]	<0.001
No	1029	613(13.7)	416(11.0)	12.867	<0.001	0.75[0.65,0.86]	<0.001
SBP (mm Hg)
<130^a	3224	1792(40.0)	1432(38.0)	3.36	0.067
≥130	5019	2685(60.0)	2334(62.0)	3.36	0.067
DBP (mm Hg)
<80^a	4270	2420(54.1)	1850(49.1)	19.717	<0.001	1.30[1.18,1.44]	<0.001
≥80	3973	2057(45.9)	1916(50.9)	19.717	<0.001	1.30[1.18,1.44]	<0.001
ALT (U/L)
<40^a	7292	3981(88.9)	3311(87.9)	1.919	0.166
≥40	951	496(11.1)	455(12.1)	1.919	0.166
GGT (U/L)
<40^a	7765	4258(95.1)	3507(93.1)	14.402	<0.001	1.38[1.13,1.70]	0.002
≥40	478	219(4.9)	259(6.9)	14.402	<0.001	1.38[1.13,1.70]	0.002
TBil (μmol/L)
<17.1^a	7064	3749(83.7)	3315(88.0)	30.296	<0.001	0.97[0.85,1.12]	0.711
≥17.1	1179	728(16.3)	451(12.0)	30.296	<0.001	0.97[0.85,1.12]	0.711
BUN (mmol/L)
<7.5^a	7643	4213(94.1)	3430(91.1)	27.288	<0.001	1.42[1.16,1.74]	0.001
≥7.5	600	264(5.9)	336(8.9)	27.288	<0.001	1.42[1.16,1.74]	0.001
TGs (mmol/L)
<1.7^a	2447	1447(32.3)	1000(26.6)	175.42	<0.001	1.71[1.54,1.89]	<0.001
≥1.7	5796	3030(67.7)	2766(73.4)	175.42	<0.001	1.71[1.54,1.89]	<0.001
TC (mmol/L)
<4.5^a	4519	2753(61.5)	1766(46.9)	32.319	<0.001	0.99[0.87,1.12]	0.835
≥4.5	3724	1724(38.5)	2000(53.1)	32.319	<0.001	0.99[0.87,1.12]	0.835
LDL - C (mmol/L)
<2.6^a	2757	1595(35.6)	1162(30.9)	20.707	<0.001	1.13[1.00,1.28]	0.046
≥2.6	5486	2882(64.4)	2604(69.1)	20.707	<0.001	1.13[1.00,1.28]	0.046
HDL - C (mmol/L)
Male>1.0 / Female>1.3^a	5745	3400(75.9)	2345(62.3)	180.48	<0.001	1.23[1.10,1.37]	<0.001
Male ≤1.0 / Female ≤1.3	2498	1077(24.1)	1421(37.7)	180.48	<0.001	1.23[1.10,1.37]	<0.001
FPG (mmol/L)
<7.0^a	4529	2403(53.7)	2126(56.5)	6.265	0.012	0.77[0.70,0.85]	<0.001
≥7.0	3714	2074(46.3)	1640(43.5)	6.265	0.012	0.77[0.70,0.85]	<0.001
eGFR (mL/min×1.73 m²)
≥ 60^a	613	204(4.6)	409(10.9)	117.16	<0.001	2.59[2.12,3.19]	<0.001
<60	7630	4273(95.4)	3357(89.1)	117.16	<0.001	2.59[2.12,3.19]	<0.001

Notes: ^areference; BMI: body mass index; SBP: systolic blood pressure; DBP: diastolic blood pressure; ALT: alanine aminotransferase; GGT: gamma-glutamyl transpeptidase; TBil: total bilirubin; BUN: blood urea nitrogen; TGs: triglycerides; TC: total cholesterol; LDL-C: low-density lipoprotein cholesterol; HDL-C: high-density lipoprotein cholesterol; FPG: fasting glucose; eGFR: estimated glomerular filtration rate. eGFR and SCr, two similar variables, were selected for eGFR. OR = odds ratio; CI = confidence interval.

LASSO regression

A minimum absolute shrinkage and selection operator (LASSO) regression algorithm was used for 10-fold cross-validation to select potential influencing factors with nonzero coefficients (Figure 2).

Figure 2.

Feature variable selection using least absolute shrinkage and selection operator (LASSO) regression. (a) A LASSO-based ordinal logistic model with 10-fold cross-validation based on the minimal mean squared error (MSE) was employed to find the optimal parameter (λ). The log (λ_1se) of - 4.05 and λ_1se of 0.017 were considered optimal. (b) The resulting nine influencing factors with nonzero coefficients were identified based on the log (λ_1se) value. LASSO: least absolute shrinkage and selection operator.

Boruta characteristics

Variable importance identification results: one potentially important variable (importance score not significantly different from the best shadow variable score), exercise status; 18 important variables, including gender, DBP, central obesity, tobacco use, alcohol use, diabetes medication use, ALT, AGE, GGT, BMI, BUN, FPG, TGs, TC, TBil, LDL-C, HDL-C, and eGFR; two insignificant variables, including duration of diabetes and systolic blood pressure (SBP) (Figure 3).

Figure 3.

Feature selection based on the Boruta characteristics. The horizontal axis is the name of each variable, and the vertical axis is the Z value of each variable. The box plot shows the Z value of each variable during model calculation. The green boxes represent the first 18 important variables, yellow represents tentative variables and red represents unimportant variables.

Performance comparison of three variable screening methods

The optimal influencing factor selection algorithm was determined according to Occam's razor method, and the LASSO regression algorithm was optimal (Table 3). LASSO regression models showed that nine statistically significant factors were obtained. Hyperuricemia in patients with type 2 diabetes mellitus was positively associated with female gender, waist circumference ≥90 cm in men (≥85 cm in women), DBP ≥80 mmHg, GGT ≥40 U/L, BUN ≥7.5 mmol/L, TGs ≥1.7 mmol/L, HDL-C ≤1.0 mmol/L in men (≤ 1.3 mmol/L) and eGFR<60 (mL/min×1.73 m²) were positively correlated (P < 0.05); they were negatively correlated with FPG≥7.0 mmol/L and no diabetes medication (P < 0.05; Figure 4).

Figure 4.

Forest plot of odds ratios (ORs) for influencing factors included in the risk assessment model.

Table 3.

Performance of three feature variable selection methods.

Feature selection method	No. of feature variables	AUC (95% CI)	Accuracy (95% CI)	Nagelkerke R²	RMSE	BIC
Univariate and multivariate stepwise regression	11	0.732(0.721–0.743)	0.688(0.688–0.688)	0.121	0.456	10093.65
LASSO regression	9	0.726(0.715–0.737)	0.686(0.686–0.686)	0.146	0.460	−12600.87
Boruta and stepwise regression	18	0.732(0.721–0.743)	0.688(0.687–0.688)	0.122	0.456	10153.99

LASSO: least absolute shrinkage and selection operator; AUC: areas under curve.

Risk assessment model building and evaluation

LR, ANN, K-nn, NB, DT and RF were used to build the risk assessment models. By using Grid-optimization, the hyperparameters of the optimal ANN model were as follows: hidden = c(7), Activation = 0.01, stepmax = logistic, Alpha = 0.0001, Learning_rate = 0.01, Power_t = 0.5, Max_iter = 1000, Tol = 1.0E-4, Learning_rate_int = 0.001. The hyperparameters of the remaining five risk assessment models are detailed in Supplemental material Table S2. The AUC and 95% CI of each algorithm in both sets were obtained using 10-fold cross-validation in the training-testing set and tested with an independent test set, as shown in Figure 5b and c. According to Table 4, the ANN model performed best with the highest AUC value of 0.736 on the test set, and the accuracy, sensitivity, specificity, precision, F1-score and KAPPA are 0.683, 0.650, 0.722, 0.736, 0.690 and 0.369, respectively.

Figure 5.

Development and application of a model for evaluating the risk of type 2 diabetes mellitus combined with hyperuricemia. (a) The ANN risk assessment model. (b, c) Receiver-operating characteristic curves showing the performance of the risk assessment model in evaluating diabetes mellitus combined with hyperuricemia in the (b) training-testing set and (c) independent validation set. ANN: artificial neural network.

Table 4.

The performance of six risk assessment models of type 2 diabetes mellitus combined with hyperuricemia based on machine learning.

Algorithm	Discrimination tests
Algorithm	AUC (95% CI)	Accuracy (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	Precision (95% CI)	F1-score (95% CI)	KAPPA (95% CI)
Logistic regression	0.728 (0.703–0.752)	0.692 (0.691–0.692)	0.708 (0.679–0.738)	0.672 (0.638–0.706)	0.720 (0.690–0.749)	0.714 (0.684–0.743)	0.380 (0.335–0.425)
Artificial neural network	0.736 (0.712–0.760)	0.683 (0.683–0.684)	0.650 (0.619–0.682)	0.722 (0.690–0.754)	0.736 (0.705–0.767)	0.690 (0.659–0.722)	0.369 (0.324–0.413)
Naïve Bayes	0.724 (0.699–0.748)	0.686 (0.687–0.687)	0.673 (0.642–0.703)	0.703 (0.670–0.735)	0.729 (0.698–0.759)	0.700 (0.669–0.730)	0.372 (0.328–0.417)
K-nearest neighbor	0.680 (0.656–0.705)	0.660 (0.660–0.660)	0.657 (0.626–0.688)	0.664 (0.630–0.689)	0.699 (0.668–0.730)	0.677 (0.646–0.708)	0.319 (0.273–0.365)
Random forest	0.733 (0.709–0.757)	0.687 (0.687–0.687)	0.723 (0.694–0.752)	0.644 (0.610–0.678)	0.707 (0.678–0.737)	0.715 (0.686–0.744)	0.368 (0.323–0.413)
Decision tree	0.701 (0.677–0.725)	0.662 (0.662–0.662)	0.603 (0.571–0.635)	0.732 (0.700–0.763)	0.728 (0.696–0.760)	0.660 (0.627–0.692)	0.330 (0.285–0.374)

AUC: areas under curve.

An intuitive comparison of the consistency risk assessment model between the calibration curves and the ideal calibration curve was made, and the consistency was further evaluated in terms of calibration slope (ideal value of 1) and Brier score (ideal value of 0, >0.3 value indicates poor calibration). ANN has the smallest Brier score; therefore, the ANN algorithm is the best algorithm. See Table 5 and Figure 6.

Figure 6.

Calibration curves for independent validation the stability of six risk assessment models in the study.

Table 5.

Consistency evaluation of the six risk assessment models.

Algorithm	Calibration
Algorithm	Brier score	Slope	Intercept
Logistic regression	0.209	1.031	−0.020
Artificial neural network	0.206	0.942	0.088
Naïve Bayes	0.211	0.892	−0.002
K-nearest neighbor	0.297	0.366	0.446
Random forest	0.207	1.007	−0.021
Decision tree	0.216	0.941	−0.007

To determine the clinical usefulness of the models, decision curve analysis and clinical impact curve analysis were performed on the risk assessment models. The clinical decision curves (Figure 7) showed that when clinical decisions were made using the LG, ANN, NB, K-nn, RF and DT risk assessment models, the net benefit of diabetes mellitus combined with hyperuricemia was greater than that of the “no treatment” or “all treatment” regimens. The thresholds were 0.78, 0.79, 0.78, 0.78, 0.78, 0.78 and 0.77.

Figure 7.

Decision curve analysis of six risk assessment models.

The clinical impact curve (Figure 8) analysis shows the clinical effectiveness of the six risk assessment models. When the threshold probabilities were greater than 63%, 62%, 62%, 71%, 69% and 67%, respectively, the LG, ANN, NB, K-nn, RF and DT models were determined to be a high match between the population at high risk for diabetes combined with hyperuricemia and the population actually experiencing diabetes combined with hyperuricemia, confirming the high clinical efficiency of this risk assessment model.

Figure 8.

Clinical impact curve analysis of six risk assessment models.

Discussion

In the results of this study, we estimated the prevalence of hyperuricemia in the population with type 2 diabetes mellitus in Fuzhou, obtained the influencing factors for developing hyperuricemia in this population, and established and evaluated a risk assessment model for hyperuricemia in patients with type 2 diabetes mellitus using various methods. In our study population, about 45.68% of patients with type 2 diabetes mellitus suffered from hyperuricemia, which is a very high percentage. Although it has been shown that type 2 diabetics are more prone to hyperuricemia than healthy people, and the prevalence in their study population ranged from 10% to 40%,^21,22 we cannot rule out the possibility that there may be a bias in our selected population.

In this study, we screened variables using LR, LASSO regression and Boruta features. Based on the performance comparison of the three algorithms and Occam's razor rule of machine learning, we finally identified nine variables screened by LASSO regression that influence the development of hyperuricemia in patients with type 2 diabetes mellitus. Among the risk factors were female gender, central obesity, elevated SBP, abnormal BUN, abnormal TGs, abnormal HDL-C and decreased eGFR.

Previous studies have confirmed that oestrogens^23,24 are effective in increasing UA excretion from the kidneys to the extent that there is a difference in the prevalence of hyperuricemia between women and men. Similarly, in our study population, women (33.86%) were more likely to develop hyperuricemia than men (11.83%) and also had a 4.387 times higher risk of developing hyperuricemia than men (95% Cl: 3.898–4.937). Women in the type 2 diabetes mellitus population, then, should be especially aware of their UA levels.

At the same time, waist circumference, blood pressure, BUN, TGs, HDL-C and eGFR are some of the common markers used in physical examination and their abnormalities are important risk factors for hyperuricemia in the type 2 diabetic population. The possible mechanism of elevated UA in type 2 diabetic patients who are centrally obese is due to dysregulation of adipocytokines and chronic low-grade inflammation.^25,26 In addition, weight loss in obese individuals is accompanied by reduced UA levels and xanthine oxidoreductase activity, which is responsible for the breakdown of hypoxanthine and xanthine into UA.²⁷

Elevated blood pressure in patients with type 2 diabetes mellitus may result in renal vascular resistance accompanied by a decrease in renal blood flow.²⁸ With this decrease in renal blood flow, proximal sodium and UA absorption increases, which may contribute to increased serum UA levels.^29,30 Secondly, microvascular damage associated with hypertension may result in local tissue ischemia.³⁰ Under ischemic conditions, the degradation of adenosine triphosphate to adenine and xanthine and the concomitant increase in the conversion of xanthine dehydrogenase to xanthine oxidase lead to increased UA production.^31,32

UA, urea nitrogen and glomerular filtration rate are all measures of kidney function,³³ and this study also shows a correlation between the three. This also suggests that abnormalities in the function of an organ may be accompanied by abnormalities in all the indicators associated with it.

Typically, abnormalities in one of TC, TGs, HDL-C and LDL-C are diagnosed as dyslipidemia. In our study, dyslipidemia (TGs and HDL-C) was also a risk factor for hyperuricemia in a type 2 diabetic population, and possible mechanisms underpinning these associations include, on the one hand, the possibility that elevated levels of TGs may accelerate the overproduction of UA via the classical free fatty acid metabolic pathway.^34,35 On the other hand, hypertriglyceridemia leads to the overproduction of free fatty acids, whose biosynthesis in the liver is associated with the de novo synthesis of purines, thus accelerating UA production.³⁶ In addition, HDL cholesterol has anti-inflammatory, antioxidant and anti-apoptotic properties,³⁷ and studies have also found that HDL cholesterol reduces the inflammatory response induced by urate crystals, suggesting that HDL cholesterol is involved in UA-induced inflammatory responses.³⁸

Interestingly, protective factors include nonuse of diabetic medications and FPG ≥7.0 mmol/L, but studies^39,40 have shown a reduction in blood UA levels in diabetes patients on glucose-lowering medications, which may be related to the renoprotective effects of glucose-lowering medications.^41,42 For example, SGLT-2 inhibitors not only promote anti-inflammatory and antifibrotic pathways and improve renal oxygenation, but also reduce glomerular hypertension and hyperfiltration rate, thereby increasing UA excretion. However, some studies^5,43 have also shown that poor medication adherence and irrational use of medication may have an effect on the metabolism and excretion of UA, leading to elevated UA levels, which may lead to hyperuricemia. And we did not have the patient's medication information and medication intake in hand. Therefore, with such unexpected results, it shows us the direction of our future work. It has also been shown⁴⁴ that when glycemic control is poor in diabetes patients, UA levels are reduced at this time due to the permeability of glucose, which causes an increase in urinary glucose excretion, which in turn leads to a competitive inhibition of UA reabsorption, which may account for the findings of this investigation.

In conclusion, in such a population, women may be more susceptible to hyperuricemia. Patients with type 2 diabetes mellitus should also strengthen the monitoring and management of risk factors such as abdominal obesity, elevated blood pressure, renal hypoplasia, and dyslipidemia, in order to minimize the harm caused by diabetes combined with hyperuricemia.

In this study, after obtaining nine influencing factors affecting the development of hyperuricemia in a population with type 2 diabetes mellitus, risk assessment models were developed using LR, ANN, DT, K-nn, NB and RF. In the study of risk assessment models, ANN had the best overall performance, not only obtaining better AUC but also the highest accuracy in the accuracy assessment, but its shortcoming was the average performance in terms of accuracy and sensitivity. The present study is similar to the ANN model reported by Kalliopi Dalakleidi et al.,⁴⁵ where ANN outperformed other algorithms such as NB, DT and LR models, which may be attributed to its better computational power in dealing with imbalances in medical datasets. The ANN also has strong self-learning capabilities and is a useful tool for evaluating complications of diabetes mellitus,⁴⁶ tuberculosis⁴⁷ and hypertension.⁴⁸ The balanced performance of the six models in the DCA and CIC analyses all had good clinical utility, but the ANN model performed best in the Brier score. In conclusion, our findings support the use of the ANN model as the best option for evaluating the risk of hyperuricemia in patients with type 2 diabetes.

Strengths and limitations

In the present study, although the variables we chose are very common, we believe that their role in evaluating risk should not be overlooked. This is because in China, whether in hospitals, medical check-up centers, or in the community, the health check-up items and questionnaires asked by these variables are relatively basic and economical. Therefore, our initial intention is to use an efficient and scientific machine learning method to evaluate the risk factors of hyperuricemia in type 2 diabetes mellitus patients and build a corresponding risk model based on such basic and economical data. Through such an approach, clinicians will be able to more accurately evaluate patients’ conditions based on routine physical examination data and questionnaire results, and then through personalized management and intervention, the risk of hyperuricemia can be effectively controlled, thus improving treatment outcomes and patient satisfaction. In addition, it can help patients with type 2 diabetes mellitus better understand which indicators they need to pay particular attention to based on their own physical examination results, so as to guide their lives.

However, there are some shortcomings in this study. The ANN algorithm obtained the highest AUC in this study, but its accuracy and sensitivity performance were not optimal. Therefore, further prospective studies are needed to improve the accuracy and generalize the use of the model and validate our conclusions. This study is a cross-sectional investigation and cannot determine the causal relationship between type 2 diabetes mellitus and hyperuricemia. Finally, our study did not investigate the dietary factors in this population.

Conclusions

We developed an ANN model to better evaluate the risk of hyperuricemia in the type 2 diabetes population. In the type 2 diabetes population, women should pay particular attention to their UA levels, and type 2 diabetics should not neglect their obesity level, blood pressure, kidney function and lipid profile during their regular medical check-ups, in order to do their best to avoid the risks associated with the combination of type 2 diabetes and hyperuricemia.

Supplemental Material

sj-docx-1-dhj-10.1177_20552076241241381 - Supplemental material for Evaluating the risk of developing hyperuricemia in patients with type 2 diabetes mellitus using least absolute shrinkage and selection operator regression and machine learning algorithm

Supplemental material, sj-docx-1-dhj-10.1177_20552076241241381 for Evaluating the risk of developing hyperuricemia in patients with type 2 diabetes mellitus using least absolute shrinkage and selection operator regression and machine learning algorithm by Qingquan Chen, Haiping Hu, Qing He, Xinfeng Huang, Huanhuan Shi, Xiangyu Cao, Xiaoyang Zhang and Youqiong Xu in DIGITAL HEALTH

Footnotes

Authors’ contributions

YQX led the conceptualization, methodology design, and managed data curation. XYZ was involved in the project's conceptualization, manuscript editing, supervision, administration, and funding acquisition. QQC and HPH both contributed to writing and editing the manuscript. XYC was involved in validating the results, formal analysis. XFH and HHS both provided supervision and writing. QQC, HPH, and QH were responsible for data curation and project supervision. All authors read and approved the final manuscript.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Ethics approval and consent to participate

This study was reviewed by the Ethics Committee of the Fuzhou Center for Disease Control and Prevention (approval number: 2019006 and 2022002). Informed consent was obtained from all respondents for this study.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Fuzhou Science and Technology Program, (grant number No. 2019-SZ-63, No. 2022-S-032).

Author information

Qingquan Chen and Haiping Hu contributed equally to this work.

Availability of data and material

The datasets used and/or analyzed during the current study available from the corresponding author on reasonable request.

Informed consent statement

Informed consent was obtained from all subjects involved in the study.

Availability of data and material

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Guarantor

Youqiong Xu.

ORCID iD

Youqiong Xu

Supplemental material

Supplemental material for this article is available online.

References

International Diabetes Federation. IDF diabetes atlas 10th edition. https://diabetesatlas.org/data/en/country/42/cn.html.

Shanshan

Zhihua

Chengcheng

, et al. Analysis of the burden of disease and predicted incidence of diabetes mellitus in China from 1990 to 2019. Chin General Med 2023; 26: 2013–2019.

Jiang

Zhang

Liu

, et al. Prevalence of diabetes in patients with hyperuricaemia and gout: a systematic review and meta-analysis. Curr Diab Rep 2023; 23: 103–117.

Liu

Zhang

. Analysis of factors influencing asymptomatic hyperuricaemia and disease knowledge in the middle-aged and young physical examination population in Daqing city. Evidence Based Nurs 2021; 7: 894–901.

Dehlin

Jacobsson

Roddy

. Global epidemiology of gout: prevalence, incidence, treatment patterns and risk factors. Nat Rev Rheumatol 2020; 16: 380–390.

Kodama

Saito

Yachi

, et al. Association between serum uric acid and development of type 2 diabetes. Diabetes Care 2009; 32: 1737–1742.

Sun

Chang

Bian

, et al. Adipose tissue insulin resistance is positively associated with serum uric acid levels and hyperuricemia in Northern Chinese adults. Front Endocrinol (Lausanne) 2022; 13: 835154.

Vuorinen-Markkola

Yki-Järvinen

. Hyperuricemia and insulin resistance. J Clin Endocrinol Metab 1994; 78: 25–29.

Cui

, et al. Hyperuricemia predisposes to the onset of diabetes via promoting pancreatic β-cell death in uricase-deficient male mice. Diabetes 2020; 69: 1149–1163.

10.

Jia

Xing

Ding

, et al. Hyperuricemia causes pancreatic β-cell death and dysfunction through NF-κB signaling pathway. PLoS One 2013; 8: e78284.

11.

Andrade

Kang

Greffin

, et al. Serum uric acid and disorders of glucose metabolism: the role of glycosuria. Braz J Med Biol Res 2014; 47: 917–923.

12.

Sharaf El Din

UAA

Salem

Abdulazim

. Uric acid in the pathogenesis of metabolic, renal, and cardiovascular diseases: a review. J Adv Res 2017; 8: 537–548.

13.

Lee

Hsieh

, et al. Hyperuricemia and diabetes mellitus when occurred together have higher risks than alone on all-cause mortality and end-stage renal disease in patients with chronic kidney disease. BMC Nephrol 2022; 23: 157.

14.

Wei

Sun

Shan

, et al. Environmental chemical exposure dynamics and machine learning-based prediction of diabetes mellitus. Sci Total Environ 2022; 806: 150674.

15.

Cai

, et al. Machine learning for predicting the 3-year risk of incident diabetes in Chinese adults. Front Public Health 2021; 9: 626331.

16.

Lee

Choe

Park

. Exploration of machine learning for hyperuricemia prediction models based on basic health checkup tests. J Clin Med 2019; 8: 72.

17.

Ichikawa

Saito

Ujita

, et al. How can machine-learning methods assist in virtual screening for hyperuricemia? A healthcare machine-learning approach. J Biomed Inform 2016; 64: 20–24.

18.

Chinese Medical Association, Chinese Medical Association Journal , Chinese Medical Association Branch of General Practice , et al. Guidelines for primary diagnosis and treatment of type 2 diabetes mellitus (practice edition, 2019). Chin J General Pract 2019; 18: 810–818.

19.

Levey

Stevens

Schmid

, et al. A new equation to estimate glomerular filtration rate. Ann Intern Med 2009; 150: 604–612.

20.

Chinese Guidelines for prevention and treatment of type 2 diabetes mellitus (2020 edition) (II). Chin J Pract Internal Med 2021; 41: 757–784.

21.

Mantovani

Rigolon

Pichiri

, et al. Hyperuricaemia is associated with an increased prevalence of atrial fibrillation in hospitalized patients with type 2 diabetes. J Endocrinol Invest 2016; 39: 159–167.

22.

Wang

Chen

Lei

, et al. Prevalence and determinants of hyperuricemia in type 2 diabetes mellitus patients with central obesity in Guangdong Province in China. Asia Pac J Clin Nutr 2013; 22: 590–598.

23.

Antón

García Puig

Ramos

, et al. Sex differences in uric acid metabolism in adults: evidence for a lack of influence of estradiol-17 beta (E2) on the renal handling of urate. Metabolism 1986; 35: 343–348.

24.

Lin

Lee

, et al. Sex-specific association of hyperuricemia with cardiometabolic abnormalities in a military cohort: the CHIEF study. Medicine (Baltimore) 2020; 99: e19535.

25.

Martyniak

Masternak

. Changes in adipose tissue cellular composition during obesity and aging as a cause of metabolic dysregulation. Exp Gerontol 2017; 94: 59–63.

26.

Trayhurn

. Hypoxia and adipose tissue function and dysfunction in obesity. Physiol Rev 2013; 93: 1–21.

27.

Nielsen

Bartels

Henriksen

, et al. Weight loss for overweight and obese individuals with gout: a systematic review of longitudinal studies. Ann Rheum Dis 2017; 76: 1870–1882.

28.

Ponnuchamy

Khalil

. Cellular mediators of renal vascular dysfunction in hypertension. Am J Physiol Regul Integr Comp Physiol 2009; 296: R1001–R1018.

29.

Ryu

Chang

Zhang

, et al. A cohort study of hyperuricemia in middle-aged South Korean men. Am J Epidemiol 2012; 175: 133–143.

30.

Johnson

Kivlighn

Kim

, et al. Reappraisal of the pathogenesis and consequences of hyperuricemia in hypertension, cardiovascular disease, and renal disease. Am J Kidney Dis 1999; 33: 225–234.

31.

Maiuolo

Oppedisano

Gratteri

, et al. Regulation of uric acid metabolism and excretion. Int J Cardiol 2016; 213: 8–14.

32.

Mazzali

Kanbay

Segal

, et al.

Uric acid and hypertension: cause or effect?

Curr Rheumatol Rep 2010; 12: 108–117.

33.

Ejaz

Nakagawa

Kanbay

, et al. Hyperuricemia in kidney disease: a major risk factor for cardiovascular events, vascular calcification, and renal damage. Semin Nephrol 2020; 40: 574–585.

34.

Matsuura

Yamashita

Nakamura

, et al. Effect of visceral fat accumulation on uric acid metabolism in male obese subjects: visceral fat obesity is linked more closely to overproduction of uric acid than subcutaneous fat obesity. Metabolism 1998; 47: 929–933.

35.

Seyed-Sadjadi

Berg

Bilgin

, et al.

Visceral fat mass: is it the link between uric acid and diabetes risk?

Lipids Health Dis 2017; 16: 142.

36.

Balasubramanian

. Uric acid or 1-methyl uric acid in the urinary bladder increases serum glucose, insulin, true triglyceride, and total cholesterol levels in wistar rats. ScientificWorldJournal 2003; 3: 930–936.

37.

Nicholls

Nelson

. HDL and cardiovascular disease. Pathology 2019; 51: 142–147.

38.

Scanu

Luisetto

Oliviero

, et al. High-density lipoproteins inhibit urate crystal-induced inflammation in mice. Ann Rheum Dis 2015; 74: 587–594.

39.

Hussain

Elahi

Hussain

, et al. Sodium-glucose cotransporter-2 (SGLT-2) attenuates serum uric acid (SUA) level in patients with type 2 diabetes. J Diabetes Res 2021; 2021: 9973862.

40.

Ueno

. Urate-lowering therapy ameliorates kidney function in type 2 diabetes patients with hyperuricemia. J Clin Med Res 2017; 9: 1007–1012.

41.

Eliseev

Panevin

Zhelyabina

, et al. Advantages of the use of metformin in patients with impaired uric acid metabolism. Ter Arkh 2021; 93: 71520.

42.

Heerspink

HJL

Kosiborod

Inzucchi

, et al. Renoprotective effects of sodium-glucose cotransporter-2 inhibitors. Kidney Int 2018; 94: 26–39.

43.

Jacobs

Stern

. An unusual case of gout in the wrist: the importance of monitoring medication dosage and interaction. A case report. Chiropr Osteopat 2007; 15: 16.

44.

Chen

Qiu

Guo

, et al. Increased urinary glucose excretion is associated with a reduced risk of hyperuricaemia. Diabet Med 2019; 36: 902–907.

45.

Dalakleidi

Zarkogianni

Thanopoulou

, et al. Comparative assessment of statistical and machine learning techniques towards estimating the risk of developing type 2 diabetes and cardiovascular complications. Expert Syst 2017; 34: e12214.

46.

Tan

Seng

JJB

Kwan

, et al. Evaluation of machine learning methods developed for prediction of diabetes complications: a systematic review. J Diabetes Sci Technol 2023; 17: 474–489.

47.

Elveren

Yumuşak

. Tuberculosis disease diagnosis using artificial neural network trained with genetic algorithm. J Med Syst 2011; 35: 329–332.

48.

Huang

Yue

, et al. Evaluating the risk of hypertension using an artificial neural network method in rural residents over the age of 35 years in a Chinese area. Hypertens Res 2010; 33: 722–726.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.02 MB