Development and Internal Validation of a Risk Prediction Model for Carotid Atherosclerosis in the Hyperuricemia Population

Introduction

Stroke ranks second among the leading causes of disability and death globally.¹ According to the 2019 stroke burden in China, ischemic stroke affected 2.87 million people and caused 1.03 million deaths.² Carotid atherosclerosis (CAS) is a potential cause of ischemic stroke, and approximately 18–25% of ischemic strokes are attributed to thromboembolism caused by CAS.³ Hyperuricemia (HUA) is a metabolic condition characterized by increased uric acid (UA) synthesis or decreased excretion. Over the past few years, HUA incidence has increased globally due to increased intake of purine-rich foods. An updated study showed that 16.4% of mainland Chinese individuals had HUA.⁴ Several studies have demonstrated that HUA is a risk factor for CAS^5–7 and that UA could be a potential therapeutic target. According to a systematic meta-analysis, HUA can increase the risk of atherosclerosis (AS) by at least 50%.⁸ In addition, the risk of stroke increased by 10% for each 1 mg/dL increase in UA level.⁹ UA maintains the AS process by interfering with lipid metabolism, decreasing the ability of endothelial cells to produce nitric oxide, accelerating the growth of arterial smooth muscle cells, and overwhelms inflammation.¹⁰

With the advancement of precision medicine, more studies utilized risk prediction models to predict the occurrence of certain diseases and their complications to prevent or reduce adverse prognostic events. For example, Huang et al¹¹ constructed a CAS risk model among the Chinese population and Feng et al¹² developed a risk model for CAS in patients with type 2 diabetes. All the above models exhibited excellent clinical predictive values. However, there are no published reports on CAS risk prediction models in a population with HUA. The purpose of this research was to develop a CAS risk prediction model using statistical algorithms in the HUA population that could contribute to early CAS identification.

Methods

Participants

This retrospective study included 3623 HUA individuals who underwent health examinations, including carotid ultrasonography, at the Zhenhai Lianhua Hospital in Ningbo, China, in 2020. The sample size was calculated based on the rule of thumb recommendation for logistic regression analysis, which suggests using at least 10 cases for each predictor variable.¹³ Relevant clinical data were obtained from the electronic medical record system. Inclusion criteria: participants aged ≥ 18 years and who satisfy the diagnostic criteria for HUA, defined by a serum UA level > 420 μmol/L in males and > 360 μmol/L in females.¹⁴ Exclusion criteria: individuals with severe abnormalities in kidney function, those who have taken uric acid-lowering or lipid-lowering medications in the past 6 months, currently have an acute infection, malignant tumors, psychiatric disorders, or are pregnant. Each data point was checked to ensure the accuracy and completeness of data collection. Those with missing severe data (over 20% of the overall) were excluded, and multiple interpolations were used to fill in those with fewer missing data (less than 20% of the overall) (Figure S1). After processing, complete data were obtained from 3579 patients with HUA. CAS was defined as increased carotid intima-media thickness ≥ 1 mm or the presence of plaques,¹⁵ and carotid ultrasound was performed independently by two skilled sonographers. A flow diagram of the research design is shown in Figure 1.

Figure 1 Flow diagram of the research design. HUA, hyperuricemia.

Clinical Baseline Data

The demographic and clinical data for this study were primarily based on their availability in the electronic medical record system, including gender, age, body mass index, systolic blood pressure (SBP), diastolic blood pressure (DBP), heart rate (HR), white blood cell count (WBC), neutrophil count, eosinophil count, basophil count (BAS), lymphocyte count (LYM), red blood cell count (RBC), hemoglobin (HGB), red blood cell distribution width (RDW), mean red blood cell volume (MCV), platelet count (PLT), platelet distribution width, mean platelet volume, alanine aminotransferase (ALT), aspartate aminotransferase (AST), total bilirubin (T-BIL), direct bilirubin (D-BIL), indirect bilirubin (I-BIL), total protein (TP), albumin (ALB), gamma-glutamyl transpeptidase (GGT), total bile acids (TBA), blood urea nitrogen (BUN), serum creatinine (SCR), UA, fasting blood glucose (FBG), total cholesterol, triglycerides (TG), high-density lipoprotein (HDL), low-density lipoprotein, apolipoprotein-A (Apo-A), apolipoprotein-B, homocysteine (HCY), thyroid stimulating hormone, total triiodothyronine (TT3), total tetraiodothyronine, free triiodothyronine (FT3), free tetraiodothyronine (FT4).

Statistical Analysis

Continuous variables were expressed as mean and standard deviation for normally distributed data and as median and interquartile range for skewed data. Percentages (%) were used to describe categorical variables. The data were analyzed using the R software package (version 4.1.2; https://www.R-project.org). All tests were two-sided, and statistical significance was defined as P < 0.05.

Selected individuals were randomly assigned to the training and internal validation sets at a ratio of 7:3.¹¹ Multivariable logistic regression analysis was performed to identify the independent risk factors. The risk model’s characteristic variables were screened by the least absolute shrinkage and selection operator (LASSO) combined with 10-fold cross-validation and the model was presented using a nomogram. The discriminative ability, calibration, and clinical utility of the model were estimated by the receiver operating characteristic (ROC) curve, calibration curves, and decision curve analysis (DCA), respectively.

Results

Characteristics of the Study Population

A total of 3579 individuals were included in this study, comprising 2676 HUA participants without CAS and 903 HUA participants with CAS. The proportion of CAS in the HUA population was 25.2%. In the univariate analysis, a higher percentage of males was observed in the control group compared to the CAS group (86.1% vs 75.9%, P < 0.001). The median age of the CAS group was higher than that of the control group [median age 67 (IQR:58–77) vs 43 (IQR:32–54), P < 0.001]. Additionally, statistically significant differences were observed between the groups in terms of blood pressure (SBP, DBP), blood cell counts (WBC, BAS, LYM, RBC, HGB, RDW, MCV, PLT), liver function (ALT, AST, T-BIL, I-BIL, TP, ALB, TBA), renal function (BUN, SCR, UA), metabolic markers (FBG, TG, HDL, Apo-A, HCY), and thyroid function (TT3, FT3, FT4) (P < 0.05) (Table 1). Randomization sampling was used to distribute 2507 and 1072 individuals to the training and internal validation sets, respectively, and the basic characteristics of participants between the two sets did not differ (P > 0.05) (Table 1).

Table 1 Characteristics of the Study Population

Independent Risk Factors

Candidate variables with P < 0.05 in the univariate analysis (Table 1) were included in the independent risk factors screening. To avoid the impact of multicollinearity on the outcome, we computed the variance inflation factor (VIF) for each variable and found severe multicollinearity among RBC, HGB, T-BIL, and I-BIL with a VIF greater than 10. Thereafter, we used stepwise backward logistic regression analysis to exclude these multicollinear variables. The most representative set of variables for CAS was selected when the Akaike information criterion was −7661.14. Furthermore, multivariable logistic regression analysis found that sex, age, MCV, and FBG were independent risk factors for CAS (Table 2).

Table 2 Multivariable Logistic Regression Analysis

Establishment of a CAS Risk Prediction Model

In the training set, we used LASSO combined with 10-fold cross-validation to screen out 19 non-zero characteristic variables (Figure 2 and Table 3). Afterward, we removed variables with low weights (scores < 15) from the risk model. Finally, a CAS risk prediction model was built using age, GGT, SCR, FBG, TT3, and D-BIL as predictors and was presented visually as a nomogram. For instance, using the risk model, a participant aged 84 years with HUA, a GGT of 23 U/L, a SCR of 105 μmol/L, a FBG of 5.5 mmol/L, a TT3 of 1.45 nmol/L, and a D-BIL of 3.6 μmol/L has an estimated probability of developing CAS of 90% (Figure 3).

Table 3 Coefficients and Lambda.min Value of the LASSO Regression

Figure 2 Characteristic variables were screened using LASSO regression analysis. (A) Trajectories of change in coefficients for each variable in the LASSO regression. (B) The best parameter (lambda) selection in the LASSO model uses 10-fold cross-validation with the lowest standard. The relationship curve between partial likelihood deviation (binomial deviation) and log (lambda) was plotted. Dotted vertical lines were drawn at the optimal values by using the minimum criteria and the 1 SE of the minimum criteria (the 1 SE criteria). LASSO, least absolute shrinkage and selection operator; SE, standard error.

Figure 3 A nomogram for predicting the probability of developing CAS in the HUA population. The nomogram is used by scoring each variable on its corresponding score scale. The scores for all variables are then summed up to obtain the total score, and a vertical line is drawn from the total point row to indicate the estimated probability of the development of CAS in the HUA population. The red dot on the scale represents the corresponding score of the variable. The asterisks indicate the level of statistical significance of each variable, with * P < 0.05, ** P < 0.01, and *** P < 0.001. CAS, carotid atherosclerosis; HUA, hyperuricemia; GGT, gamma-glutamyl transpeptidase; SCR, serum creatinine; FBG, fasting blood glucose; TT3, total triiodothyronine; D-BIL, direct bilirubin.

Internal Validation of a CAS Risk Prediction Model

The discriminative ability of the risk model was assessed by the ROC curve. The area under the ROC curve (AUC) was 0.891 in the training set and 0.901 in the internal validation set, which indicates an excellent discriminative ability of the model (Figure 4).

Figure 4 Receiver operating characteristic (ROC) curves. (A) Training set; (B) Internal validation set.

Calibration curves were used to assess the agreement between the estimated probabilities and observed results in the model. Our study revealed that the predicted probabilities were well-aligned with the observed outcomes in both the training and internal validation sets, indicating a high level of model fitting (Figure 5).

Figure 5 Calibration curves. (A) Training set; (B) Internal validation set. The x-axis represents the predicted CAS risk. The y-axis represents the actual diagnosed CAS. The diagonal dotted line represents a perfect prediction by an ideal model. The solid line represents the performance of the nomogram, of which a closer fit to the diagonal dotted line represents a better prediction.

The clinical utility of the risk model was evaluated by DCA. The risk threshold probabilities for the training and internal validation sets were 1–87% and 1–100%, respectively, which suggested that our risk model was clinically beneficial in this range (Figure 6).

Figure 6 Decision curve analysis. (A) Training set; (B) Internal validation set. The black line represents the net benefit when none of the participants are considered to develop CAS, while the light gray line represents the net benefit when all participants are considered to develop CAS. The area between the blue and light gray lines in the model curve indicates the clinical utility of the model.

Discussion

HUA was a widely recognized risk factor for CAS. As a major potential risk factor for ischemic stroke, CAS constantly threatens life and health. Therefore, early prediction and prevention of CAS are vital. This study enrolled 3579 HUA individuals, including 903 with CAS. The incidence of CAS among the HUA population was 25.2%, similar to the 26% reported in the 2009 NHLBI Family Heart Study.¹⁶ Multivariable logistic regression analysis identified sex, age, MCV, and FBG as independent risk factors for CAS in a population with HUA. A risk prediction model was established using age, GGT, SCR, FBG, TT3, and D-BIL as predictors. The risk model showed an excellent discriminative ability with a comparatively high AUC in the training and internal validation sets (0.891 and 0.901, respectively) and a good degree of fit with the calibration curves. Meanwhile, the DCA results indicated that the risk model clinically benefited when the risk threshold probabilities were 1–87% and 1–100% for the training and internal validation sets, respectively.

This study showed that sex [odds ratio: 2.137 (95% CI: 1.580–2.904)] was an independent risk factor for CAS. Males have a higher risk of CAS than females, possibly because of sex hormone levels.¹⁷ Multiple studies confirmed that age is a risk factor for CAS.^18–20 However, more consideration should be given to other controllable factors, because age is an uncontrolled factor. We also observed FBG as an independent risk factor for CAS, in agreement with previous studies^7,21 and FBG plays a significant role in ischemic stroke.^22,23 Studies have shown that blood lipids were significantly associated with CAS.²⁴ However, based on the available data, we did not observe a statistically significant difference between them. One possible explanation for this is that our study population primarily consisted of individuals who underwent routine health examinations. As a result, they may have been in the early stages of CAS, when the elevation of blood lipids may not yet appear to be a risk factor for CAS.

In clinical practice, the diagnosis of CAS relies primarily on carotid ultrasonography.²⁵ However, difficulties remain with the large-scale availability of ultrasound in less-medically developed regions or countries. In recent years, with the growing demand for high-quality healthcare, machine learning has become a powerful tool in clinical medicine. For instance, one study developed explainable machine learning models for the early screening of CAS, demonstrating excellent clinical performance with an AUC of 0.860.²⁶ Additionally, Chen et al²⁷ identified CAS endotypes by machine learning methods, and their results showed that these endotypes could be applied in precision medicine to prevent atherosclerotic cardiovascular disease. In this study, we have developed a risk prediction model for CAS in the HUA population, which could contribute to the early identification of individuals at high risk.

Our study has significant clinical implications and potentially represents the first application of machine learning methods to predict CAS risk in the HUA population. By offering an alternative approach for the early detection of CAS, our model identifies individuals at elevated risk, enabling physicians to selectively perform further tests. Such a strategy promises to optimize healthcare resource deployment and engender economic efficiencies. We utilized variables such as age, GGT, SCR, FBG, TT3, and D-BIL to establish a risk prediction model for CAS. These variables are readily available in routine clinical practice, ensuring the model’s widespread applicability in the healthcare domain. Furthermore, we developed a nomogram to facilitate the use of our predictive model by healthcare professionals in daily clinical practice.

Although the risk prediction model we developed showed good predictive value, limitations were inevitable. First, the study population was regional, which may affect the extrapolation of the prediction model. Second, clinical baseline data were not collected comprehensively enough, and potential clinical predictors may have been overlooked. Third, the risk prediction model was only validated using internal datasets, while the validation of external datasets is necessary. In the future, We will conduct studies with multiple centers and large samples to further revise and improve the model.

Conclusion

This study identified sex, age, MCV, and FBG as independent risk factors for CAS in a population with HUA. Additionally, we developed a CAS risk prediction model for the early identification of high-risk individuals.