We retrospectively enrolled patients diagnosed with IS who were admitted to the Department of Neurology at Suining Central Hospital between January 2023 and July 2025. During hospitalization, all participants were required to have undergone at least two separate fasting SUA measurements on non-consecutive days, with the first conducted on the day of admission and the second on day 3 after admission. The inclusion criteria for IS patients were as follows: age ≥ 18 years and a first-time IS diagnosis confirmed by neuroimaging, such as cranial computed tomography (CT) or magnetic resonance imaging (MRI). Exclusion criteria included: tumor; chronic kidney disease (CKD) caused by diverse etiologies, severe dysfunction of major organs (heart, liver, kidney, or lungs); severe systemic infection or sepsis; autoimmune diseases; receipt of in-hospital anticoagulation, thrombolysis, or reperfusion therapy; a previous history of stroke or psychiatric disorders, or those who underwent only one SUA measurement after admission. Furthermore, we also concurrently included 2,459 relatively healthy subjects, defined as those who presented with IS-like symptoms but were confirmed non-IS via imaging examinations and had no tumors, CKD, severe organ function impairment, or infection. Based on prior studies (23, 24), HUA was defined as a SUA level ≥ 7.0 mg/dL (420 μmol/L) among males and ≥ 6.0 mg/dL (360 μmol/L) among females. This study was approved by the Ethics Committee of the Suining Central Hospital (No. KYLLKS20250126). Informed consent of all participants were involved. The flow diagram illustrating all analytical procedures in this study is presented in Figure 1.

We extracted general clinical characteristics of the participants from electronic medical records, including age, gender, marital status, ethnicity, height, weight, systolic blood pressure (SBP), diastolic blood pressure (DBP), and medical histories of DM, coronary heart disease (CHD), atrial fibrillation (AF), hypertension (HTN), hyperlipidemia (HLP); medication histories of antiplatelet therapy (APT), antihypertensive therapy, antidiabetic therapy, statin therapy, and urate-lowering therapy; as well as smoking and alcohol use. Hyperlipidemia (HLP) was defined as TG ≥ 150 mg/dl (1.70 mmol/L), TC ≥ 240 mg/dl (6.22 mmol/L), LDL-C ≥ 160 mg/dl (4.14 mmol/L), HDL-C < 40 mg/dl (1.04 mmol/L), or a medical diagnosis (25). All laboratory parameters, including routine blood tests, a coagulation profile, renal function tests, electrolytes, and a lipid panel, all of which were all measured on admission, were retrieved from the Laboratory Information System (LIS). Since the proportion of missing D-dimer values exceeded 60%, we did not incorporate it into the coagulation profile. Meanwhile, we also collected SUA of all participants on the 3rd day after admission. For patients with IS, we additionally collected the time from symptom onset to admission; Trial of Org 10172 in Acute Stroke Treatment (TOAST) classification; and modified Rankin Scale (mRS), National Institutes of Health Stroke Scale (NIHSS), and Glasgow Coma Scale (GCS) scores at admission. FBG was measured by the hexokinase method; UA was detected by the uricase-peroxidase method; TC was measured by the cholesterol oxidase method; TG was quantified by the glycerol-3-phosphate oxidase-peroxidase (GPO-POD) method; HDL-C and LDL-C were detected by direct methods, with all assays performed on a Beckman Coulter AU5800 analyzer. Besides, other relevant derived indices are provided below: neutrophil-to-lymphocyte ratio (NLR) = neutrophil (NEU)/lymphocyte (LYM), lymphocyte-to-monocyte ratio (LMR) = LYM/monocyte (MON), systemic inflammation index (SII) = platelet (PLT) * NEU/LYM (26), system inflammation response index (SIRI)=NEU*MON/LYM (27), platelet-to-neutrophil ratio (PNR) = PLT/NEU (28), platelet-to-lymphocyte ratio (PLR) = PLT/LYM (28), monocyte to high-density lipoprotein cholesterol ratio (MHR) = MON/high-density lipoprotein cholesterol (HDL-C) (29), neutrophil to high-density lipoprotein ratio (NHR) = NEU/HDL-C (30), platelet to high-density lipoprotein cholesterol ratio (PHR) = PLT/HDL-C (31), hemoglobin-to-red blood cell distribution width ratio (HRR) = hemoglobin (HGB)/red blood cell distribution width (RDW) (32), Hemoglobin, Albumin, Lymphocyte, Platelet Score (HALP) = hemoglobin (HGB) * albumin (ALB) * LYM/PLT (33), triglyceride-glucose index (TyG) = ln(fasting blood glucose [FBG] (mg/dL) * triglyceride [TG] (mg/dL)/2) (34), atherogenic index of plasma (AIP) = lg(TG/HDL-C) (35), atherogenic coefficient (AC) = non-high-density lipoprotein cholesterol (non-HDL-C)/HDL-C (36), lipoprotein combine index (LCI) = total cholesterol (TC) * TG * low-density lipoprotein cholesterol (LDL-C)/HDL-C (37, 38), Castelli’s risk index I (CRI-I) = TC/HDL-C (36), and Castelli’s risk index II (CRI-II) = LDL-C/HDL-C (36).

For data with missing values below 5%, we performed multiple imputation using the “mice” package, with the number of iterations (parameter m) set to 5, which effectively handled the missing values. Normality of continuous variables was assessed using the Kolmogorov-Smirnov test. Variables conforming to a normal distribution were expressed as mean ± standard deviation (SD) and compared using the Student’s t-test; otherwise, they were summarized as median (Q1, Q3) and compared using the Mann-Whitney U test. Categorical variables were described as frequencies and percentages and compared by the chi-square test. To mitigate potential confounding effects, we performed 1:1 propensity score matching (PSM) (39), matching the comorbidity and non-comorbidity groups to 1,031 subjects. The covariates included age, gender, marriage, nationality, SBP, DBP, BMI, the bad habits of drinking and smoking; the histories of HTN, DM, AF and CHD, HLP; medication histories of APT, antihypertensive therapy, antidiabetic therapy, statin therapy, and urate-lowering therapy; and inflammatory markers (WBC count and SII). The propensity score was calculated for each participant using a caliper width of 0.02. Balance between groups after matching was evaluated with standardized mean difference (SMD) and quartile comparisons (40), using the Love plot for balance assessment. Post-matching results indicated well-balanced baseline characteristics between the two groups. Univariate and multivariate logistic regression were employed to assess the association between differential metabolic parameters and the risk of HUA-IS comorbidity before and after PSM. We categorized significant variables identified from the multivariate logistic regression analysis to quartile-based logistic regression. Three models were employed: Model 1 (crude model) included no covariates; Model 2 (partially adjusted model) was adjusted for demographic factors (age, gender, marital status, nationality, SBP, DBP, drinking status, smoking status, histories of HTN, DM, AF, CHD, HLP, and body mass index [BMI]); Model 3 (fully adjusted model) further incorporated medication histories of APT, antihypertensive therapy, antidiabetic therapy, statin therapy, urate-lowering therapy, and inflammatory status (WBC and SII). Subsequently, we performed restricted cubic spline (RCS) analysis on variables that remained significant in the quartile-based logistic regression to evaluate the dose-response relationship, with 4 knots placed at the 5th, 35th, 65th, and 95th percentiles. The matched dataset was split into training and internal validation sets in a 7:3 ratio. For variables with significant differences in the training set, we applied the LASSO method for regularization and coefficient shrinkage, thereby shrinking the coefficients of irrelevant features to zero and selecting features for subsequent model construction (41). In the training set, based on the selected features, 11 ML algorithms, including Generalized Linear Model (glm), Recursive Partitioning and Regression Trees (rpart), Naive Bayes (nb), Bayesian Generalized Linear Model (bayesglm), Random Forest (rf), eXtreme Gradient Boosting Tree (xgbTree), Support Vector Machine with Radial Basis Function Kernel (svmRadial), Support Vector Machine with Linear Kernel (svmLinear), Gradient Boosting Machine (gbm), Multivariate Adaptive Regression Splines (earth), and Elastic Net Regularized Regression (glmnet), were utilized to construct Clinlabomics models for HUA-IS comorbidity. Hyperparameter tuning was performed via grid search in the training set. Model performance was evaluated using the accuracy (ACC), F1-score, and area under the curve (AUC), with the optimal Clinlabomics model identified. Receiver operating characteristic (ROC) curves, calibration curve, and decision curve analysis (DCA) were used to evaluate the discriminative ability, the agreement between predicted probabilities and actual probabilities, and the clinical utility of the optimal model, respectively. Based on the optimal model, we also calculated its Brier score to further quantify the model’s calibration performance. Subsequently, we enrolled 108 participants from August to September 2025 as the independent validation cohort to evaluate the accuracy of the optimal Clinlabomics model. In addition, to interpret the contribution of individual features to the model, the SHAP algorithm was applied to compute SHAP values from the best-performing Clinlabomics model (42). SHAP values quantify the marginal contribution of each feature to the estimated outcome. This approach supports both local and global interpretability of the model. The entire process for Clinlabomics model development was performed using the “Icare” package (https://github.com/OmicsLY/Icare). All statistical analyses were performed using RStudio (Version 4.3.0). A two-sided P-value < 0.05 was considered statistically significant.

A total of 2,164 IS patients and 2,459 healthy controls (HCs) were included in this study. Based on two SUA levels, subjects were stratified into four groups: non-HUA HCs (n=1,145), HUA HCs (1,314), non-HUA IS (n=1,082), and HUA-IS comorbidity (n=1,082) group, with differences in characteristics across groups presented in Table 1. A detailed flowchart is provided in Figure 1. Supplementary Table S1 presents missing data proportions and data distributions before and after imputation, with no significant differences detected in the latter. Meanwhile, we compared the differences of indicators between non-HUA IS patients and non-HUA HCs (Supplementary Table S2), as well as between HUA HCs and non-HUA HCs (Supplementary Table S3). Through these comparisons, we identified differentially expressed indicators closely associated with IS and HUA, respectively. To further investigate the potential pathogenic mechanism of the comorbidity group, we merged all groups except the HUA-IS comorbidity group into a non-comorbidity group and compared the differences in indicators (Supplementary Table S4). Subsequently, we took the intersection of the differentially expressed indicators from these three groups (Supplementary Table S5), identifying 33 indicators closely associated with IS, HUA, and comorbidity (Supplementary Figure 1). Among these, ten metabolic indicators, including UA_admission, UA_3d, TyG, TG, HDL-C, AIP, AC, LCI, CRI-I, and CRI-II, were found to be closely correlated with the comorbidity, indicating no significant biological confounding. After PSM, both the comorbidity and non-comorbidity groups included 1,031 cases each, with 2,561 participants (55.4%). The standardized mean difference (SMD) values of these covariates were all less than 0.1, as visualized via a Love plot (Supplementary Table S6, Supplementary Figure 2). Significant differences (all P < 0.05) were observed between the two groups in nationality, DBP, MON, NLR, LMR, SII, SIRI, PLR, MHR, HRR, HALP, RBC, HGB, HCT, RDW-CV, TG, AIP, LCI, TyG, UREA, CREA, UA_admission, UA_3d, and TT (Table 2). Additionally, the dataset was randomly split into a training set and a testing set at a 7:3 ratio, with intergroup differences in the training set presented in Table 3. Furthermore, the detailed characteristics of all clinical and laboratory results in the independent validation cohort were presented in Table 4.

Ten metabolic indicators (lipid, glucose, and uric acid metabolism) were included, followed by univariate and multivariate analyses. Before PSM, UA_admission (OR: 1.02, 95% CI: 1.02-1.02, P < 0.001), UA_3d (OR: 1.01, 95% CI: 1.01-1.01, P < 0.001), TyG (OR: 1.32, 95% CI: 1.17-1.48, P < 0.001), TG (OR: 1.05, 95% CI: 1.01-1.09, P = 0.027), AIP (OR: 2.21, 95% CI: 1.59-3.07, P < 0.001), and LCI (OR: 1.00, 95% CI: 1.00-1.00, P = 0.022) showed significant associations with the comorbidity in multivariate models. After PSM, only five metabolic parameters, namely, UA_3d (OR: 1.02, 95% CI: 1.02-1.02, P < 0.001), TyG (OR: 1.40, 95% CI: 1.21-1.62, P < 0.001), TG (OR: 1.13, 95% CI: 1.05-1.22, P = 0.002), AIP (OR: 2.74, 95% CI: 1.80-4.19, P < 0.001), and LCI (OR: 1.01, 95% CI: 1.00-1.01, P = 0.001) were significantly associated with increased risk of HUA-IS comorbidity in fully adjusted models. All associations were statistically significant (P < 0.05), visually summarized in Figures 2A, B, Supplementary Table S7.

Subsequently, we performed quantile stratification on the aforementioned five statistically significant metabolic parameters following PSM (Supplementary Table S8). After PSM, the quartiles of UA_3d were not significantly associated with the comorbidity (all P-values > 0.05, in all models). Elevated risks of HUA-IS comorbidity were observed in the highest quartile (Q4) of the following indicators in fully adjusted multivariate models: TyG (OR: 1.92, 95% CI: 1.40-2.62, P < 0.001), TG (OR: 2.60, 95% CI: 1.90-3.56, P < 0.001), AIP (OR: 2.00, 95% CI: 1.42-2.81, P < 0.001), LCI(OR: 1.68, 95% CI: 1.25-2.27, P = 0.001). Furthermore, significant dose-response associations were confirmed for all above indicators, with a trend P-value < 0.05 across increasing quartiles (Figures 3A, B, Supplementary Table S9).

Although significant linear associations were observed for four indicators before (Figure 4) and after PSM (Figure 5), only TG and LCI demonstrated significant nonlinear relationships (nonlinear P < 0.05). Specifically, for TG, the risk of comorbidity showed a pronounced increase with increasing TG levels below the threshold of 3.69 mmol/L (pre-PSM) and 3.81 mmol/L (post-PSM). Above these cutoffs (TG > 3.69 and 3.81 mmol/L), the risk continued to decrease (Figure 4B, Figure 5B). To verify whether the inverted U-shaped effect was due to the small TG sample, we identified subsets with TG ≥ 3.69 (pre-PSM) and ≥ 3.81 (post-PSM) mmol/L, which included 360 patients (8%) and 160 patients (8%), respectively. Excluding these patients, repeated RCS analyses showed the nonlinear relationship was no longer significant (nonlinear P-values: 0.544 and 0.246, Supplementary Figures S3A, B). Thus, the U-shaped effect may reflect the sparse high-TG sample rather than a true biological effect. Besides, regarding the LCI, its inflection points were 18.79 (before PSM) and 19.17 (following PSM), respectively. When LCI values were below these thresholds, the risk of comorbidity increased in a statistically significant manner with the elevation of LCI; once LCI exceeded these inflection points, the rate of risk elevation began to decelerate, reflecting a nonlinear dose-response relationship between LCI and comorbidity risk (Figures 4D, 5D).

A total of 23 differentially expressed variables from the training set were subjected to LASSO regression analysis, which identified 11 feature variables for subsequent model development (Figures 6A-C). However, considering the associations of age, gender, TyG, and TG with the comorbidity, we incorporated these variables into subsequent model construction. Eleven​ algorithms were used to develop Clinlabomics models based on the training set, with the optimal Clinlabomics model (established by the rpart algorithm) identified (F1-score: 0.960; sensitivity: 0.965; specificity: 0.954; AUC: 0.986, 95% CI: 0.981-0.992) (Figure 6D). Subsequently, in the training set, hyperparameter tuning was conducted on the optimal model, with the complete tuning procedure depicted in Figure 6E. The optimal hyperparameter for the rpart algorithm was identified as cp=0.0017. In the training set, ROC curves were generated and AUC values were calculated for the untuned and tuned models, demonstrating a slightly higher AUC in the tuned model (0.987, 95% CI: 0.982-0.993) compared to the untuned model (0.986, 95% CI: 0.981-0.992) (Figure 6F). Ultimately, the optimally tuned Clinlabomics model was selected, with ROC curves generated across the training set, testing set (AUC: 0.955, 95% CI: 0.939-0.972), and validation set (AUC: 0.957, 95% CI: 0.915-0.999) (Figure 7A). Calibration curves demonstrated that, for the training, testing, and validation sets, the slopes of the relationship between estimated and observed probabilities approximated to 1 (with corresponding intercepts near 0), indicating robust calibration performance across these datasets and high concordance between estimated and actual event probabilities (Figure 7B). In addition, the Brier score was low for the training (0.034), testing (0.084), and validation (0.068) sets, indicating the model’s accurate assessment with minimal bias across these datasets. Furthermore, for the three datasets, their DCA curves exhibited significantly higher standardized net benefit than the “none” (no risk stratification) strategy across a wide high-risk threshold range, indicating the model’s robust clinical decision-making net benefit across these datasets (Figure 7C). Subsequently, feature importance ranking (Figure 7D) and SHAP-based interpretation (Figure 7E) were performed for the optimal model, revealing the UA_admission, UA_3d, TG, LCI, AIP, TyG as the influential metabolic factors. Randomized​ SHAP visualization of a subject mitigated the model’s black-box limitation by elucidating impactful variables, aiding individualized diagnosis and treatment (Figure 7F). Subsequently, the validation set was used as a new dataset. Without group labels, it was assessed by the model, classifying 57 as comorbidity and 51 as non-comorbidity (Figure 7G). Probabilities < 40%, 40–70%, and >70% were defined as low-risk, intermediate-risk, and high-risk, respectively. Comparison with original labels showed that 50 (87.7%, probability < 40%) non-comorbidity and 50 (98.0%, probability > 70%) comorbidity patients were accurately identified (Figure 7H), indicating favorable discriminative ability of our optimal Clinlabomics model. To identify comorbid populations from healthy controls, we excluded hyperuricemia-negative IS patients from the non-comorbid group and analyzed the data via the aforementioned methods. The rpart algorithm remained optimal (Supplementary Figure 4A), with an optimal cp of 0.00015 after tuning (Supplementary Figure 4B), though the adjusted model’s AUC did not differ from the unadjusted model’s AUC (Supplementary Figure 4 C). The optimal Clinlabomics model exhibited AUCs of 0.974 (95%CI: 0.968–0.980), 0.916 (95%CI: 0.896–0.937), and 0.976 (95%CI: 0.949–1.00) across three datasets, with corresponding Brier scores of 0.051, 0.093, and 0.087 (Supplementary Figure 5A). Calibration and DCA curves (Supplementary Figures 5B, C) confirmed its strong performance in distinguishing comorbid patients from healthy controls. Feature importance ranking and SHAP plots highlighted UA, UA_3d, LCI, AIP, TyG, and TG as key predictors (Supplementary Figures 5D-F). In the validation set (72 unlabeled subjects), the model identified 45 comorbid patients and 27 healthy controls (Supplementary Figure 5G), with correct classification rates of 95.2% (20/21) for healthy controls and 78.4% (40/51) for comorbid patients (Supplementary Figure 5H). Meanwhile, we also attempted to build models for the IS population, but failed due to overfitting of ML algorithms, so no IS-specific comorbidity model was established.