Association of the triglyceride-to-high-density lipoprotein cholesterol ratio with kidney stone status: cross-sectional and machine-learning analyses

Abstract

This study investigated the association between triglyceride-to-high-density lipoprotein cholesterol ratio (TG/HDL-C) and kidney stone status and evaluated its contribution to multivariable classification models. A total of 1,174 patients from the First Affiliated Hospital of Anhui Medical University were included, with kidney stone status confirmed by urological imaging. Multivariable analyses were used to assess the association between TG/HDL-C and kidney stone status. Supplementary population-based analyses were conducted using weighted data from the National Health and Nutrition Examination Survey (NHANES). Seven machine-learning models were developed and internally evaluated within the clinical cohort, and SHapley Additive Explanations (SHAP) were used to interpret model outputs. In the clinical cohort, a higher TG/HDL-C ratio was associated with greater odds of kidney stone status after multivariable adjustment (OR, 2.16; 95% CI, 1.73–2.70). The association was generally consistent across subgroups defined by age, sex, body mass index, hypertension, and diabetes, with no significant interactions observed. A directionally similar association was identified in the weighted NHANES analysis (OR, 1.04; 95% CI, 1.01–1.07). Among the evaluated machine-learning models, the random forest model achieved the highest area under the receiver operating characteristic curve in the internal validation set (AUROC, 0.818; 95% CI, 0.775–0.859). SHAP analysis identified the TG/HDL-C ratio as an important contributor to model classification. These findings indicate that the TG/HDL-C ratio is an accessible metabolic marker associated with kidney stone status. However, the observational design precludes causal inference, and prospective studies with external model validation are warranted.

Introduction

Kidney stones, which consist of both crystalline and organic components, are mineral deposits that form within the renal calyces and renal pelvis. When urinary minerals become supersaturated, they may form free particles or adhere to the junctions among the renal calyces, renal pelvis, and ureter [1]. Kidney stones are a common condition, with a global prevalence estimated at 7.2%–7.7%, according to data from the 2007–2010 National Health and Nutrition Examination Survey (NHANES) in the United States [2,3]. The prevalence of kidney stones in the general population increased from 5.2% in the 1988–1994 survey cycle to 8.4% in the 2007–2010 survey cycle [4]. Various treatment modalities are used for kidney stones, including open lithotomy, laparoscopic lithotripsy, percutaneous nephrolithotomy, ureteroscopy, extracorporeal shock wave lithotripsy, and medical therapy [5]. Despite these treatment options, kidney stones are not always completely resolved, thereby contributing to a substantial societal burden [6]. Therefore, the clinical prevention of kidney stones has become an increasingly important focus.

Previous research has suggested that unhealthy lifestyle behaviors, such as adherence to a Western diet, may contribute to the increasing incidence of diseases associated with metabolic syndrome [7]. This study builds on this understanding by investigating the potential association between metabolic dysfunction and kidney stones. Previous studies have shown that metabolic syndrome and its components, including hypertension, diabetes, and insulin resistance, are potential risk factors for the development of kidney stones [4,6,8]. Insulin resistance is clinically defined as a state in which a given amount of exogenous or endogenous insulin is unable to promote glucose uptake and utilization as effectively as in healthy individuals. Research has demonstrated that insulin resistance increases the risk of urinary calcium stones by reducing urinary citrate excretion [7]. Currently, the hyperinsulinemic–euglycemic clamp (HIEC) is considered the gold standard for assessing insulin sensitivity in peripheral tissues; however, this method is complex, time-consuming, and resource-intensive [9]. In previous studies, HOMA-IR has been widely used as a noninvasive method for assessing insulin resistance; however, because it depends on insulin measurements, its utility is limited in subjects receiving insulin therapy or in those without functional pancreatic β-cells [10]. Therefore, there is a clear need to develop new and effective methods for assessing insulin resistance. In a study of overweight and obese children, Eman G. Behiry found that a TG/HDL-C ratio of ≥1.36 was a more sensitive indicator of insulin resistance in children than HOMA-IR [11]. The TG/HDL-C ratio (triglyceride-to-HDL cholesterol ratio), calculated by dividing the triglyceride concentration by the high-density lipoprotein cholesterol concentration, is considered a simple, reliable, and reproducible indicator associated with insulin resistance [8,9]. It has been identified as a marker of cardiometabolic health and risk of type 2 diabetes, and additional studies have linked TG/HDL-C to the prevalence of hypertension [8,10,12]. The standard TG/HDL-C ratio may help identify patients with insulin resistance as well as those exhibiting dyslipidemic features [13].

Although previous studies have linked metabolic syndrome, insulin resistance, obesity, hypertension, diabetes, and dyslipidemia to kidney stone disease, the specific association between the TG/HDL-C ratio and kidney stones remains insufficiently characterized. In addition, most prior studies have relied mainly on conventional regression analyses and have provided limited information on whether TG/HDL-C contributes to multivariable classification models for kidney stone status.

To address these gaps, this study investigated the association between TG/HDL-C and kidney stone status in a single-center clinical cohort with imaging-confirmed kidney stones. We further examined this association in NHANES as a supplementary population-based analysis. In addition, we developed and compared multiple machine-learning models and used SHAP analysis to evaluate the contribution of TG/HDL-C and other clinical variables to model classification [14,15]. Through this approach, our study aimed to provide a clearer observational assessment of TG/HDL-C as a clinically accessible metabolic marker associated with kidney stones.

Materials and methods

Clinical center study population data collection

This study was conducted in accordance with the Declaration of Helsinki (2013 revision) and was approved by the Ethics Committee of the First Affiliated Hospital of Anhui Medical University (approval no. PJ2025-10-71). The requirement for written informed consent was waived by the Ethics Committee of the First Affiliated Hospital of Anhui Medical University owing to the retrospective study design. The clinical cohort was derived from a health examination population at the First Affiliated Hospital of Anhui Medical University. We retrospectively screened participants who underwent routine health examinations and had available urological ultrasonography or CT results between June 2021 and June 2025. Urological imaging was performed as part of the routine health examination or screening assessment rather than because of specific urological symptoms or disease-based indications. A total of 1,174 eligible participants were included in the final analysis. According to imaging findings, 600 participants with evidence of kidney stones were classified as stone formers, whereas 574 participants without imaging evidence of kidney stones were classified as non-stone participants. The inclusion criteria were: (1) age ≥18 years and (2) complete data for eligibility-defining variables. The exclusion criteria were: (I) immunosuppressive diseases; (II) renal abnormalities, such as horseshoe kidney and transplanted kidney; (III) ureteral anatomical abnormalities, such as duplicated ureter and congenital ureteral stenosis; and (IV) infections of other systems.

Clinical data collection included the following variables: (I) demographic characteristics, including sex, age, body mass index (BMI), history of diabetes, and hypertension; and (II) laboratory tests. Venous blood samples were collected the following morning after an overnight fast and analyzed in the Laboratory Department of the First Affiliated Hospital of Anhui Medical University. All tests were performed by professional laboratory technicians to ensure quality and accuracy. The biochemical parameters included total cholesterol (TC), triglycerides (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), very-low-density lipoprotein (VLDL), ApoA, ApoB, total bilirubin (TBIL), creatinine (Cre), urea, uric acid (UA), estimated glomerular filtration rate (eGFR, mL/[min·1.73 m²]), calcium (Ca), and other biochemical indicators. Participants with missing data in variables required for eligibility assessment and cohort definition, including the TG/HDL-C ratio and urological imaging data, were excluded during screening. After the inclusion and exclusion criteria were applied, missing values in the remaining analytical variables were addressed using multiple imputation with the MICE package in R. The extent of missingness for key variables is presented in Supplementary Figure S1. A prespecified directed acyclic graph was constructed to clarify the assumed relationships among demographic factors, metabolic factors, renal function-related biomarkers, Cys-C, and kidney stone status, and to guide the interpretation of confounders, potential intermediate variables, colliders, and disease-related consequences. The DAG is presented in Supplementary Figure S2.

A total of 1,230 participants were initially screened for eligibility. According to the predefined inclusion and exclusion criteria, 19 participants aged <18 years were excluded. In addition, 24 participants with missing TG/HDL-C ratio data, 7 with missing urological imaging data, and 6 with horseshoe kidney, solitary kidney, or abnormal ureteral anatomy were excluded. Ultimately, 1,174 participants were included in the final analysis (Figure 1).

Figure 1.

Flowcharts of participant selection. (A) Clinical cohort. Of the 1,230 participants screened, 1,174 were included after the predefined exclusion criteria were applied. (B) NHANES analysis. Of the 59,842 participants initially assessed, 14,849 were included after applying the age eligibility criterion and excluding participants with missing data on kidney stone history, the TG/HDL-C ratio, or relevant covariates.

The TG/HDL-C ratio was calculated as TG (mmol/L)/HDL-C (mmol/L). All participants underwent urological ultrasonography or CT to screen for kidney stones. Participants with evidence of kidney stones on imaging were classified as having kidney stones, regardless of stone size or number.

NHANES dataset study design and participants

The baseline clinical data analyzed in this study were obtained from the National Health and Nutrition Examination Survey (NHANES) cycles from 2007 to 2018. NHANES is a biennial cross-sectional survey of the US population conducted by the Centers for Disease Control and Prevention (CDC). The NHANES protocol was approved by the Institutional Review Board of the National Center for Health Statistics (NCHS), and all participants provided written informed consent. NHANES was used only for supplementary weighted association analyses. It was not used for machine-learning model development, testing, or external validation. We included six consecutive 2-year survey cycles that administered the kidney stone questionnaire. For this cross-sectional analysis, data were extracted from the NHANES database for the period 2007–2018. A total of 59,842 participants were initially screened. Participants were excluded if they were younger than 20 years (n = 25,072), lacked information on kidney stone history (n = 91), had missing data on triglycerides (TG) or high-density lipoprotein cholesterol (HDL-C) (n = 19,704), or had missing data on serum creatinine (n = 53), educational level (n = 16), marital status (n = 5), physical activity (n = 12), hypertension(n = 20), diabetes (n = 8), smoking status (n = 12). After these exclusions, 14,849 participants were included in the final analysis. The exclusion process is shown in Figure 1.

NHANES collection and definition of data

The TG/HDL-C ratio was defined as the exposure variable and calculated as triglycerides (mg/dL) divided by high-density lipoprotein cholesterol (HDL-C; mg/dL). Triglyceride and HDL-C concentrations were measured using an enzymatic method with an automated biochemical analyzer.

Kidney stone status was determined using the NHANES KIQ026 questionnaire item: ‘Have you ever had a kidney stone?’ [2]. Participants who answered ‘yes’ were classified as having a history of kidney stones. Kidney stone occurrence (yes/no) was defined as the outcome variable.

Potential confounders in the association between the TG/HDL-C ratio and kidney stone were prespecified and included in the multivariable-adjusted models. The covariates included sex (male/female), age (years), ethnicity, education level, poverty income ratio (PIR), marital status (married or living with a partner/single), alcohol consumption (yes/no), physical activity (vigorous/moderate/less than moderate), BMI (kg/m²), total cholesterol (mg/dL), LDL-C (mg/dL), blood calcium (mg/dL), albumin-creatinine ratio (mg/g), smoking status (yes/no), hypertension, diabetes, water consumption (<2,000 mL, ≥2,000 to <2,500 mL, and ≥2,500 mL), and asthma, all of which were identified as risk factors for kidney stone [16]. Detailed measurement procedures for all variables are available in the NHANES documentation.

Statistical analysis

Continuous variables are presented as mean ± standard deviation (SD), and categorical variables as percentages. The Kolmogorov-Smirnov (KS) test was used to assess the normality of continuous variables. Depending on the data distribution, t-tests were used for continuous variables, whereas chi-square tests were used for categorical variables and non-normally distributed data to compare differences between the two groups. According to established guidelines [17], multivariable logistic regression was used to evaluate the independent associations of the TG/HDL-C ratio and its tertiles with kidney stone disease, and three models were constructed. Trend tests were performed to assess whether the association between the TG/HDL-C ratio and kidney stones showed a linear pattern across increasing TG/HDL-C levels (P for trend). Model 1 was unadjusted. Model 2 was adjusted for demographic and physiological variables, including sex, age, hypertension, and diabetes. Model 3 was adjusted for total cholesterol, age, sex, BMI, hypertension, diabetes, total bilirubin (TBIL), creatinine (Cre), urea, calcium (Ca), and uric acid (UA). To assess potential multicollinearity, correlation matrices were generated for lipid-related variables and covariates included in the multivariable models. Variance inflation factors (VIFs) were calculated for the fully adjusted model. In addition, sensitivity analyses were conducted by excluding TC from the model and by replacing TC with LDL-C.

Possible nonlinear relationships between TG/HDL-C and group status were examined using logistic regression with restricted cubic splines (RCS). Models with 3 to 7 knots were tested, and the model with the lowest Akaike information criterion (AIC) was selected. Finally, an RCS model with 3 knots placed at the 10th, 50th, and 90th percentiles was used.

NHANES statistical analyses were conducted in accordance with the Centers for Disease Control and Prevention (CDC) guidelines, For the six combined NHANES cycles from 2007 to 2018, the 2-year fasting subsample weight (WTSAF2YR) was divided by six to generate the combined 12-year fasting subsample weight. The NHANES survey design was specified using the combined fasting subsample weight, strata, and primary sampling units. Based on the dataset-specific weights, the NHANES complex multistage sampling design was incorporated using the Survey Design package in R. Weighted means and 95% confidence intervals (CIs) were calculated for continuous variables, and weighted proportions and 95% CIs were calculated for categorical variables. For continuous variables, survey-weighted chi-square tests were used to assess differences between groups. Three logistic regression models were constructed to examine whether the TG/HDL-C ratio was independently associated with kidney stone. Model 1 was unadjusted. Model 2 was adjusted for sex, age, and ethnicity. Model 3 was adjusted for all covariates. The relationship between the TG/HDL-C ratio and kidney stones was further evaluated using smooth curve fitting (penalized spline method) and weighted generalized additive models (GAMs). In addition, log-likelihood ratio tests were used to assess heterogeneity across subgroups by including interaction terms. Statistical significance was defined as p < 0.05. All analyses were performed using EmpowerStats (X&Y Solutions, Inc., Boston, MA, USA) and R version 4.5.2 (R Foundation for Statistical Computing, Vienna, Austria).

Machine learning algorithms

In the machine-learning analysis, all prediction models were developed and internally evaluated only within the single-center clinical cohort. Before model preprocessing, the clinical cohort was randomly divided into a training set and an internal validation set at a 7:3 ratio. Variable selection was first performed in the training set using LASSO regression with 10-fold cross-validation. Based on the LASSO results and the clinical relevance of the prediction model, 10 variables were selected for model construction: sex, hypertension, diabetes, age, body mass index (BMI), total cholesterol (TC), total bilirubin (TBIL), triglyceride/high-density lipoprotein cholesterol ratio (TG/HDL-C), calcium (Ca), and uric acid (UA).

Based on these variables, seven machine learning models were constructed, including logistic regression, XGBoost, Random Forest, Decision Tree, LightGBM, Support Vector Machine (SVM), and Artificial Neural Network (ANN). The performance of each model in the validation set was evaluated using the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, accuracy, and F1 score, with AUROC as the primary evaluation metric. The model’s performance was evaluated using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), Brier score, and 95% confidence intervals for these metrics. The detailed performance metrics are provided in Supplementary Table S1.

Additionally, a calibration curve was used to assess the consistency between the predicted probabilities and the actual observed outcomes, and decision curve analysis (DCA) was employed to evaluate the clinical net benefit of the models.

Based on overall discrimination and calibration performance, the Random Forest model was selected as the optimal model for this study. For this model, hyperparameters were tuned in the training set using grid search with 5-fold cross-validation, with the area under the receiver operating characteristic curve (AUC) used as the optimization metric. The search range was defined as follows: n_estimators from 50 to 450 in increments of 50, and max_features from 2 to the square root of the number of candidate predictor variables. The optimal hyperparameters were n_estimators = 450 and max_features = 2. The predictive performance of the final model was then evaluated in the validation set. Model interpretability was assessed using SHapley Additive Explanations (SHAP).

For all tests, a significance level of p < 0.05 was considered statistically significant. All analyses were conducted using R software version 4.5.2 (http://www.R-project.org, The R Foundation) and all machine-learning analyses were performed using Python version 3.10.4, and we added the relevant Python package names and versions for LASSO, Random Forest, XGBoost, LightGBM, and SHAP analysis.

Results

Table 1 shows the baseline characteristics of 1,174 study participants, with continuous variables expressed as median and interquartile range (Q1, Q3) and categorical variables as frequency and percentage. Stone formers were significantly older and had a higher BMI than non-stone formers (both p < 0.001). The prevalence of hypertension and diabetes was notably higher in the stone group (41% and 15.7%, respectively) than in the non-stone group (19.3% and 8.4%, both p < 0.001). Regarding serum biochemical parameters, stone formers showed higher levels of total cholesterol, triglycerides, and LDL-C, but lower HDL-C concentrations (all p < 0.01). Serum cystatin C, creatinine, urea, calcium, and uric acid levels were also significantly increased, whereas eGFR values were lower in the stone-forming group (p < 0.001), indicating renal impairment. Total bilirubin levels were lower in stone formers compared with non-stone formers (p < 0.001). The TG/HDL-C ratio was significantly elevated among stone formers (p < 0.001). Across quartiles of TG/HDL-C, the prevalence of kidney stones increased progressively from Q1 to Q4 (11.8% → 37%).

Table 1.

Baseline Table of clinical center participants.

Characteristic	Overall (N = 1,174)	Non-stone Formers (N = 574)	Stone formers (N = 600)	P-value
Age (years)	54 (39, 61)	54 (33, 61)	55 (44, 61)	<0.001
BMI (kg/m2)	24.1 (21.7, 26.4)	23.3 (21.2, 25.4)	24.7 (22.5, 27.1)	<0.001
Hypertension, n (%)				<0.001
No	817 (69.6%)	463 (80.7%)	354 (59.0%)
Yes	357 (30.4%)	111 (19.3%)	246 (41.0%)
Diabetes, n (%)				<0.001
No	1,032 (87.9%)	526 (91.6%)	506 (84.3%)
Yes	142 (12.1%)	48 (8.4%)	94 (15.7%)
Sex, n (%)				0.276
Female	370 (31.5%)	182 (31.7%)	188 (31.3%)
Male	804 (68.5%)	392 (68.3%)	412 (68.7%)
TC (mmol/L)	4.60 (3.98, 5.33)	4.29 (3.69, 4.90)	4.96 (4.39, 5.61)	<0.001
TG (mmol/L)	1.32 (0.95, 1.82)	1.10 (0.82, 1.51)	1.55 (1.14, 2.07)	<0.001
HDL-C (mmol/L)	1.18 (1.00, 1.36)	1.24 (1.08, 1.44)	1.11 (0.94, 1.29)	<0.001
LDL-C (mmol/L)	2.82 (2.30, 3.34)	2.76 (2.25, 3.26)	2.89 (2.36, 3.45)	0.002
VLDL (mmol/L)	0.49 (0.35, 0.68)	0.41 (0.30, 0.56)	0.58 (0.42, 0.77)	<0.001
ApoA (g/L)	1.30 (1.11, 1.46)	1.32 (1.15, 1.48)	1.27 (1.10, 1.45)	0.007
ApoB (g/L)	0.87 (0.73, 1.00)	0.81 (0.68, 0.95)	0.91 (0.78, 1.02)	<0.001
TBIL (μmol/L)	12.9 (10.3, 16.2)	13.8 (11.1, 17.0)	12.1 (9.4, 15.3)	<0.001
IBIL (μmol/L)	9.0 (7.2, 11.3)	9.8 (8.0, 12.1)	8.1 (6.4, 10.3)	<0.001
DBIL (μmol/L)	3.95 (2.66, 5.20)	3.90 (2.60, 5.20)	4.00 (2.70, 5.25)	0.889
Cys-C (mg/L)	0.99 (0.75, 1.24)	0.77 (0.67, 0.95)	1.20 (1.01, 1.37)	<0.001
AST (U/L)	18.0 (16.0, 22.0)	18.0 (16.0, 21.0)	19.0 (16.0, 22.0)	0.024
ALT (U/L)	18 (14, 25)	18 (14, 24)	19 (14, 25)	0.109
Cre (μmol/L)	74 (62, 88)	69 (58, 78)	80 (67, 99)	<0.001
Urea (mmol/L)	5.28 (4.38, 6.31)	5.02 (4.14, 5.94)	5.53 (4.58, 6.69)	<0.001
Ca (mmol/L)	2.31 (2.24, 2.38)	2.28 (2.22, 2.35)	2.33 (2.26, 2.42)	<0.001
UA (μmol/L)	343 (283, 400)	334 (275, 390)	348 (291, 409)	0.034
TG/HDL-C	1.11 (0.74, 1.70)	0.89 (0.62, 1.28)	1.40 (0.94, 2.05)	<0.001
Q1	294 (25.0%)	223 (38.9%)	71 (11.8%)
Q2	293 (25.0%)	155 (27.0%)	138 (23.0%)
Q3	293 (25.0%)	124 (21.6%)	169 (28.2%)
Q4	294 (25.0%)	72 (12.5%)	222 (37.0%)
eGFR mL/(min·1.73 m²)	102 (85, 115)	109 (98, 123)	91 (72, 106)	<0.001

Abbreviations: BMI-Body Mass Index; TC-Total Cholesterol; TG-Triglycerides; HDL-C-High-Density Lipoprotein Cholesterol; LDL-C-Low Density Lipoprotein Cholesterol; Cys-C-Cystatin C; Cre-Creatinine; TBIL-Total Bilirubin; UA-Uric Acid; eGFR -estimated Glomerular Filtration Rate.

As shown in Table 2, in logistic regression analyses, a higher TG/HDL-C ratio was significantly associated with greater odds of kidney stone prevalence. In the unadjusted model (Table 2, Model 1), each increase in TG/HDL-C was associated with higher odds of kidney stones (OR, 3.01; 95% CI, 2.47–3.67). This association remained significant after adjustment for age, sex, hypertension, and diabetes (OR, 2.84; 95% CI, 2.33–3.47) (Table 2, Model 2) and after further adjustment for total cholesterol (TC), BMI, total bilirubin (TBIL), creatinine (Cre), urea, uric acid (UA) and calcium (Ca) (OR, 2.16; 95% CI, 1.73–2.70) (Table 2, Model 3). When TG/HDL-C was categorized into quartiles, higher quartiles were associated with progressively increased odds of kidney stones compared with the lowest quartile (Q1) across all models. In the fully adjusted model, the odds ratios were 2.11 (95% CI, 1.38–3.22) for Q2, 2.74 (95% CI, 1.78–4.21) for Q3, and 5.43 (95% CI, 3.43–8.62) for Q4 (p for trend < 0.001).

Table 2.

Logistic regression analysis between TG/HDL-C ratio and kidney stone prevalence.

Characteristic	Model 1 OR (95% CI)	Model 2 OR (95% CI)	Model 3 OR (95% CI)
TG/HDL-C ratio	3.01 (2.47, 3.67)	2.84 (2.33, 3.47)	2.16 (1.73, 2.70)
Categories
Q1 [0.23, 0.699)	1	1	1
Q2 [0.699, 1.06)	2.80 (1.97, 3.98)	2.45 (1.71, 3.52)	2.11 (1.38, 3.22)
Q3 [1.06, 1.65)	4.28 (3.01, 6.10)	3.78 (2.63, 5.43)	2.74 (1.78, 4.21)
Q4 [1.65, 6.26]	9.68 (6.64, 14.12)	8.57 (5.83, 12.58)	5.43 (3.43, 8.62)
P for trend	<0.001	<0.001	<0.001

Abbreviations: CI = Confidence Interval, OR = Odds Ratio.

Model 1: no covariates were adjusted.

Model 2: adjusted for age(years), sex, hypertension, and diabetes.

Model 3: adjusted for TC (mmol/L), age (years), sex, BMI (kg/m²), hypertension, diabetes, TBIL (μmol/L), Cre (μmol/L), urea (mmol/L), Ca (mmol/L) and UA (μmol/L).

To further assess whether the observed association was influenced by multicollinearity among lipid-related variables and covariates, correlation matrix and variance inflation factor (VIF) analyses were performed. The lipid correlation matrix showed expected correlations among lipid-related variables, particularly between TG/HDL-C and TG, VLDL, and HDL-C. In contrast, the correlations among covariates included in the fully adjusted model were generally low to moderate. All VIF values were below 5, suggesting no substantial multicollinearity among the variables included in the multivariable model. In addition, sensitivity analyses were conducted by removing TC from the model and by replacing TC with LDL-C. The association between TG/HDL-C and kidney stone status remained statistically significant and directionally consistent across these sensitivity analyses, indicating that the main findings were not solely driven by collinearity among lipid-related variables. The detailed correlation matrices, VIF results, and sensitivity analyses are presented in Supplementary Figure S3, Supplementary Figure S4, Supplementary Table S2, and Supplementary Table S3.

To address the possibility of overadjustment or collider bias related to renal function-related variables, we performed additional sensitivity analyses. First, the association between TG/HDL-C and kidney stone status was reexamined after excluding renal function-related biomarkers, including creatinine, urea, uric acid, and calcium, from the fully adjusted model. Second, we fitted a model adjusted only for demographic and metabolic factors. Third, we repeated the analysis after excluding participants with impaired renal function. Across these sensitivity analyses, the association between TG/HDL-C and kidney stone status remained directionally consistent with the main analysis. The detailed results are presented in Supplementary Table S4.

As shown in Figure 2, the prevalence of kidney stones increased significantly with increasing TG/HDL-C ratios. The red curve indicated a statistically significant nonlinear relationship (P for nonlinearity = 0.008). The overall association was also significant (P for overall association < 0.001), suggesting that TG/HDL-C was significantly associated with the prevalence of kidney stones.

Figure 2.

Dose-response association between TG/HDL-C and kidney stone prevalence. Restricted cubic spline analysis showed a significant nonlinear association between TG/HDL-C and kidney stone prevalence (P for nonlinearity = 0.008; P for overall association < 0.001), after adjustment for relevant covariates: (TC, age, sex, BMI, hypertension, diabetes, TBIL, Cre, urea, Ca and UA).

As shown in Figure 3, this forest plot presents the adjusted odds ratios (ORs) and 95% confidence intervals (CIs) for the association between TG/HDL-C and kidney stone prevalence in the overall population and across prespecified subgroups. In the overall population (N = 1,174), higher TG/HDL-C was associated with increased odds of kidney stones (OR, 2.16; 95% CI, 1.73–2.70; p < 0.001). No significant effect modification was observed by age (P for interaction = 0.483), sex (P for interaction = 0.494), BMI (P for interaction = 0.831), hypertension (P for interaction = 0.398), or diabetes status (P for interaction = 0.330). The association was significant among participants without elevated cystatin C (OR, 2.31; 95% CI, 1.79–2.99; p < 0.001), but not among those with elevated cystatin C (OR, 1.09; 95% CI, 0.67–1.78; p = 0.733). However, the interaction was not statistically significant (P for interaction = 0.072).

Figure 3.

Subgroup analyses of the association between TG/HDL-C and kidney stone prevalence. The forest plot shows adjusted odds ratios (ORs) and 95% confidence intervals (CIs) for the overall population and prespecified subgroups. Higher TG/HDL-C was associated with increased odds of kidney stones in the overall population, with no significant interactions across subgroups.

As shown in Table 3, the weighted demographic baseline characteristics of the included participants shown in Table 3. TG/HDL-C ratio was 3.16 (2.94,3.38) in the stone group, much higher than the 2.69 (2.60,2.77), p < 0.001.

Table 3.

NHANES database population characteristics baseline Table.

Characteristic	Nonstone formers (n = 13431)	Stone formers (n = 1418)	P value
Age (years)	46.76 (46.25, 47.27)	53.27 (52.24, 54.29)	<0.001
PIR	2.92 (2.84, 3.00)	2.93 (2.80, 3.05)	0.928
BMI (kg/m2)	28.88 (28.67, 29.08)	30.77 (30.30, 31.23)	<0.001
Serum Creatinine (mg/dL)	0.87 (0.86, 0.88)	0.92 (0.89, 0.96)	0.003
Urine Albumin Creatinine Ratio (mg/g)	33.95 (28.82, 39.08)	61.19 (36.29, 86.09)	0.041
TG/HDL-C ratio	2.69 (2.60, 2.77)	3.16 (2.94, 3.38)	<0.001
Serum Cholesterol (mg/dL)	192.12 (190.94, 193.30)	190.32 (187.34, 193.29)	0.225
Serum Calcium (mg/dL)	9.34 (9.33, 9.36)	9.33 (9.29, 9.36)	0.368
Sex (%)			<0.001
Male	47.76 (46.84, 48.68)	53.71 (50.39, 57.01)
Female	52.24 (51.32, 53.16)	46.29 (42.99, 49.61)
Race (%)			<0.001
Mexican American	8.87 (7.47, 10.50)	6.55 (5.03, 8.49)
White	71.16 (68.75, 73.46)	80.86 (77.71, 83.65)
Black	11.80 (10.31, 13.46)	5.55 (4.36, 7.06)
Other Race	8.17 (7.28, 9.15)	7.04 (5.52, 8.93)
Education Level (%)			0.536
Less than high school	16.11 (14.82, 17.49)	17.55 (15.02, 20.41)
High school	22.87 (21.44, 24.36)	22.66 (19.68, 25.93)
More than high school	61.02 (58.80, 63.19)	59.79 (55.95, 63.52)
Marital Status (%)			0.005
Cohabitation	63.27 (61.70, 64.81)	68.78 (65.00, 72.32)
Solitude	36.73 (35.19, 38.30)	31.22 (27.68, 35.00)
Physical Activity (%)			0.004
Never	26.55 (25.38, 27.75)	32.40 (28.97, 36.02)
Moderate	32.59 (31.42, 33.78)	30.47 (27.08, 34.09)
Vigorous	40.86 (39.45, 42.29)	37.13 (33.42, 41.01)
Alcohol (%)			0.574
Yes	18.92 (17.78, 20.12)	20.49 (17.49, 23.85)
No	60.29 (58.37, 62.18)	58.82 (54.56, 62.94)
Unclear	20.79 (19.21, 22.46)	20.70 (17.69, 24.06)
Hypertension (%)			<0.001
No	68.82 (67.40, 70.21)	52.91 (49.21, 56.58)
Yes	31.18 (29.79, 32.60)	47.09 (43.42, 50.79)
Smoked (%)			0.023
Yes	43.65 (42.17, 45.15)	48.60 (44.08, 53.15)
No	56.35 (54.85, 57.83)	51.40 (46.85, 55.92)
Total Water (%)			0.490
<2,000 ml	48.80 (47.53, 50.07)	51.11 (47.01, 55.19)
≥2,000 < 2,500 ml	8.09 (7.52, 8.69)	8.05 (6.33, 10.18)
≥2,500 ml	28.37 (27.10, 29.67)	28.05 (24.40, 32.00)
Unclear	14.74 (13.78, 15.76)	12.79 (10.73, 15.19)
Diabetes (%)			<0.001
No	91.31 (90.53, 92.02)	81.81 (79.17, 84.19)
Yes	8.69 (7.98, 9.47)	18.19 (15.81, 20.83)
Asthma (%)			0.361
No	85.22 (84.22, 86.16)	84.13 (81.62, 86.36)
Yes	14.78 (13.84, 15.78)	15.87 (13.64, 18.38)

For continuous variables: survey-weighted mean (95% CI), P-value was by survey-weighted linear regression.

For categorical variables: survey-weighted percentage (95% CI), P-value was by survey-weighted Chi-square test.

BMI: Body mass index (kg/m²); PIR: Ratio of family income to poverty.

As shown in Table 4, a higher TG/HDL-C ratio was associated with greater odds of kidney stones. This association remained statistically significant in the fully adjusted model (Model 3) (OR, 1.04; 95% CI, 1.01–1.07), indicating that each one-unit increase in the TG/HDL-C ratio was associated with a 4% increase in the odds of kidney stones. When the TG/HDL-C ratio was categorized into tertiles, participants in the higher tertile had significantly higher odds of kidney stones than those in the lower tertile in the fully adjusted model (OR, 1.18; 95% CI, 1.01–1.39), whereas the association for the middle tertile was not statistically significant (OR, 1.11; 95% CI, 0.95–1.29). A significant linear trend across increasing tertiles was observed in the fully adjusted model (P for trend = 0.032).

Table 4.

Higher TG/HDL-C ratio was associated with a higher prevalence of kidney stones.

Characteristic	Model 1 OR (95% CI)	Model 2 OR (95% CI)	Model 3 OR (95% CI)
TG/HDL-C ratio	1.02 (1.01, 1.03)	1.02 (1.00, 1.03)	1.04 (1.01, 1.07)
Categories
Tertile 1 (Lower)	1	1	1
Tertile 2 (Middle)	1.42 (1.23, 1.64)	1.25 (1.08, 1.45)	1.11 (0.95, 1.29)
Tertile 3 (Higher)	1.77 (1.54, 2.03)	1.45 (1.26, 1.68)	1.18 (1.01, 1.39)
P for trend	<0.001	<0.001	0.032

Model 1: Unadjusted.

Model 2: Adjusted for age, race sex.

Model 3: Adjusted for all covariates.

As shown in Figure 4, the relationship between the TG/HDL-C ratio and the prevalence of kidney stone was further investigated using a generalized additive model and smooth curve fitting. The results showed a positive linear association between the TG/HDL-C ratio and kidney stone prevalence.

Figure 4.

Dose-response and threshold effect analysis of TG/HDL-C and kidney stone prevalence. Smooth curve fitting using a generalized additive model showed a positive association between TG/HDL-C and the prevalence of kidney stones.

As shown in Table 5, Subgroup analyses were performed to assess the robustness of the association between the TG/HDL-C index and kidney stones (Table 5). We found that both of these factors, male (OR, 1.037; 95% CI,1.001–1.075), female (OR, 1.055; 95% CI, 1.004–1.109), white (OR, 1.051; 95% CI, 1.014–1.089), BMI ≥ 25kg/m²(OR, 1.041; 95% CI, 1.009–1.073), ≥60 years (OR, 1.047; 95% CI, 1.001–1.095), non-hypertensive group (OR, 1.048; 95% CI, 1.009–1.088) and the diabetic group (OR, 1.037; 95% CI, 1.003–1.073) showed a positive correlation. We also tested for an interaction with sex, age, BMI, hypertension, and diabetes. However, no p-correlation was detected with meeting a statistically significant interaction, indicating that the association was independent of sex, age, BMI, hypertension, and diabetes (p > 0.05 for all interactions).

Table 5.

Subgroup analysis between TG/HDL-C ratio with kidney stone prevalence.

Characteristic	Model 1 OR (95% CI)	Model 2 OR (95% CI)	Model 3 OR (95% CI)	P for interaction*
Sex				0.322
Male	1.011 (0.997, 1.024)	1.011 (0.997, 1.024)	1.037 (1.001, 1.075)
Female	1.041 (1.018, 1.065)	1.034 (1.011, 1.057)	1.055 (1.004, 1.109)
Race				0.270
Mexican American	1.003 (0.969, 1.038)	0.998 (0.960, 1.036)	0.980 (0.907, 1.060)
White	1.016 (1.004, 1.029)	1.015 (1.002, 1.028)	1.051 (1.014, 1.089)
Black	1.014 (0.955, 1.077)	1.015 (0.954, 1.080)	1.058 (0.960, 1.165)
Other Race	1.017 (1.006, 1.028)	1.046 (1.002, 1.091)	1.093 (0.986, 1.212)
Hypertension				0.574
Yes	1.020 (1.006, 1.035)	1.015 (1.000, 1.031)	1.030 (0.985, 1.076)
No	1.015 (0.999, 1.032)	1.006 (0.988, 1.023)	1.048 (1.009, 1.088)
BMI				0.874
<25 kg/m2	1.039 (0.977, 1.106)	1.014 (0.948, 1.085)	1.042 (0.959, 1.133)
≥25 kg/m2	1.014 (1.003, 1.026)	1.008 (0.996, 1.020)	1.041 (1.009, 1.073)
Age (years)				0.945
<60	1.019 (1.006, 1.032)	1.014 (1.001, 1.028)	1.035 (0.997, 1.075)
≥60	1.039 (1.016, 1.062)	1.024 (1.001, 1.047)	1.047 (1.001, 1.095)
Diabetes				0.841
Yes	1.017 (1.003, 1.030)	1.011 (0.996, 1.026)	1.037 (1.003, 1.073)
No	1.020 (0.999, 1.041)	1.013 (0.992, 1.035)	1.062 (0.999, 1.129)

*Means only in model 3.

Model 1 = Unadjusted.

Model 2 = Adjusted for age, race, and sex.

Model 3 = Adjusted for sex, age, race, education level, marital status, PIR, asthma, physical activity, alcohol consumption, BMI, blood calcium, serum creatinine, diabetes, smoking status, LDL-C, TC, hypertension, urine albumin-to-creatinine ratio, and total water intake.

The selection of candidate variables from the clinical center data was based on the results of the univariable analyses, combined with their clinical relevance, potential roles as confounders, and the available variables in the clinical center data. Variables such as age, sex, BMI, hypertension, diabetes, TC, ApoB, TBIL, Cys-C, Cr, Urea, UA, Ca, and TG/HDL-C were entered into the least absolute shrinkage and selection operator (LASSO) regression to reduce dimensionality and identify the optimal set of predictors for model development.

1. Ten-fold cross-validation for selecting the optimal penalty parameter (λ).

As shown in Figure 5, ten-fold cross-validation was used to determine the optimal value of the regularization parameter λ in the LASSO regression model. The curve shows the relationship between the binomial deviance and log(λ). The vertical dotted line indicates the value of λ that minimized the cross-validated deviance, representing the optimal penalty used for variable selection.

Figure 5.

Selection of characteristic variables using LASSO regression. (A) Ten-fold cross-validation was used to select the optimal penalty parameter (λ). (B) LASSO coefficient profiles of candidate predictors, with variables retaining non-zero coefficients at the optimal λ selected for subsequent modeling.

1. LASSO coefficient profiles for all candidate predictors.

As shown in Figure 5, the coefficient profiles for each predictor are plotted against log(λ). As the degree of penalization increased, the coefficients shrank progressively toward zero. At the optimal λ (vertical dashed line), a subset of predictors retained non-zero coefficients, indicating variables selected for inclusion in the subsequent predictive modeling.

Based on the LASSO regression results and clinical relevance, we included sex, hypertension, diabetes,age, BMI, total cholesterol (TC), total bilirubin (TBIL), TG/HDL-C ratio, calcium (Ca), uric acid (UA) in the development of seven machine learning models.

1. Receiver Operating Characteristic (ROC) Curves for Test Dataset

As shown in Figure 6A, the receiver operating characteristic (ROC) curves compare the discriminative performance of seven prediction models: logistic regression, decision tree, random forest, extreme gradient boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), support vector machine (SVM), and artificial neural network (ANN). Among these models, the random forest model showed the best discrimination, with an area under the ROC curve (AUC) of 0.818 (95% CI, 0.775–0.859).

Figure 6.

Construction of Seven Machine Learning Models for Classification of Kidney Stone Status (A) Receiver operating characteristic curves comparing model discrimination in the test dataset. (B) Calibration curves assessing agreement between predicted probabilities and observed outcomes. (C) Decision curve analysis evaluating the clinical net benefit of each model across different threshold probabilities.

1. Calibration Curves for Test Dataset

As shown in Figure 6B, the calibration curves assess the agreement between predicted probabilities and observed event rates for the models in the validation set. The models showed varying degrees of calibration. Logistic regression, random forest, XGBoost, and LightGBM demonstrated good calibration, with Brier scores ranging from 0.175 to 0.208. Among these models, random forest had the lowest Brier score, at 0.175 (95% CI, 0.154–0.194).

1. Decision Curve Analysis for Test Dataset

As shown in Figure 6C, Decision curve analysis showed that the random forest and logistic regression models provided greater net benefit than the treat-all and treat-none strategies across a range of threshold probabilities. In the present study, this net benefit should be interpreted in the context of selecting individuals for additional kidney stone screening or preventive metabolic evaluation, rather than initiating direct treatment. These findings suggest that the models may have potential value for risk stratification in health examination settings.

These results suggest that the machine learning models showed different levels of classification performance and predictive consistency in the test dataset. Among the models evaluated, random forest performed best, achieving an AUC of 0.818 (95% CI, 0.775–0.859) and showing the greatest net clinical benefit in the decision curve analysis. In addition to AUC, the random forest model achieved a sensitivity of 0.756, specificity of 0.732, precision of 0.747, an F1 score of 0.751, and a Brier score of 0.175 (95% CI, 0.154–0.194), indicating good overall discrimination, calibration, and clinical utility (Supplementary Table S1). Based on these results, the random forest model was selected as the final model for subsequent SHAP analysis.

 

1. Feature Importance for Random Forest Model

As shown in Figure 7A, panel A presents the feature importance of each variable in the random forest model. TG/HDL-C (triglyceride/high-density lipoprotein cholesterol ratio) was identified as the most important predictor, followed by age, total cholesterol (TC), calcium (Ca), and hypertension. In contrast, total bilirubin (TBIL), BMI, uric acid (UA), sex, and diabetes made relatively smaller contributions to the model.

Figure 7.

SHAP-based interpretation of feature contributions to kidney stone classification in the Random Forest Model (A) Feature importance ranking of variables in the Random Forest model. (B) SHAP summary plot showing the contribution of each variable to model output. (C) SHAP dependence plot illustrating the relationship between TG/HDL-C and its SHAP value. (D) SHAP waterfall plot showing the contribution of individual variables to a single prediction.

1. SHAP Value Distribution for Random Forest Model

As shown in Figure 7B, panel B presents the distribution of SHAP (Shapley additive explanations) values for each feature in the random forest model. The color gradient represents feature values, with warmer colors indicating higher values and cooler colors indicating lower values. TG/HDL-C showed a wide distribution of SHAP values, indicating a strong contribution to model predictions. Age, TC, and Ca also showed considerable variation in SHAP values, suggesting substantial predictive contributions within the model. Hypertension and TBIL had moderate contributions to model output, whereas BMI, UA, sex, and diabetes showed relatively smaller changes in SHAP values, indicating a lower impact on model output.

1. SHAP Dependence Plot for TG/HDL-C and kidney stone

As shown in Figure 7C, panel C illustrates the relationship between the SHAP values of TG/HDL-C and the interacting feature. The color gradient represents the values of the interacting feature. As TG/HDL-C increased, its SHAP values gradually changed from negative to positive, indicating a positive contribution to the model’s predictions. Specifically, the SHAP value increased rapidly at low to moderate TG/HDL-C levels and then tended to plateau at higher levels, suggesting a nonlinear pattern between TG/HDL-C values and model output on model output.

1. SHAP Explanation for a Single Prediction Using Random Forest

As shown in Figure 7D, panel D explains the prediction for a specific instance in the random forest model. The waterfall plot presents the contribution of each feature and its cumulative effect on the prediction. The final prediction for this instance was f(x) = 0.733, representing the model-predicted probability. TC contributed the most to the prediction, increasing it by +0.15, while age and TG/HDL-C also made positive contributions, further increasing the predicted probability. TBIL, Ca, and BMI showed smaller positive contributions to the predicted probability.

Discussion

This study systematically evaluated the association between the TG/HDL-C ratio and kidney stones. The results indicated that elevated TG/HDL-C levels were significantly associated with kidney stone status. After adjusting for demographic characteristics and metabolic factors, each 1-unit increase in TG/HDL-C was associated with higher odds of kidney stones (OR, 2.16; 95% CI, 1.73–2.70). In stratified analyses, this association remained consistent across subgroups defined by age, sex, BMI, hypertension, and diabetes, with no significant interactions detected. A similar pattern was also observed in the NHANES population, where higher TG/HDL-C levels were associated with kidney stone prevalence, supporting the robustness of the observed association across different datasets. However, this comparison should be interpreted cautiously because kidney stone status was defined by imaging in the clinical cohort but by self-reported history in NHANES. In the machine-learning analysis conducted within the clinical cohort, the random forest model showed the best overall performance, with an AUC of 0.82 and favorable calibration and clinical net benefit. SHAP analysis further suggested that TG/HDL-C had high predictive importance within the fitted Random Forest classification model. However, this finding reflects model-specific predictive contribution rather than etiological importance or evidence of a causal mechanism. Taken together, these findings suggest that TG/HDL-C may serve as a clinically accessible metabolic marker associated with kidney stone status and may help identify individuals with a higher likelihood of stones.

Previous studies have shown that the prevalence of kidney stones is increasing and is closely associated with the components of metabolic syndrome (MetS), including obesity, hypertension, dyslipidemia, and glucose intolerance [18].

Additional correlation studies have found that kidney stones and hyperuricemia are strongly associated [19]. Furthermore, some researchers have suggested that insulin resistance (IR) is a central pathogenic mechanism of metabolic syndrome (MetS), and the bidirectional relationship between hyperuricemia and IR is recognized as an important mechanism underlying metabolic disorders [20,21]. As a major component of MetS, and given the kidney’s susceptibility to disturbances in glucose metabolism, diabetic nephropathy has become a major public health burden [22]. Reducing the prevalence of obesity and type 2 diabetes plays an important role in the prevention of kidney stones [23]. Longer durations of diabetes, poor glycemic control, and the higher prevalence of kidney stones are closely related.

In 2003, McLaughlin was the first to demonstrate the clinical utility of the TG/HDL-C ratio for identifying generally healthy individuals with insulin resistance [24]. Additionally, some studies suggest that TG/HDL-C is not only simple to use but also comparable to other indices derived from combined biochemical and anthropometric measures. Furthermore, it has shown high sensitivity and specificity in obese individuals [9]. Other studies have shown that IR is often associated with dyslipidemia, including elevated triglyceride levels and low HDL-C levels. In individuals with hyperuricemia, TG/HDL-C has also been reported to be strongly associated with the condition [25].

Blanca Guadalupe Baez-Duarte et al. found that TG/HDL-C is highly sensitive for identifying metabolic syndrome (MetS), thereby simplifying the identification and prediction of MetS while reducing the need for additional laboratory tests [26]. However, the relationship between the TG/HDL-C ratio and kidney stones requires further investigation. Several possible explanations may account for this association. Patients with kidney stones often have IR. Japanese researchers have reported that metabolic syndrome increases insulin resistance, thereby increasing the risk of kidney stone formation [27]. It has also been shown that insulin resistance lowers urinary pH and impairs ammonia production under acid load. Additionally, insulin resistance is associated with elevated plasma free fatty acid levels. These free fatty acids can enter proximal renal tubular cells and interfere with the utilization of glutamine for ammonia production, particularly when urinary pH is below 5.0. This may promote the formation of uric acid stones by reducing the solubility of uric acid [28,29]. Prior studies have suggested that hyperlipidemia may activate resident renal cells and increase the production of profibrotic cytokines and chemokines, which may in turn be associated with increased cystatin C expression. These inflammatory and profibrotic signals may promote monocyte/macrophage infiltration into glomerular tissue and extracellular matrix deposition, thereby contributing to local renal inflammation. Renal inflammation has been implicated in the initiation and progression of kidney stone disease [28,30]. Alternatively, inflammation and oxidative stress may play important roles in the association between the TG/HDL-C ratio and kidney stones. Studies have shown that insulin resistance reduces the release and synthesis of nitric oxide (NO) while increasing oxidative stress and inflammatory cytokine levels in endothelial cells [31]. In the renal microenvironment, pro-inflammatory factors such as osteopontin, the heavy chain of inter-alpha inhibitor, collagen, and zinc contribute to inflammatory damage in the nephron. This urothelial injury further increases crystal adhesion [32]. Additional experimental studies have shown that statins, owing to their antioxidant effects, may reduce the development of kidney stones. Using a rat model, Tsujihata et al. found that atorvastatin treatment significantly reduced urinary levels of renal tubular injury markers (NAG and OS) and decreased renal tubular cell apoptosis [33]. Overall, this study suggests that the TG/HDL-C ratio is significantly associated with kidney stone prevalence. As an inexpensive marker ofIR, the TG/HDL-C ratio may help identify high-risk individuals for early intervention, thereby preventing or delaying the progression of kidney stone disease. Large cohort studies involving different ethnic populations are needed to confirm these findings.

This study has several notable strengths. First, it conducted an in-depth analysis using real-world clinical data and further examined the association between TG/HDL-C and kidney stones in an independent NHANES population. Second, through both univariate and multivariate logistic regression analyses, the study identified factors associated with kidney stone status. Additionally, the RCS model explored nonlinear relationships, and LASSO regression was employed for variable selection in model development. To further evaluate multivariable classification performance, seven machine-learning algorithms were developed and compared. In addition to conventional regression analysis, the machine-learning framework provided complementary value by integrating multiple clinical variables, capturing potential nonlinear relationships, and enabling comparison of discrimination, calibration, and clinical net benefit across alternative models. Another strength of this study is the use of interpretable artificial intelligence methods, specifically SHAP, to explain the output of complex machine-learning models. By quantifying the contribution of each feature to model output, SHAP enhances transparency and model interpretability. Furthermore, SHAP values allow estimation of the relative importance of each feature in the model. Compared with traditional feature-importance methods, SHAP provides a more consistent and informative framework for interpreting model behavior. Moreover, we used SHAP to explain individual-level model outputs, which is particularly valuable for improving the interpretability of model classification at the individual level. In supplementary exploratory analyses, Cys-C was statistically associated with both the TG/HDL-C ratio and kidney stone status (Supplementary Methods S1 and Supplementary Tables S5–S8). Because these variables were assessed within a retrospective cross-sectional framework, the Cys-C findings are hypothesis-generating and do not establish temporal ordering, mediation, or a mechanistic role. These findings should be interpreted cautiously given the retrospective cross-sectional nature of the study.

However, this study has several notable limitations. First, although multivariable adjustment was performed for known confounding factors, including demographic characteristics and relevant clinical indicators, residual confounding due to unmeasured or incompletely measured variables cannot be completely excluded. Because this study was based on a retrospective health examination cohort, several potentially important clinical factors were not systematically collected, including physical activity, medication use, genetic background, environmental exposures, and detailed medical history. In particular, stone-related information such as stone size, number, anatomical location, laterality, recurrence status, obstruction, hydronephrosis, stone composition, urine chemistry, and previous treatment history was unavailable. Therefore, kidney stone status was analyzed only as a binary imaging-based outcome rather than according to stone burden, stone phenotype, disease severity, or recurrence pattern. The lack of these detailed variables may limit the clinical interpretation of our findings and prevent further assessment of whether the association between TG/HDL-C and kidney stones differs across specific stone characteristics. Future prospective studies should collect comprehensive imaging, biochemical, compositional, and treatment-related information to validate and extend our findings. Second, although exploratory non-causal analysis was performed, the retrospective cross-sectional design did not allow us to establish temporality among TG/HDL-C, Cys-C, and kidney stones. Therefore, reverse causation remains possible. Third, kidney stone status was defined differently in the two datasets: in the clinical cohort, it was based on current imaging findings, whereas in NHANES it was based on self-reported lifetime history of kidney stones. Finally, further validation in larger, more representative datasets with more complete clinical information is warranted to reduce potential selection bias and to confirm the robustness and generalizability of the proposed framework.

Conclusion

This retrospective cross-sectional study found that a higher TG/HDL-C ratio was associated with greater odds of kidney stone status after adjustment for potential confounding factors, with a directionally similar association observed in the weighted NHANES analysis. Within the internally evaluated machine-learning models, the TG/HDL-C ratio contributed to the classification of kidney stone status. These findings support the TG/HDL-C ratio as an accessible metabolic marker associated with kidney stone status. However, the observational design precludes causal inference, and prospective studies with external model validation are needed to confirm these findings and evaluate their clinical utility.

Funding Statement

This work was supported by the National Natural Science Foundation of China (82370768), Excellent Innovation Team in Scientific Research of Anhui Provincial Colleges (2022AH010071), and Natural Science Research Project of Anhui Educational Committee (2023AH040372).

Disclosure statement

The authors have no relevant financial or non-financial interests to disclose.

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki (2013 revision) and was approved by the Ethics Committee of the First Affiliated Hospital of Anhui Medical University (approval no. PJ2025-10-71). The requirement for written informed consent was waived by the Ethics Committee of the First Affiliated Hospital of Anhui Medical University owing to the retrospective study design.

Data availability statement

The NHANES data that support the findings of this study are publicly available from the Centers for Disease Control and Prevention (CDC) National Health and Nutrition Examination Survey repository at https://www.cdc.gov/nchs/nhanes/. These data were derived from publicly available resources. The clinical cohort data that support the findings of this study are available on reasonable request from the corresponding author, [Zongyao Hao]. The clinical cohort data are not publicly available due to privacy and ethical restrictions, as they contain information that could compromise the privacy of research participants. Access to these data is subject to approval by the relevant ethics committee.