Association between the triacylglycerol–glucose index, TyG-BMI and hypertension in children and adolescents: a cross-sectional study with machine-learning-based risk stratification

In a finding that could reshape how doctors think about high blood pressure in the young, a research team in Henan Province, China, has traced a striking connection between two simple metabolic markers and hypertension in children and adolescents. The study, published in Frontiers in Endocrinology, analyzed data from 25,886 children and adolescents aged 7 to 17 years across 18 prefecture-level cities in Henan Province, and found that higher values of the triacylglycerol–glucose (TyG) index and its composite with body mass index (TyG-BMI) were associated with a higher prevalence of hypertension. The association was stronger for TyG-BMI than for TyG alone, suggesting that the combination of adiposity and glucose–lipid dysregulation identifies a particularly high-risk group among the young.

The TyG index is calculated from two routinely measured blood analytes — serum triglycerides and fasting plasma glucose — using the formula ln[triglycerides (mg/dL) × fasting plasma glucose (mg/dL)/2]. When insulin resistance develops, the suppressive effect of insulin on adipocyte lipolysis is attenuated, increasing the release of free fatty acids, which in turn promotes hepatic triglyceride synthesis; at the same time, insulin-mediated glucose uptake in muscle and adipose tissue is reduced, raising fasting plasma glucose. By integrating these two analytes, the TyG index reflects the interaction between glucose and lipid metabolism and serves as a practical surrogate marker of insulin resistance. Combining the TyG index with body mass index (TyG-BMI) has been reported to improve the assessment of insulin resistance relative to TyG alone, because it captures adiposity in addition to glucose–lipid dysregulation.

Hypertension in children and adolescents has emerged as a public health issue that can no longer be ignored. An updated systematic review and meta-analysis of 96 population-based studies, covering 443,914 children and adolescents across 21 countries, estimated the pooled global prevalence of childhood hypertension at 4.28% when hypertension was defined from repeated in-office measurements, at 6.67% when out-of-office measurements were also taken into account, and reported that the prevalence nearly doubled between 2000 and 2020, increasing from 3.40 to 6.53% in boys and from 3.02 to 5.82% in girls. This increase is closely linked to rising obesity rates, reduced physical activity, higher dietary sodium intake and urbanization. Childhood hypertension is closely associated with subclinical cardiovascular disease, including left ventricular hypertrophy, increased carotid intima-media thickness, elevated pulse wave velocity, atherosclerotic change, retinal microangiopathy and proteinuria. A long-term follow-up analysis of the International Childhood Cardiovascular Cohort Consortium, published in 2024, reported that childhood hypertension was associated with an approximately 2.3-fold higher risk of fatal or non-fatal cardiovascular events over more than 30 years.

The research team recruited a total of 28,844 students from primary, junior high and senior high schools across 18 prefecture-level cities in Henan Province, China, between 2023 and 2024, and collected baseline demographic, anthropometric, lifestyle and laboratory data. After exclusions for missing data and for participants outside the reference age range, 25,886 participants aged 7 to 17 years remained in the primary analysis. Blood pressure was measured with validated oscillometric electronic monitors using a protocol that required a quiet, separate room at an ambient temperature of 18–24 °C, no strenuous exercise or food and no caffeine-containing drink in the preceding hour, and at least five minutes of seated rest. To reduce the influence of the white-coat effect, blood pressure was measured three times on non-consecutive days at approximately the same time of day, and the participant’s blood pressure was taken as the mean of the last two of the three measurements. Hypertension was defined as a systolic or diastolic blood pressure at or above the 95th percentile for the corresponding age, sex and height, according to the Chinese national reference values for children and adolescents aged 7 to 17 years.

Among the 25,886 participants, 1,798 (6.95%) had hypertension according to the primary definition. Median TyG was 8.05 (interquartile range 7.77–8.35) overall and was higher in participants with hypertension than in those without (8.23 versus 8.03; p < 0.001). Median TyG-BMI was 154 (interquartile range 133–180) overall and was also higher in participants with hypertension (186 versus 152; p < 0.001). Participants with hypertension were older, had a higher heart rate, higher low-density lipoprotein cholesterol, total cholesterol, total protein, serum creatinine, uric acid and hemoglobin, a higher BMI, a body shape index and waist-to-height ratio, and lower high-density lipoprotein cholesterol; they were more frequently male, reported snoring more often, slept less on workdays and spent more time sedentary.

Each 1-standard-deviation increment in TyG was associated with hypertension in all models, with the fully adjusted Model IV yielding an odds ratio of 1.306 (95% confidence interval 1.145–1.489). The association was graded across quartiles: relative to the first quartile, the odds ratio in the fourth quartile was 1.45 (95% confidence interval 1.17–1.81) in Model IV, with p for trend < 0.001. TyG-BMI showed stronger associations than TyG at every level of adjustment. Each 1-standard-deviation increment in TyG-BMI was associated with hypertension with an odds ratio of 2.109 (95% confidence interval 1.985–2.241) in Model IV, and across quartiles the fourth quartile was associated with a more than threefold higher odds of hypertension relative to the first quartile (3.79, 95% confidence interval 3.12–4.61; p for trend < 0.001). In a head-to-head comparison, the area under the receiver-operating characteristic curve (AUC) of TyG-BMI for hypertension was 0.7087 and that of TyG 0.6250, with the DeLong test for correlated curves giving Z = −12.28 (p < 0.001).

Restricted cubic spline analysis showed a monotonic positive association between TyG and hypertension that departed modestly from linearity (p for non-linearity 0.004) and no evidence of non-linearity for TyG-BMI (p for non-linearity 0.346); neither curve had an internal minimum. For heart rate, the spline indicated a markedly non-linear association with hypertension (p for non-linearity < 0.001); the curve was essentially flat, with an odds ratio close to 1.0, up to approximately 70 beats per minute and then increased steeply, reaching an odds ratio of about 4.5 at 120 beats per minute. However, when the interaction was tested formally, the multiplicative interaction between heart rate and the exposure was not statistically significant, and no interior cut-point was identified; heart rate was therefore not dichotomized for the main regression models.

Because the study has a cross-sectional design, the contributions of waist-to-height ratio, uric acid and heart rate were evaluated as effect modifiers rather than as mediators. On the multiplicative scale, no product term survived correction for the six interaction tests performed (Bonferroni-adjusted significance level 0.0083). On the additive scale, however, all six relative excess risk due to interaction (RERI) estimates were positive with confidence intervals excluding zero, with the largest being 0.444 (95% confidence interval 0.314–0.575) for heart rate in the TyG-BMI panel. Because the odds ratio is non-collapsible, an additive-scale departure from additivity can coexist with a null multiplicative product term, so the two scales answer different questions and neither is intrinsically more correct; on cross-sectional data neither can be read as biological interaction. The researchers therefore no longer state that there was no effect modification: the data provide no consistent evidence of interaction on the multiplicative scale, but several additive-scale estimates are positive, indicating departure from additivity.

The team also developed eleven machine-learning models on a stratified 70:30 split, with imputation and feature selection performed strictly within the training set and discrimination evaluated in the held-out validation set. In the held-out validation set, discrimination was moderate, with the best-performing model being XGBoost in both panels (AUC 0.7544, 95% confidence interval 0.7302–0.7777, for TyG and 0.7534, 0.7298–0.7759, for TyG-BMI). Adding either index to the same baseline model already containing BMI increased discrimination only marginally (about 0.003 in AUC). Interpretability analyses with SHapley Additive exPlanations (SHAP) identified BMI, heart rate and the corpulence index as the most influential features. Because the outcome is imbalanced (approximately 7%), the classification metrics obtained at the default probability threshold of 0.50 are dominated by the majority class, and the very low sensitivities obtained at that threshold reflect the threshold and the low prevalence of the outcome rather than an absence of discriminative ability.

The researchers emphasize that the cross-sectional design precludes any conclusion about temporal order or causality; all estimates describe associations, and prospective cohort studies are needed to determine whether changes in TyG precede the development of hypertension. The machine-learning models built from routinely available variables showed only moderate discrimination and remain hypothesis-generating tools for risk stratification rather than validated prediction models. The findings are consistent in direction with the small number of comparable pediatric studies, including an analysis of 12,087 children and adolescents from the 2017 China National Nutrition and Health Surveillance that reported associations of TyG with several hypertension phenotypes, although effect sizes differed plausibly because of differences in outcome definition and covariate adjustment. As obesity is a central risk factor for childhood hypertension, and is already associated with adverse cardiovascular and renal measures in youth, the stronger association observed for TyG-BMI is plausible. Longitudinal studies with cluster-level information and detailed dietary assessment are required to clarify whether these associations are causal and to establish their clinical utility.

Subject of Research: TyG and TyG-BMI indices linked to hypertension in Chinese children and adolescents

Article Title: Association between the triacylglycerol–glucose index, TyG-BMI and hypertension in children and adolescents: a cross-sectional study with machine-learning-based risk stratification

Article References: Zhu, J., Zhang, C., Sun, J., Du, H., Zhao, W., Tian, G., Zhao, J., Tian, D., Wang, M., Liu, Z., Liu, M., & Hao, Y. (2026). Association between the triacylglycerol–glucose index, TyG-BMI and hypertension in children and adolescents: a cross-sectional study with machine-learning-based risk stratification. Frontiers in Endocrinology, 17, Article 1964709. https://doi.org/10.3389/fendo.2026.1964709

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1964709

Keywords: Association, between, triacylglycerol, glucose, index, TyG-BMI, hypertension, children, adolescents, cross-sectional, machine-learning-based, risk