Association between the triglyceride-glucose-waist circumference index and obstructive sleep apnoea: a hospital-based case–control study in China and the development of machine learning models

Abstract

Objectives

We aimed to examine the association between the triglyceride-glucose-waist circumference (TyG-WC) index and obstructive sleep apnoea (OSA) and to develop machine learning (ML) models to identify OSA.

Design

Case–control study and development of six ML models.

Setting

A sleep medicine centre at a tertiary hospital in China.

Participants

534 patients with OSA diagnosed by polysomnography between 2017 and 2025, matched with 534 OSA-free individuals.

Primary and secondary outcome measures

The primary outcome was the association between TyG-WC and OSA, assessed using multivariable logistic regression adjusted for age, sex, education, marital status, hypertension, diabetes, smoking and alcohol use. Secondary outcomes were the dose-response and threshold effects, subgroup interactions, external consistency testing in the National Health and Nutrition Examination Survey (NHANES), sensitivity analyses and development of ML models (LightGBM, XGBoost, Random Forest, Support Vector Machine, K-Nearest Neighbours, Naive Bayes) incorporating TyG-WC and the same clinical variables, evaluated for discrimination, calibration and clinical utility, interpreted using SHapley Additive exPlanations (SHAP).

Results

TyG-WC was linearly associated with OSA (adjusted OR=1.067, 95% CI 1.055 to 1.079; p<0.001; p for non-linearity=0.827), and this association remained robust in NHANES and sensitivity analyses. Significant interactions were observed by age and hypertension status. In the test set, six ML models demonstrated good predictive performance (area under the receiver operating characteristic curve (AUC)=0.732–0.803). XGBoost achieved the highest discrimination (AUC=0.803, 95% CI 0.755 to 0.851), with no evidence of miscalibration (Brier score=0.183; Hosmer-Lemeshow p>0.05) and positive net benefit across threshold probabilities. SHAP identified TyG-WC as the most influential predictor across all ML models.

Conclusions

TyG-WC was associated with OSA. ML models incorporating TyG-WC and clinical variables achieved good performance, supporting TyG-WC’s potential as an accessible risk marker.

Trial registration

Not applicable.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This large-scale case–control study examined the association between triglyceride-glucose-waist circumference (TyG-WC) and obstructive sleep apnoea (OSA) confirmed by polysomnography, with the association showing external consistency in National Health and Nutrition Examination Survey and robustness in sensitivity analyses.

• Machine learning models incorporating TyG-WC and clinical variables achieved good discriminative performance, with no evidence of miscalibration in three of the six models, highlighting the potential predictive value of TyG-WC for OSA.

• The study population was derived from patients presenting with sleep-related problems, which may have introduced selection bias and reduced the applicability of the findings to the general population.

Introduction

Obstructive sleep apnoea (OSA) is a sleep-related breathing disorder characterised by recurrent episodes of upper airway obstruction during sleep and is associated with metabolic and cardiovascular complications. Growing evidence suggests that beyond intermittent hypoxaemia and sleep fragmentation, OSA contributes to insulin resistance (IR) through sympathetic activation, oxidative stress and systemic inflammation, even in individuals without obesity.^1–4 Clinical studies have shown that IR correlates with polysomnographic features such as nocturnal desaturation severity and rapid eye movement-related apnoea burden.⁵ Dyslipidaemia, particularly elevated triglycerides, also frequently coexists with OSA and may further contribute to metabolic dysfunction.⁶

The triglyceride-glucose (TyG) index is widely recognised as a well-validated surrogate marker for IR, exhibiting high concordance with results obtained via the gold-standard hyperinsulinaemic-euglycaemic clamp.⁷ To better incorporate central adiposity, several TyG-derived anthropometric indices, including the TyG-body mass index (TyG-BMI) and TyG-waist circumference (TyG-WC) index, have been proposed and evaluated as markers of IR and diabetes risk.^{8 9} Recent analyses from the National Health and Nutrition Examination Survey (NHANES) have reported significant associations between TyG-based indicators and OSA, with TyG-WC showing the strongest discriminatory performance among these markers.¹⁰ A recent meta-analysis also found that individuals with OSA have higher TyG levels than healthy controls, with TyG showing modest diagnostic ability for OSA.¹¹ Additionally, studies in hypertensive and non-obese populations indicate that TyG is independently associated with OSA severity and that this association persists after adjustment for BMI and additional metabolic covariates.¹²

Although prior evidence¹³ suggests a potential association between TyG-WC and OSA, existing studies have predominantly used large-scale, population-based datasets such as NHANES, with limited validation in studies using polysomnographic (PSG) diagnosis. To assess external consistency, we also examined the association in NHANES.

This study aimed to examine the association between TyG-WC and OSA, including its dose-response and threshold effects and subgroup interactions. It also aimed to develop machine learning (ML) models incorporating TyG-WC and clinical variables to identify OSA and to evaluate their discrimination, calibration and clinical utility.

Methods

This study was designed in three stages. In the first stage, we examined the association between TyG-WC and OSA in a large case–control study, assessing its robustness with sensitivity analyses. In the second stage, we further evaluated the consistency of this association in NHANES. In the third stage, we developed six ML models incorporating TyG-WC and clinical variables and evaluated their predictive performance in identifying OSA. This study was reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement, and the machine learning analyses were reported according to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis+Artificial Intelligence (TRIPOD+AI) statement.

Case–control study

Study design and setting

We conducted a single-centre case–control study at the Sleep Centre of the Affiliated Hospital of Qingdao University. Patients with OSA confirmed by PSG who were admitted between July 2017 and March 2025 were included in the case group. Patients without OSA confirmed by PSG were included in the non-OSA group. For each patient with OSA, one non-OSA patient was randomly selected from those admitted during the same month, yielding a 1:1 matched ratio.

Participants

Patients diagnosed with OSA by PSG at the sleep centre were included in the case group; the non-OSA group comprised randomly matched patients without OSA. The exclusion criteria for both groups were: (1) age younger than 20 years or older than 90 years; (2) missing laboratory or anthropometric data required for TyG-WC calculation (fasting blood glucose (FBG), triglyceride (TG) or WC); or (3) pregnancy, which precludes valid WC assessment for metabolic indexing. No participants were excluded for missing covariate data.

Variables

TyG-WC index

The TyG-WC index was calculated based on WC, FBG and TG. These data were extracted from the hospital’s electronic medical records and were obtained on the first day of hospital admission. FBG and TG were measured using standard laboratory procedures, while WC was measured by trained nurses using standard procedures.

The calculation of TyG-WC incorporated TG (mg/dL), FBG (mg/dL) and WC (cm) as follows:^{14 15}

$$TyG=ln(TG\times FBG/2)$$

$$TyG-WC=TyG\times WC$$

Obstructive sleep apnoea

Participants presenting to the sleep centre with sleep-related problems underwent PSG for diagnostic evaluation. The Apnoea–Hypopnoea Index (AHI) was obtained from the PSG recordings and was used to determine the presence of OSA. OSA was defined as an AHI of ≥5 events/hour, whereas participants with an AHI of <5 events/hour were classified as non-OSA.

Covariates

Covariates were selected based on clinical relevance and previous literature.¹⁰ Demographic characteristics included sex (male or female), age, education level (less than middle school, middle school or above middle school) and marital status (unmarried or married). Clinical comorbidities included hypertension and diabetes (classified as yes/no). Lifestyle factors included smoking status and alcohol use (classified as yes/no). All covariate information was obtained from the Hospital Information System.

NHANES analysis

Study design and setting

We further examined the consistency of this association in NHANES,¹⁶ a nationally representative survey of the US population. Four cycles were included because sleep-related questionnaire data were available only in these cycles. The data used in this study were publicly available and de-identified before release.

Participants

Participants from the 2005–2006, 2007–2008, 2015–2016 and 2017–2018 NHANES cycles were included in the study. Exclusion criteria were as follows: (1) pregnancy or missing data required to calculate the TyG-WC index, including FBG, TG or WC; (2) missing data on OSA-related questionnaire indicators; and (3) missing data on covariates.

Variables

TyG-WC index

For participants from NHANES, TG data were obtained from the laboratory data, cholesterol: low-density lipoprotein, triglyceride and apolipoprotein; FBG data from the laboratory data, plasma fasting glucose and insulin; WC data from the examination data, body measures; and sleep-related information from the questionnaire, sleep disorders. All measurements were collected by trained NHANES professionals. TyG-WC was calculated using the formula described above.

Obstructive sleep apnoea

OSA classification was based on self-reported responses to the NHANES sleep questionnaire, following a previously used symptom-based definition,^{1–3 10} rather than a diagnosis confirmed by PSG. Participants were classified as having OSA if they endorsed at least one of the following symptom profiles: (1) habitual snoring occurring ≥3 nights per week over the past year; (2) witnessed apnoeas, including snorting, gasping or breathing cessation, with the same weekly frequency; or (3) significant daytime hypersomnolence (≥16 days in the prior month) occurring despite adequate sleep duration (>7 hours/night).

Covariates

Covariate selection was informed by prior literature and theoretical rationale.¹⁰ Sex was categorised as male or female. Age was included as a continuous variable. Race was classified into five groups: non-Hispanic white, non-Hispanic black, other Hispanic, Mexican American and other race (including multiracial individuals). Education was categorised as less than high school, high school graduate and above high school. The poverty income ratio was dichotomised as ≤1 or >1. Marital and residential status was classified as married/cohabiting (including legally married and living with a partner) or living alone (divorced, widowed, separated or never married). Hypertension and diabetes status were self-reported based on prior medical diagnoses. Alcohol use was defined as consuming more than 12 standard drinks per year, and smoking history as having smoked more than 100 cigarettes in a lifetime.

Machine learning models

All ML analyses were performed using R (V.4.5.1). Before model training, multicollinearity among the predictors was assessed using the variance inflation factor (VIF), with VIF <5 considered acceptable. We developed six ML models incorporating TyG-WC and clinical variables: LightGBM, XGBoost, Random Forest, Support Vector Machine (SVM), K-Nearest Neighbours (KNN) and Naive Bayes. The dataset was randomly partitioned into a training set and a test set in a 7:3 ratio using stratified sampling on OSA status, with a fixed random seed (set.seed=1234). For each model, hyperparameters were selected via fivefold cross-validation on the training set, with the area under the receiver operating characteristic curve as the optimisation metric. The R package/function, underlying algorithmic paradigm and hyperparameter specifications for each model are summarised in online supplemental table S1. The independent test set was not used for hyperparameter selection. Continuous predictors were standardised within the training folds only, and the same transformation was applied to the test set.

Model performance was evaluated across three dimensions. (1) Discrimination was assessed by the area under the receiver operating characteristic curve (AUC), with 95% CIs estimated using the DeLong method. Pairwise statistical comparisons across models were conducted using the non-parametric DeLong test, and p values from the pairwise DeLong tests were adjusted for multiple comparisons using the Holm-Bonferroni method. For all six algorithms, models were retrained with and without TyG-WC to evaluate its contribution to model performance. (2) Calibration was quantified using the Brier score and Hosmer-Lemeshow (H-L) goodness-of-fit test alongside calibration curves. (3) Clinical utility was quantified through decision curve analysis (DCA) to estimate the net clinical benefit. For visualisation only, the receiver operating characteristic curves were smoothed using the binormal method (pROC); all reported AUC values were calculated from the empirical (non-parametric) curve. Calibration curves were plotted using fixed-width probability bins for visualisation, whereas the H-L test used 10 risk deciles (equal-frequency bins).

SHapley Additive exPlanations (SHAP) analysis was used to interpret the ML models. SHAP was used to assess feature importance and visualise the effects of TyG-WC on individual predictions. SHAP values were computed for 100 test-set observations; TreeSHAP was used for LightGBM and XGBoost and KernelSHAP, fitted using a background set of 50 observations, was used for Random Forest, SVM, KNN and Naive Bayes.

Statistical analysis

Sample size adequacy for multivariable logistic regression was assessed using the events per variable (EPV) criterion. An EPV of ≥10 was considered adequate based on a previous simulation study.¹⁷

Statistical analyses were performed in R (V.4.5.1). Normality of continuous variables was assessed using the Shapiro-Wilk test. Continuous data were reported as mean±SD or median (IQR), while categorical data were presented as numbers (n) and percentages (%). Between-group differences in baseline characteristics were evaluated using the two-sample t-test, Mann-Whitney U test or χ² test.

Multivariable logistic regression models were used to assess the association between TyG-WC and OSA. Adjusted ORs (aORs) with 95% CIs were reported. TyG-WC was scaled by a factor of 10 (TyG-WC/10) in all regression models, and the aORs were interpreted per 10-unit increase in TyG-WC. Although participants were matched by admission month, this variable was not included in the regression models because it was not considered a confounder of the TyG-WC-OSA association. For trend analysis across tertiles, the median TyG-WC value of each tertile was entered into the model as a continuous variable, and p for trend was calculated.

Restricted cubic splines with four knots were used to assess potential non-linear associations between TyG-WC and OSA. When a non-linear association was detected, two-piecewise logistic regression was used to identify the inflection point and a likelihood ratio test (2 df) was used to compare the piecewise and linear models. Subgroup analyses were performed, and p for interaction was used to assess whether the association between TyG-WC and OSA differed across subgroups; ‘stronger’ and ‘weaker’ indicate statistically significant interactions.

Sensitivity analyses were performed to assess the robustness of the findings. First, the primary regression models were re-estimated after excluding participants with TyG-WC values in the upper and lower 5% of the distribution. Second, propensity score matching (PSM) was performed using 1:1 nearest-neighbour matching with a calliper of 0.2 SD of the logit propensity score, based on age, sex, education level, marital status, hypertension, diabetes, smoking and alcohol use. Covariate balance was assessed using standardised mean differences (SMDs), with SMD <0.1 indicating adequate balance. After PSM, the regression models were refitted in the matched sample. Two-sided p values<0.05 were considered statistically significant.

Patient and public involvement

Patients and/or the public were not involved in the design, conduct, reporting or dissemination plans of this research.

Results

Baseline characteristics

In the hospital population, 566 patients with PSG-confirmed OSA were initially identified. After applying the eligibility criteria, 534 patients were included in the OSA group and matched with 534 non-OSA individuals, giving a total of 1068 participants (figure 1). Of these, 657 (61.5%) were male and 411 (38.5%) were female. The overall median TyG-WC was 862.51 and was significantly higher in the OSA group than in the non-OSA group (table 1). The hospital population was ethnically homogeneous, with 99.8% of participants being Han Chinese.

Figure 1. Flowchart of participant selection. NHANES, National Health and Nutrition Examination Survey; OSA, obstructive sleep apnoea; TyG-WC, triglyceride-glucose-waist circumference index.

Table 1. Baseline characteristics of participants in the case–control study.

Variables	Total (n=1068)	Non-OSA (n=534)	OSA (n=534)	P value
Age (years)	57 (47–66)	61 (54–69)	52 (42–62)	<0.001
Age group (years)				<0.001
<60	600 (56.180)	235 (44.007)	365 (68.352)
≥60	468 (43.820)	299 (55.993)	169 (31.648)
Sex				<0.001
Male	657 (61.517)	256 (47.940)	401 (75.094)
Female	411 (38.483)	278 (52.060)	133 (24.906)
Ethnicity				>0.999
Han	1066 (99.813)	533 (99.813)	533 (99.813)
Non-Han	2 (0.187)	1 (0.187)	1 (0.187)
Education level				<0.001
Less than middle school	112 (10.487)	63 (11.798)	49 (9.176)
Middle school	288 (26.966)	173 (32.397)	115 (21.536)
More than middle school	668 (62.547)	298 (55.805)	370 (69.288)
Marital status				<0.001
Unmarried	51 (4.775)	13 (2.434)	38 (7.116)
Married	1017 (95.225)	521 (97.566)	496 (92.884)
Hypertension				<0.001
Yes	672 (62.921)	310 (58.052)	362 (67.790)
No	396 (37.079)	224 (41.948)	172 (32.210)
Diabetes				0.945
Yes	773 (72.378)	386 (72.285)	387 (72.472)
No	295 (27.622)	148 (27.715)	147 (27.528)
Smoking				<0.001
Yes	313 (29.307)	122 (22.846)	191 (35.768)
No	755 (70.693)	412 (77.154)	343 (64.232)
Alcohol use				<0.001
Yes	309 (28.933)	114 (21.348)	195 (36.517)
No	759 (71.067)	420 (78.652)	339 (63.483)
TyG-WC	862.512 (763.400–977.850)	793.180 (712.391–879.366)	944.626 (842.903–1,057.193)	<0.001

Data are presented as median (IQR) or n (%).

OSA, obstructive sleep apnoea; TyG-WC, triglyceride-glucose-waist circumference index.

Of 37 960 initially identified NHANES participants, 26 897 were excluded because data required to calculate TyG-WC were unavailable, 1123 because OSA-related questionnaire data were missing and 2550 because covariate data were missing, leaving 7390 participants for the final analysis (figure 1). Among them, 2937 participants (39.7%) were classified as having OSA. The overall mean TyG-WC was 866.79 and was significantly higher in the OSA group than in the non-OSA group (online supplemental table S2).

Association between TyG-WC and OSA

In the case–control study, 534 OSA events and 10 model parameters yielded an EPV of 53.4, which is above the conventional threshold of 10 for multivariable analysis. TyG-WC was associated with OSA (aOR=1.067; 95% CI 1.055 to 1.079; p<0.001), with a trend across TyG-WC tertiles (p for trend <0.001; table 2). Compared with the lowest tertile of TyG-WC (T₁), the highest tertile (T₃) was associated with OSA (aOR=8.143; 95% CI 5.585 to 11.992; p<0.001; table 2).

Table 2. Multivariable logistic regression analysis of TyG-WC and OSA in the case–control study.

Variables	Model 1		Model 2		Model 3
	OR (95% CI)	P value	aOR (95% CI)	P value	aOR (95% CI)	P value
TyG-WC (per 10 units)	1.075 (1.064 to 1.087)	<0.001	1.067 (1.055 to 1.079)	<0.001	1.067 (1.055 to 1.079)	<0.001
Tertiles
T1 (466.52–794.19, median=717.11)	1.000 (Reference)		1.000 (Reference)		1.000 (Reference)
T2 (794.23–936.93, median=862.51)	3.145 (2.286 to 4.353)	<0.001	2.875 (2.059 to 4.036)	<0.001	2.792 (1.987 to 3.944)	<0.001
T3 (937.11–1849.56, median=1038.54)	11.532 (8.167 to 16.470)	<0.001	8.351 (5.813 to 12.115)	<0.001	8.143 (5.585 to 11.992)	<0.001
P for trend		<0.001		<0.001		<0.001

Based on the distribution of TyG-WC, the 33.3rd and 66.7th percentiles were used as cut-off points to divide TyG-WC values into three equal groups (T₁, T₂ and T₃).

Model 1: No adjustment.

Model 2: Adjusted for age, sex, education level and marital status.

Model 3: Adjusted for age, sex, education level, marital status, hypertension, diabetes, smoking and alcohol use.

aOR, adjusted OR; OSA, obstructive sleep apnoea; TyG-WC, triglyceride-glucose-waist circumference index.

In NHANES, TyG-WC was associated with OSA (aOR=1.026; 95% CI 1.022 to 1.029; p<0.001), with a trend across TyG-WC tertiles (p for trend<0.001; online supplemental table S3). Compared with the lowest tertile of TyG-WC (T1), the highest tertile (T₃) was associated with OSA (aOR=2.708; 95% CI 2.377 to 3.086; p<0.001; online supplemental table S3).

Dose-response relationship

No non-linear association between TyG-WC and OSA was observed (p for overall <0.001; p for non-linearity=0.827; online supplemental figure S1). Sex-stratified analyses showed a consistent linear association between TyG-WC and OSA in both males and females (p for trend<0.001; p for non-linearity=0.162 in males and 0.562 in females; online supplemental figure S2).

In NHANES, a non-linear association between TyG-WC and OSA was observed (p for non-linearity=0.015; online supplemental figure S3). Segmented regression identified a threshold in the association between TyG-WC and OSA (likelihood ratio test p=0.009; online supplemental table S4). Below the threshold of 1016.547, TyG-WC was significantly associated with OSA (aOR=1.029; 95% CI 1.024 to 1.035). Above this threshold, the association remained statistically significant but was attenuated (aOR=1.014; 95% CI 1.003 to 1.025; online supplemental table S4 and figure S4).

Subgroup analysis

Significant interactions were observed for age and hypertension (online supplemental figure S5). The association was significantly stronger in participants aged <60 years (aOR=1.081; 95% CI 1.065 to 1.100) than in those aged ≥60 years (aOR=1.053; 95% CI 1.036 to 1.071; p for interaction=0.021). Similarly, the association was significantly stronger in participants without hypertension (aOR=1.087; 95% CI 1.066 to 1.110) than in those with hypertension (aOR=1.056; 95% CI 1.041 to 1.071; p for interaction=0.022). No significant interactions were observed for the remaining subgroups (online supplemental figure S5).

In NHANES, significant interactions were observed for age (p for interaction <0.001), hypertension (p for interaction<0.001) and alcohol consumption (p for interaction=0.030; online supplemental figure S6), whereas no significant interactions were observed for the remaining subgroups (all p for interaction ≥0.05). Similarly, the association was significantly stronger in participants aged <60 years, without hypertension and with a history of alcohol consumption (online supplemental figure S6).

Sensitivity analyses

After excluding participants in the upper and lower 5% of the TyG-WC distribution, the case–control study included 960 participants, and NHANES included 6650 participants. The association between TyG-WC and OSA remained robust in both studies (case–control study: aOR=1.070; 95% CI 1.056 to 1.085; p<0.001; NHANES: aOR=1.027; 95% CI 1.023 to 1.031; p<0.001).

After PSM, 364 pairs were included; SMDs were below 0.10 for all covariates except age (SMD=0.107; online supplemental table S5 and figure S7). Age was nonetheless retained as a continuous covariate in the adjusted model to account for this residual imbalance. TyG-WC remained associated with OSA after matching (aOR=1.063; 95% CI 1.050 to 1.077; p<0.001; online supplemental table S6).

Machine learning models

Multicollinearity was low among the predictors, with VIF values ranging from 1.07 to 1.73 (online supplemental table S7). Six ML models (LightGBM, XGBoost, Random Forest, SVM, KNN and Naive Bayes) incorporating TyG-WC and clinical variables were developed and evaluated for OSA identification. The training and test sets comprised 748 (374 with OSA) and 320 (160 with OSA) participants, respectively.

In the training set, AUCs ranged from 0.789 to 0.875 (figure 2A). In the test set, AUCs ranged from 0.732 to 0.803, with XGBoost showing the highest discrimination (AUC=0.803, 95% CI 0.755 to 0.851). The remaining models ranked as follows: Random Forest (AUC=0.799), LightGBM (AUC=0.795), SVM (AUC=0.790), Naive Bayes (AUC=0.777) and KNN (AUC=0.732; figure 2B). KNN showed lower discrimination than LightGBM, XGBoost, Random Forest and SVM (Holm-adjusted p<0.05), but did not differ from Naive Bayes (Holm-adjusted p=0.161); no significant differences were observed among the other five models (online supplemental table S8). Excluding TyG-WC significantly reduced the discrimination of all six models (ΔAUC ranging from −0.076 to −0.132; all DeLong p<0.01), indicating that TyG-WC contributed substantially to the discrimination across all ML models (online supplemental table S9).

Figure 2. Receiver operating characteristic curves of six machine learning models. (A) Training set. (B) Test set. LightGBM, XGBoost, Random Forest, SVM, KNN and Naive Bayes models are shown. AUC, area under the receiver operating characteristic curve; KNN, K-Nearest Neighbours; SVM, Support Vector Machine.

Calibration curves (figure 3A,B) and quantitative metrics showed no evidence of miscalibration for LightGBM (Brier score=0.185; H-L χ²=9.40, p=0.310), XGBoost (Brier score=0.183; H-L χ²=12.52, p=0.130) and SVM (Brier score=0.187; H-L χ²=9.18, p=0.327). In contrast, Random Forest (Brier score=0.188; H-L χ²=52.49, p<0.001), KNN (Brier score=0.212; H-L χ²=21.38, p=0.006) and Naive Bayes (Brier score=0.209; H-L χ²=82.83, p<0.001) showed evidence of miscalibration (online supplemental table S10). Despite these calibration differences, all six models showed positive net benefit in the test set at threshold probabilities up to 76% and exceeded both the treat-all and treat-none strategies at thresholds of approximately 31–76% (figure 4B).

Figure 3. Calibration curves of six machine learning models. (A) Training set. (B) Test set. Curves illustrate the agreement between the predicted probability of OSA and the observed event rate. KNN, K-Nearest Neighbours; OSA, obstructive sleep apnoea; SVM, Support Vector Machine.

Figure 4. Decision curve analysis of six machine learning models. (A) Training set. (B) Test set. Curves show the net clinical benefit across threshold probabilities. KNN, K-Nearest Neighbours; SVM, Support Vector Machine.

SHAP analysis showed that TyG-WC was the most influential predictor across all six models, consistently ranking higher than conventional predictors such as age and hypertension (figure 5). Higher TyG-WC values were associated with higher SHAP values, indicating a greater positive contribution to the predicted probability of OSA.

Figure 5. SHapley Additive exPlanations summary plots of six machine learning models. Feature importance is ranked from top to bottom based on the mean absolute SHAP value. KNN, K-Nearest Neighbours; SHAP, SHapley Additive exPlanations; SVM, Support Vector Machine.

Discussion

Statement of the principal findings

This study identified an association between TyG-WC and OSA, which was also observed in NHANES, supporting the consistency of this association across the two study populations. Subgroup analyses showed generally consistent associations, with significant interactions observed for age and hypertension. In addition, ML models incorporating TyG-WC and clinical variables showed good predictive performance for OSA, with TyG-WC consistently ranking as the most influential predictor in SHAP analysis. Overall, our findings suggest that TyG-WC is associated with OSA and may provide complementary predictive information beyond conventional clinical variables, highlighting the potential value of TyG-WC as a simple metabolic-anthropometric marker for OSA prediction.

Strengths and weaknesses of the study

A major strength of this study was the evaluation of the association in both a large clinical dataset and NHANES. In the hospital population, OSA was objectively assessed by PSG. In addition, six ML models were developed to identify OSA, and the contribution of TyG-WC to model prediction was further examined across different algorithms.

Nevertheless, several limitations warrant consideration. First, the case–control design of the hospital study and the cross-sectional design of the NHANES analysis preclude the determination of causal relationships. Second, OSA was assessed differently in the two datasets. PSG was used for diagnosis in the hospital population, whereas symptom-based questionnaire definitions were used in NHANES and may have introduced non-differential misclassification and self-report bias. Questionnaire-based definitions cannot replace PSG-confirmed OSA diagnosis. Third, several limitations related to the NHANES data should be acknowledged. Sleep questionnaire data were available only in four discrete cycles, and sampling weights were not applied because the NHANES analysis was intended to test the external consistency of the association observed in the hospital sample, rather than to produce nationally representative prevalence or population-level effect estimates. Therefore, NHANES descriptive statistics should be interpreted as sample statistics rather than US population estimates, and the lack of sampling weights may limit the generalisability of the findings to the target US population. Furthermore, the requirement for complete fasting laboratory measurements and covariate data may have resulted in the exclusion of participants with missing information and affected the composition of the analytical sample. Fourth, the primary analysis was based on a single centre, which may limit demographic representativeness and introduce selection bias. Both OSA and non-OSA participants were symptomatic patients who underwent PSG evaluation for sleep-related problems, rather than individuals from the general population. This clinical recruitment setting may limit the generalisability of the findings to asymptomatic or community-based populations.

Strengths and weaknesses in relation to other studies

Our findings are broadly consistent with existing evidence showing positive associations between TyG-related indices, including TyG, TyG-WC and TyG-BMI, and OSA.^{10 11 13 18} The present study extends this evidence by examining TyG-WC in two distinct settings. The association was observed in both the hospital-based case–control study, in which OSA was confirmed by PSG and NHANES. Although the strength and dose-response pattern differed between the two studies, the direction of the association was consistent. We further incorporated TyG-WC and clinical variables into six ML models for OSA identification, an approach that has been less explored in previous studies. TyG-WC was the most influential predictor across the models, suggesting that it contributed substantially to OSA prediction beyond its statistical association with OSA.

Meaning of the study

In the hospital sample, stronger associations were observed among participants aged <60 years and those without hypertension, with differences in baseline TyG-WC distribution across subgroups (online supplemental table S11). In the NHANES sample, stronger associations were observed among participants aged <60 years, those without hypertension and those with a history of alcohol use, with differences in baseline TyG-WC distribution across subgroups (online supplemental table S12). These differences may partly contribute to the observed interactions, although baseline TyG-WC distribution did not fully explain the subgroup differences. The weaker association among older participants may partly reflect increased upper-airway collapsibility with ageing,^{19 20} which reduces the relative contribution of metabolic factors to OSA.

Several biological mechanisms may contribute to the association between TyG-WC and OSA. OSA-related sympathetic activation and oxidative stress may promote lipolysis and increase circulating free fatty acids.^{21 22} In a mouse model, chronic intermittent hypoxia impaired the clearance of triglyceride-rich lipoproteins and reduced adipose lipoprotein lipase activity, which may contribute to altered triglyceride metabolism.²³ Intermittent hypoxia and chronic inflammation, including increased levels of tumour necrosis factor (TNF)-α and interleukin (IL)-6, may impair insulin signalling and glucose uptake.^24–26 Altered leptin signalling may further affect ventilatory control and respiratory stability in OSA.²⁷ At the same time, abdominal obesity can contribute to upper-airway narrowing and increased airway collapsibility, while also being associated with metabolic dysfunction.²⁸ Together, these pathways may link IR, dysregulated lipid metabolism and central adiposity to OSA, providing a potential biological basis for the observed association between TyG-WC and OSA.

TyG-WC integrates TG, FBG and WC, providing a composite measure of IR and central adiposity.²⁹ TyG is widely used as a surrogate marker of IR.³⁰ Abdominal adiposity may play an important role in OSA beyond overall obesity,³¹ and IR has also been associated with OSA in paediatric populations.³² Previous studies have reported positive associations between TyG-related indices and OSA, including TyG-WC.^{10 11 13 18} Central adiposity may narrow the upper airway and increase airway collapsibility, whereas IR is associated with metabolic and inflammatory disturbances that may further affect OSA-related physiological processes. This may partly explain the contribution of TyG-WC to OSA identification in the ML models. In the SHAP analysis (figure 5), TyG-WC was the most influential predictor in all six models. This finding suggests that TyG-WC contributed to model prediction, although SHAP analysis does not by itself establish a specific biological mechanism.

From a clinical perspective, TyG-WC may serve as a simple and low-cost auxiliary marker for identifying individuals at higher risk of OSA. Models incorporating TyG-WC and clinical variables showed good predictive performance, and the DCA demonstrated positive net clinical benefit across a wide range of threshold probabilities (figure 4). These findings suggest that TyG-WC may help identify and prioritise individuals who may benefit from further OSA evaluation, particularly in settings where access to PSG is limited. Potential management strategies for individuals at higher risk include structured exercise programmes to reduce sympathetic activity,³³ targeted weight reduction to alleviate upper-airway fat deposition and improve airway stability,³⁴ and technology-assisted strategies such as digital tools to support continuous positive airway pressure adherence.³⁵ However, whether TyG-WC can be used to guide the selection or timing of these interventions remains unclear.

Unanswered questions and future research

Several questions remain to be addressed before TyG-WC can be incorporated into routine OSA risk assessment. First, the observational nature of the present study does not allow causal relationships between TyG-WC and OSA to be established. Future experimental and prospective studies are needed to clarify the biological pathways linking IR, central adiposity and OSA, including the dynamic changes in metabolic and inflammatory markers such as TNF-α and IL-6 under intermittent hypoxia. Second, the different approaches to OSA assessment between the hospital and NHANES samples highlight the need for large, multicentre prospective studies using PSG-confirmed OSA to further validate the association and assess the generalisability of TyG-WC across diverse populations. Such studies could also determine clinically useful cut-off values and evaluate whether TyG-WC provides incremental predictive value beyond established clinical predictors. Third, although TyG-WC is a low-cost measure, formal cost-effectiveness analyses are needed to determine whether TyG-WC-based screening provides sufficient clinical benefit to support its implementation as a standardised screening approach, particularly in settings where access to PSG is limited. Given the reported association of TyG-WC with other metabolic conditions such as metabolic dysfunction-associated fatty liver disease,³⁶ future studies may also examine whether TyG-WC can provide complementary information when OSA is evaluated alongside related metabolic conditions.

Conclusions

TyG-WC was associated with OSA, and ML models incorporating TyG-WC and clinical variables achieved good predictive performance, supporting the potential of TyG-WC as an accessible marker for future OSA risk assessment.

Data availability statement

Data are available in a public, open access repository. Data are available upon reasonable request.