Insulin resistance indices for predicting circadian syndrome: estimated glucose disposal rate as a protective indicator through non-linear dose-response in Chinese adults

CHARLS is a comprehensive longitudinal survey initiated in 2011, designed to examine the socioeconomic situation and health dynamics of individuals aged 45 and older in China. This study employed a multi-stage probability sampling method proportional to size to select 17,708 participants from 10,257 families throughout 150 counties in 28 provinces nationwide. Our work utilized data from CHARLS collected in 2011 and 2015 to conduct both cross-sectional and longitudinal analyses examining the relationship between IR surrogate indices and CircS risk. The CHARLS received approval from the Biomedical Ethics Review Committee at Peking University (IRB00001052-11015), and all participants provided informed consent.

The cross-sectional study's inclusion criteria were: (1) age ≥45 years; (2) complete data for CircS diagnosis. The exclusion criteria comprised: (1) participants lacking data on any of the six IR surrogate indices (eGDR, CVAI, TyG, TyG-BMI, METS-IR, and AIP); (2) participants without marital status information; (3) participants missing smoking status; (4) participants lacking drinking status; (5) participants without diabetes data; (6) participants missing low density lipoprotein cholesterol (LDL-C) concentration data; (7) participants lacking creatinine data.

The longitudinal study's inclusion criteria aligned with those of the cross-sectional study, while the exclusion criteria encompassed two additional items: (1) participants diagnosed with CircS at baseline and (2) participants with incomplete follow-up data. The participant selection flowchart is shown in. Figure 1

Six indices were selected as proxies of IR, including: eGDR, CVAI, TyG, TyG-BMI, METS-IR, and AIP. The specific calculation formula of each index was:

The concept of CircS is derived from MetS, and its diagnosis includes the following seven components: short sleep duration (< 6 h daily), central obesity [male: waist circumference (WC) ≥85 cm; female: WC ≥80 cm], elevated triglyceride (TG) levels [≥150 mg/dl (1.7 mmol/L)], reduced high-density lipoprotein cholesterol (HDL-C) [male: < 40 mg/dl (1.0 mmol/L); female: < 50 mg/dl (1.3 mmol/L)], hypertension (systolic blood pressure ≥130 mmHg and/or diastolic blood pressure ≥85 mmHg), hyperglycemia (≥100 mg/dl), and depression. The ten-item Center for Epidemiologic Studies Depression Scale-10 (CESD-10) was employed to evaluate depressive symptoms. The overall score ranges from 0 to 30 points, with elevated scores signifying more severe depressive symptoms. Individuals fulfilling four or more components are classified as possessing CircS; otherwise, they are deemed to lack CircS (). ⁷

Covariates in this study included: (1) demographic information: age (< 60 and ≥60); gender (female and male); education levels (less than junior high school education, high school and vocational training, higher education), and marital status (unmarried or separated, marriage or cohabitation); (2) health status and functioning: smoking (yes and no), drinking (yes and no), sleep duration, BMI, WC, diabetes, and hypertension; (3) laboratory test data: total cholesterol (TC), HDL-C, LDL-C, creatinine, uric acid, blood urea nitrogen (BUN), and C-reactive protein (CRP).

To assess the robustness of associations independent of component overlap between IR indices and CircS diagnostic criteria, sensitivity analyses were conducted with modified CircS definitions. For each IR index, CircS components sharing constituent variables with that index were excluded, and modified CircS was defined as meeting ≥3 of the remaining non-overlapping criteria (). Primary analyses were then repeated using these modified definitions. Supplementary Table S2

Descriptive analysis utilized mean ± standard deviation (SD) for continuous variables and percentages (%) for categorical variables. The Wilcoxon test (continuous variables) and the Chi-square test (categorical variables) were applied to examine the differences between the groups. In the cross-sectional study, the association between IR surrogate indices and CircS risk was evaluated by the logistic regression models. Three models were developed: Model 1 was unadjusted; Model 2 adjusted for age, gender, education levels, marital status, smoking, drinking, BMI, and WC; and Model 3 adjusted for Model 2 plus sleep duration, diabetes, hypertension, TC, HDL-C, LDL-C, creatinine, uric acid, BUN, and CRP. Restricted cubic splines (RCS) analysis was conducted to determine linear or non-linear relationships between IR surrogate indices and CircS risk. Additionally, subgroup analysis was carried out to investigate the consistency of associations across subgroups (age and gender). Receiver operating characteristic (ROC) curves were generated to evaluate the predictive effectiveness of IR surrogate indices for CircS risk. The Cox regression models in the longitudinal investigation assessed the relationship between IR surrogate measures and CircS risk, with remaining procedures aligned with the cross-sectional study.

All analyses were performed using R software (version 4.4.3). A-value below 0.05 was deemed statistically significant. P

After exclusions, 8,392 participants were enrolled in our study, including 3,101 participants diagnosed with CircS and 5,291 participants without CircS (). The CircS population had a higher proportion of females compared to males (69.0 vs. 31.0%), while the non-CircS group displayed an inverse tendency (44.0 vs. 56.0%). Both groups were predominantly characterized by low educational attainment (less than junior high school: CircS 91.9%, non-CircS 89.8%) and married/cohabiting status (CircS 81.5%, non-CircS 85.0%). Smoking and drinking rates were lower in the CircS group compared to the non-CircS group. Compared to participants without CircS, CircS patients exhibited higher CVAI, TyG, TyG-BMI, METS-IR, AIP, BMI, WC, TC, LDL-C, uric acid, and CRP levels, with a higher proportion of hypertension and diabetes, while eGDR, sleep duration, HDL-C, creatinine, and BUN were lower. All baseline characteristics showed significant differences between CircS and non-CircS groups (< 0.05). Table 1 P

,present the correlations between IR surrogate indicators and CircS risk. The cross-sectional study indicated that after full adjustment (Model 3), eGDR was a protective factor for CircS occurrence (OR = 0.381, 95% CI: 0.277–0.524), while AIP, METS-IR, TyG-BMI, and TyG were risk factors for CircS occurrence (AIP: OR = 4.089, 95% CI: 3.399–4.918; METS-IR: OR = 61.517, 95% CI: 40.52–93.395; TyG-BMI: OR = 25.057, 95% CI: 18.007–34.865; TyG: OR = 3.515, 95% CI: 3.089–4.000;). However, the association between CVAI and CircS was not significant (= 0.548). Tables 2 3 Table 2 P

The longitudinal study demonstrated that following adjustment (Model 3), eGDR was negatively associated with CircS risk (HR = 0.602, 95% CI: 0.523–0.693), while CVAI, TyG-BMI, METS-IR, and AIP showed opposite trends (CVAI: HR = 4.480, 95% CI: 2.744–7.314; TyG-BMI: HR = 1.189, 95% CI: 1.121–1.260; METS-IR: HR = 1.181, 95% CI: 1.112–1.254; AIP: HR = 1.232, 95% CI: 1.027–1.478;). However, the correlation between TyG and CircS was not marked (= 0.134). Table 3 P

Therefore, except for CVAI (non-significant in cross-sectional analysis) and TyG (non-significant in longitudinal analysis), the remaining indices achieved significance in both analyses.

These findings were further explored through conservative sensitivity analyses using modified CircS definitions (excluding overlapping components;,). In unadjusted models, all six indices maintained significant associations with modified CircS (cross-sectional: 10to 10; longitudinal: 10to 10). After full adjustment, TyG, TyG-BMI, and AIP retained cross-sectional significance (≤ 0.002), while eGDR (HR = 0.578,= 0.004) retained longitudinal significance. Supplementary Tables S3 S4 P P P P −15 −236 −5 −29

Restricted cubic spline (RCS) analysis was conducted to examine linear or non-linear relationships between IR surrogate indices and CircS risk. All IR surrogate indices showed significant non-linear associations with CircS risk in the cross-sectional analysis (< 0.001,< 0.001;–). P P −overall −non − linear Figures 2A F

In the longitudinal analysis, only eGDR, TyG-BMI, and METS-IR exhibited significant non-linear associations with CircS risk (< 0.001,< 0.001), while AIP showed a linear relationship (–). P P −overall −non − linear Figures 3A F

We conducted subgroup and interaction analyses based on age and gender to examine the relationship between IR surrogate indices and CircS risk across various demographic parameters. Results demonstrated that in the cross-sectional study, an increase in eGDR was consistently associated with a decrease in the prevalence of CircS, and the remaining indices showed opposite trends (). Interaction analysis revealed that age and gender moderated the relationship between CVAI and CircS risk (interaction: age:= 0.002; gender:< 0.001), and gender moderated the relationships between TyG and CircS risk (interaction:= 0.001), TyG-BMI and CircS risk (interaction:= 0.001), and AIP and CircS risk (interaction:= 0.016). Figure 4A P P P P P

In the longitudinal study, all IR surrogate indices exhibited consistent correlations with incident CircS (). Notable interactions were detected: age and gender moderated the association between TyG-BMI and CircS (age interaction:= 0.006; gender interaction:= 0.016), and age moderated the associations between eGDR and CircS (= 0.024) as well as between METS-IR and CircS (< 0.001). Figure 4B P P P P

ROC curves were generated to evaluate the discriminative ability of IR surrogate indicators for CircS. The results indicated that eGDR achieved an AUC of 0.893, while other indices all exceeded 0.9 (detailed data in;). In both sexes, all IR surrogate indices exhibited strong discriminative ability for CircS (detailed data in;,). Furthermore, except for METS-IR, the other indices demonstrated superior discriminative performance in the female group relative to the male group (,). When further stratified by age (< 60 vs ≥60 years), both sexes demonstrated better discrimination in the younger age group (–). Supplementary Table S1 Figure 5A Supplementary Table S1 Figures 5B C Figures 5B C Figures 5D G

This study systematically evaluated six IR surrogate indices for identifying prevalent and incident CircS in 8,392 Chinese adults aged ≥45 years. While all indices demonstrated comparable discriminative ability (AUC: 0.893–0.907), their performance diverged markedly between cross-sectional and longitudinal analyses, with single-component indices showing particularly limited longitudinal predictive value. Our findings complement previous studies that established associations between IR indices and traditional cardiometabolic outcomes (–). Whereas, prior work focused on IR surrogates' utility for predicting complications of diabetes or cardiovascular events, we demonstrate their specific application to CircS—a metabolic-circadian phenotype that uniquely integrates sleep and mood disturbances with metabolic dysfunction. This application to metabolic-circadian identification and prediction offers clinicians practical tools for screening patients at risk for a syndrome particularly relevant to modern lifestyle-related health challenges. ^{16 18}

Among the evaluated indices, eGDR emerged as particularly noteworthy, demonstrating both excellent discriminative capacity and unique protective characteristics (OR = 0.381, HR = 0.602). By integrating three critical dimensions—glycemic control (HbA1c), vascular health (hypertension status), and central adiposity (waist circumference)—eGDR captures a comprehensive insulin sensitivity phenotype that extends beyond simple glucose metabolism. Our findings align with recent work (), which demonstrated eGDR's inverse association with MetS prevalence and mortality. The robust inverse association between eGDR and CircS suggests a potential mechanistic link: preserved insulin sensitivity may protect against circadian disruption through maintenance of peripheral clock gene expression in metabolically active tissues. This hypothesis gains support from clinical observations in type 1 diabetes mellitus, where eGDR selectively identifies pathological sleep disorders requiring intervention while remaining stable in individuals with physiological sleep variations (). This selective responsiveness—distinguishing pathological from normal sleep variability—represents a critical advantage for CircS screening, as it minimizes false positives that might arise from benign sleep pattern variations common in aging populations. ^{19 20}

In contrast to eGDR's protective effects, TyG-BMI and METS-IR showed strong associations with prevalent CircS (OR = 25.057 and 61.517, respectively) and maintained significant, though more modest, associations with incident CircS (HR = 1.189 and 1.181, respectively). The substantial attenuation from cross-sectional ORs to longitudinal HRs merits explanation. Component overlap between IR indices and CircS criteria—while partly reflecting shared pathophysiology—may contribute to the larger cross-sectional effect sizes where exposure and outcome are measured simultaneously. The longitudinal design, with temporal separation between baseline exposure and incident outcome, reduces this contribution. Additionally, TyG-BMI remained significant in both analyses, whereas TyG lost longitudinal significance (HR=1.097,= 0.134), underscoring the added value of incorporating adiposity measures. This finding aligns with recent evidence, which demonstrated that composite IR indices (TyG-WHtR, TyG-WC) maintained predictive value while the standalone TyG index failed to predict mortality (). Together, these observations reveal fundamental principle: incorporating adiposity measures transforms glycemic-lipid indices into robust predictors by capturing the full spectrum of metabolic dysfunction—including adipose tissue distribution and ectopic lipid accumulation that drive progressive IR (). We also noted that, CVAI exhibited marked reduction after full adjustment in cross-sectional analysis, while the other five indices maintained consistent effect directions and significance in fully adjusted models. CVAI's instability likely relates to its having the highest number of constituent variables and the greatest overlap with adjusted covariates. P ^{21 22}

To rigorously assess the independence of these associations from component overlap, we conducted sensitivity analyses using modified CircS definitions that excluded overlapping components—a deliberately conservative approach. In unadjusted models, all indices maintained highly significant associations (< 10), indicating predictive capacity beyond shared variance. In fully adjusted models, several indices retained significance: TyG, TyG-BMI, and AIP in cross-sectional analysis (< 0.01), and eGDR (HR = 0.578,= 0.004) in longitudinal analysis. The attenuation of certain associations likely reflects over-adjustment: overlapping components were first excluded from the outcome definition, then related metabolic parameters were additionally controlled as covariates—a dual removal that inevitably diminishes residual signal. These sensitivity findings reinforce our primary analyses—particularly for eGDR, which demonstrated robust associations even under this conservative framework. eGDR's sustained longitudinal significance strengthens confidence in its reliability for long-term risk stratification. P P P −5

A striking finding was the three-fold higher diabetes prevalence among CircS patients compared to non-CircS individuals (11.2 vs. 3.6%,< 0.001), confirming a strong metabolic-circadian interconnection. This observation carries important clinical implications. First, patients with diabetes warrant prioritized CircS screening given their elevated risk. Second, the substantial overlap suggests shared pathophysiological pathways involving IR, explaining why IR indices effectively predict both conditions. Third, the high comorbidity burden indicates that isolated management of either condition may yield suboptimal outcomes. These findings align with emerging evidence of bidirectional causality—diabetes may precipitate CircS through metabolic disruption, while CircS can exacerbate glycemic dysregulation (,). For diabetes clinics, implementing IR-based CircS screening could involve: (1) automated calculation of eGDR at each visit using routine HbA1c and anthropometric data; (2) flagging patients whose IR indices deviate significantly from population norms for comprehensive CircS evaluation; (3) intensifying metabolic control in those with confirmed CircS, recognizing the need for integrated metabolic-circadian management beyond standard diabetes care. This approach transforms CircS from an underrecognized comorbidity to an actionable clinical target within existing diabetes care pathways. P ^{23 24}

Our subgroup analyses revealed age-dependent effect modifications with important implications for clinical implementation. The protective effect of eGDR was more pronounced in adults < 60 years (HR = 0.63) compared to those ≥60 years (HR = 0.74;= 0.024), identifying middle age as a critical intervention window. This age-related attenuation likely reflects the natural decline in metabolic flexibility with aging—older adults have reduced ability to switch between burning fats and carbohydrates for energy due to decreased mitochondrial function (). Conversely, risk indices (TyG-BMI, METS-IR) demonstrated stronger associations with CircS in older adults in longitudinal analysis (TyG-BMI:= 0.006; METS-IR:< 0.001). Aging disrupts circadian rhythms through reduced NAD+/SIRT1 signaling, impairing the clock's metabolic responsiveness (). This mechanistic insight may explain why IR indices show stronger associations with CircS in older adults—their compromised circadian-metabolic interface cannot adequately buffer metabolic disruptions. These age-specific modifications suggest tailored intervention strategies: middle-aged adults with preserved metabolic flexibility may benefit from aggressive IR management to prevent CircS development, while older adults require integrated approaches addressing both metabolic dysfunction and circadian desynchronization. P P P −interaction −interaction −interaction ^{25 26}

The female predominance in CircS (69.0%) reflects complex endocrine-metabolic interactions beyond ascertainment bias. In our cohort (mean age 59.75 years), postmenopausal estrogen decline disrupts metabolic-circadian coupling. Estrogen normally enhances insulin sensitivity via ERα-PI3K-Akt-Foxo1 signaling, suppressing hepatic gluconeogenesis while promoting peripheral glucose uptake (,). It also maintains pancreatic β-cell function by promoting insulin secretion and preventing apoptosis (). Its postmenopausal loss therefore promotes dual metabolic insults: peripheral IR and β-cell dysfunction. Importantly, IR indices showed better discriminative ability for CircS in women than in men, with optimal performance in those aged < 60 years—suggesting that screening during the perimenopausal transition, before metabolic deterioration becomes fully established, may maximize detection efficiency. These findings suggest that women in midlife and beyond as represent a priority population for IR-based screening. ^{27 28 29}

Longitudinal RCS analyses revealed distinct dose-response patterns among IR indices, with eGDR and AIP demonstrating a meaningful contrast. Specifically, eGDR showed a significant non-linear protective association with CircS risk: HR remained relatively stable at lower eGDR values, but declined steeply as eGDR increased—suggesting that greater insulin sensitivity is associated with more pronounced risk reduction. In contrast, AIP exhibited a linear positive association, with CircS risk increasing proportionally across its entire range without apparent thresholds. This difference in patterns may reflect distinct metabolic characteristics: eGDR integrates multidimensional information including glycemic control, blood pressure, and central adiposity, whereas AIP primarily reflects lipid metabolism status.

This study's strengths include the use of a nationally representative cohort with rigorous sampling methods, ensuring generalizability to Chinese middle-aged and elderly adults, while the integration of cross-sectional and longitudinal designs strengthens causal inference beyond what either approach could achieve alone. The comprehensive evaluation of six IR indices, compared to the 2–3 typically examined in prior studies, enables robust comparison of predictive efficacy, and the application of restricted cubic splines alongside subgroup interaction analyses provides nuanced insights into non-linear relationships and population heterogeneity. Despite these methodological advantages, several limitations warrant consideration. First, component overlap between IR indices and CircS diagnostic criteria represents an inherent methodological challenge. The high cross-sectional ORs may reflect both biological associations and shared components. Sensitivity analyses using modified CircS definitions supported the robustness of eGDR. However, these modified definitions deviate from the original CircS framework, and the primary analyses remain the core basis for this study. Second, the absence of fasting insulin data in CHARLS precluded calculation of HOMA-IR and direct comparison with this established IR measure. This limitation reflects the reality of community-based research—fasting insulin is not routinely measured in primary care settings. Third, the focus on adults ≥45 years restricts generalizability to younger populations—although CircS is increasingly prevalent among younger adults, research in this group remains limited.