The obesity paradox effect of triglyceride-glucose-body mass index (TyG-BMI) on all-cause mortality in patients with malignant tumors: a machine learning-driven cohort study

Introduction

Malignant tumors represent a major global public health challenge and impose a substantial burden on healthcare systems worldwide (1). As of 2022, cancer accounted for approximately 16.8% of all deaths and 22.8% of deaths due to non-communicable diseases globally (2,3). In 177 countries and regions, cancer ranks among the top three causes of death in individuals aged 30–69 years, significantly reducing life expectancy, straining medical resources, and exerting a marked impact on the global economy (4,5). According to recent epidemiological projections, the annual number of new cancer cases is expected to rise to 35 million by 2050, underscoring cancer as a central challenge to global public health security (6,7). Despite advances in early detection and therapeutic strategies, cancer prognosis remains unsatisfactory. Thus, identifying modifiable risk factors and reliable prognostic markers is essential for improving outcomes and reducing the disease burden.

Metabolic dysfunction, characterized by abnormalities in glucose and lipid metabolism, is a key driver in the pathogenesis of many chronic diseases. The triglyceride-glucose (TyG) index, a simple and validated surrogate of insulin resistance, has been widely studied (8). Accumulating evidence suggests that dysregulation of glucose and lipid metabolism is biologically linked to cancer outcomes (9,10). Elevated TyG index, reflecting insulin resistance, may promote inflammation and alter the vascular microenvironment, thereby stimulating cell proliferation and inhibiting apoptosis (11,12). Furthermore, the TyG index may accelerate cancer progression and reduce treatment sensitivity through mechanisms involving oxidative stress, hormonal dysregulation, and aberrant lipid signaling pathways (13).

Increasing evidence has demonstrated that higher TyG index levels are associated with greater risks of cardiovascular disease (CVD), type 2 diabetes, and all-cause mortality (14-16). Beyond the conventional TyG index, several modified indices have recently been proposed, such as TyG-waist circumference (TyG-WC), TyG-waist-to-height ratio (TyG-WHtR), and TyG-body mass index (TyG-BMI) (17-20). By integrating anthropometric and metabolic information, these composite indices may provide a more comprehensive assessment of metabolic risk (21). Although preliminary studies have examined the association between the TyG index and cancer incidence or mortality (22), the prognostic value of TyG-WC, TyG-WHtR, and TyG-BMI for cancer-specific, all-cause, and cardiovascular mortality in cancer patients or high-risk populations remains largely unexplored. In this context, we hypothesize that TyG-BMI may exhibit a nonlinear relationship with mortality, while other TyG-derived indices might demonstrate distinct cardiovascular risk implications. Accordingly, this study first employs weighted Cox regression to characterize etiological associations and identify risk thresholds. Subsequently, machine learning (ML) algorithms are utilized to systematically evaluate and compare the prognostic accuracy of these indices. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-aw-2385/rc).

Methods

Data source

Data for this study were obtained from the National Health and Nutrition Examination Survey (NHANES) spanning 1999–2018, available at: https://www.cdc.gov/nchs/nhanes/index.html. Conducted by the National Center for Health Statistics (NCHS) under the Centers for Disease Control and Prevention (CDC), NHANES is a continuous cross-sectional survey that employs a multistage probability sampling design. It systematically collects nationally representative data on demographics, clinical examinations, laboratory measures, and health outcomes in the U.S. population, providing high-quality evidence to support epidemiological research and public health policy-making. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Participant selection

This study adopted a prospective cohort design to explore the associations of TyG-related indices with all-cause, cancer-specific, and CVD mortality in cancer patients. A total of 101,316 participants were initially identified from the NHANES database (1999–2018). The participant selection process followed sequential steps: (I) age restriction: participants under 18 years old were excluded (n=42,112), leaving 59,204 adults. (II) Definition of malignancy: among adults, cancer status was determined based on self-reported medical history (i.e., answering “yes” to the question “Ever told you had cancer or malignancy?”). Individuals without a self-reported history of malignancy were excluded (n=54,038), resulting in 5,166 cancer survivors. (III) Data availability: participants with missing measurements for fasting plasma glucose (FPG) or triglycerides (TG) were further excluded (n=3,015). Ultimately, 2,151 participants were included in the final analysis (Figure 1).

Figure 1 Flow diagram of study population selection. FPG, fasting plasma glucose; NHANES, National Health and Nutrition Examination Survey; TG, triglyceride.

Assessment of malignancy

Cancer status was determined based on self-reported medical history from the NHANES questionnaire item MCQ220 (“Ever told you had cancer or malignancy?”). Participants who responded “yes” were considered to have a history of malignancy and were included in the analysis cohort. Consistent with the nature of NHANES data, this population is referred to as “participants with a self-reported history of malignancy” or “cancer survivors”, a definition that does not differentiate between active disease and the survivorship phase.

Measurement of TyG-related indices

The TyG index was calculated from FPG and TG levels as previously described (23):

• TyG = ln [TG (mg/dL) × FPG (mg/dL)/2];
• TyG-WC = TyG × waist circumference (cm);
• TyG-WHtR = TyG × waist circumference (cm)/height (cm);
• TyG-BMI = TyG × body mass index (kg/m²).

Each index was categorized into quartiles (Q1–Q4) based on the distribution of the study population, with Q1 serving as the reference.

Covariates

Potential confounders were included from three domains: (I) demographics: age (categorized as ≤68 and >68 years, with 68 representing the median age of mortality events in our cohort to ensure adequate statistical power), sex, race/ethnicity (Mexican American, other Hispanic, non-Hispanic White, non-Hispanic Black, other), marital status (married/cohabiting, separated/divorced/widowed, never married), education (<high school, high school, ≥college), and poverty-income ratio (PIR: <1.3, 1.3–3.5, >3.5). (II) Lifestyle factors: to minimize residual confounding and address the potential influence of the “obesity paradox,” smoking and alcohol consumption were finely stratified. Smoking status was categorized into three distinct groups: never (<100 cigarettes in lifetime), former (≥100 cigarettes but quit), and current smokers. Alcohol consumption was quantified by drinks per week and categorized into four groups: never, former, current light/moderate, and current heavy drinkers. (III) Medical history: diabetes and coronary heart disease (self-reported, diagnosed/undiagnosed).

Outcomes

The primary endpoints were all-cause mortality, cancer-related mortality, and CVD-related mortality. All events were ascertained from the publicly available, linked mortality files of the NHANES, probabilistically matched by the NCHS to the National Death Index (NDI) with follow-up censored on 31 December 2019. Cause-specific deaths were classified according to the International Classification of Diseases, Tenth Revision (ICD-10): cancer-related deaths were defined by malignant neoplasm codes (C00–C97), and CVD-related deaths by circulatory system disease codes (I00–I99), following the NCHS coding specifications.

Handling of missing values

In response to the complex sampling design of the NHANES database and the characteristics of this study’s core exposure indicators, data were processed with strict methodological considerations. Cases with missing values for core variables essential for calculating TyG-related indices (i.e., FPG, TG, and BMI) were excluded utilizing a complete case analysis approach. Importantly, missingness in these fasting laboratory data in NHANES is primarily considered “missing by design”, as these specific metabolic tests are only administered to a randomly selected morning fasting subsample (approximately one-third to one-half of the total cohort).

For missing data pertaining to non-core covariates (e.g., demographic characteristics, lifestyle factors, and clinical comorbidities), Multiple Imputation by Chained Equations (MICE) was employed to minimize potential bias and maximize statistical power. We generated five imputed datasets using the Random Forest algorithm within the MICE framework. Crucially, the imputation models included all variables used in the final analytical Cox models. This encompassed not only the covariates but also the complex survey design variables [strata, primary sampling units (PSU), and sample weights], as well as the outcome variables (survival time and mortality status, utilizing the Nelson-Aalen estimator for cumulative baseline hazard) to prevent attenuation of the estimated associations. Final regression results were obtained by pooling the estimates from the five imputed datasets according to Rubin’s rules.

Statistical analysis

All analyses strictly incorporated NHANES complex sampling design elements, including sample weights, strata, and PSU. Because the calculation of TyG-related indices requires FPG and TG—which were exclusively measured in a randomly selected subset of participants—we specifically utilized the morning fasting subsample weights (WTSAF2YR) rather than the standard Mobile Examination Center (MEC) weights, to ensure national representativeness. Furthermore, to account for the concatenation of 10 continuous survey cycles from 1999 to 2018, a new 20-year fasting weight was constructed by dividing the original 2-year fasting weights by 10, strictly following the analytical guidelines of the NCHS. Variance estimation was adjusted using the specific design variables for pseudo-stratum (SDMVSTRA) and pseudo-PSU (SDMVPSU) utilizing Taylor series linearization. Both the baseline distribution characteristics and the core association analyses of this study employed weighted statistical methods. Continuous variables are described according to their distribution: normally distributed variables are presented as mean ± standard deviation, while non-normally distributed variables are described using the median (25th, 75th percentiles). Categorical variables are presented as percentages [95% confidence intervals (CIs)], and between-group comparisons were conducted using non-parametric tests and the χ² test.

TyG-related indices (TyG, TyG-WC, TyG-WHtR, TyG-BMI) were categorized into quartiles (Q1–Q4) based on the 25th, 50th, and 75th percentiles of the overall baseline cohort distribution, with Q1 serving as the reference group. Cox proportional hazards (CoxPH) regression models were used to assess the associations between TyG-related indices and all-cause, cancer-related, and CVD mortality. Three hierarchical models were constructed: Model 1 was unadjusted; Model 2 was minimally adjusted for age and sex; Model 3 was further adjusted for race, marital status, education level, PIR, smoking status, alcohol consumption, diabetes, and coronary heart disease on the basis of Model 2. TyG-related indices were included as continuous variables in the models to test for linear trends, and their potential non-linear associations with outcome events were analyzed using restricted cubic spline (RCS) models (24).

Furthermore, to rigorously address potential residual confounding—particularly regarding the “obesity paradox”, where current or former heavy smokers/drinkers might exhibit lower body weight but higher mortality—comprehensive sensitivity analyses were conducted. Interaction and stratified subgroup analyses were performed across the finely categorized strata (including the updated 3-tier smoking and 4-tier alcohol categories) to verify the robustness of the associations between TyG-related indices and mortality outcomes. The Wald test was used to calculate the P for interaction. All tests were two-tailed, with a P value <0.05 considered statistically significant.

To model non-linear relationships and time-to-event outcomes, we employed a comprehensive suite of survival-specific ML algorithms, including random survival forests (RSF) (25,26), gradient boosting survival analysis (GBSA), and fast survival support vector machines (survivalSVM) (27). Additionally, we concurrently constructed standard CoxPH and penalized Cox regression (CoxNet, employing an elastic net penalty) models to serve as methodological benchmarks. Model discrimination was quantified using Harrell’s concordance index (C-index), evaluated on an independent 30% hold-out test set. To provide robust measures of uncertainty, we generated 95% CIs for all C-indices via 1,000 bootstrap resamples.

Crucially, due to algorithmic constraints in incorporating complex survey designs into standard tree-based and SVM survival architectures, the ML pipeline and its derived C-indices were unweighted. This distinguishes them from the primary survey-weighted Cox models utilized for etiological inference. To interpret the best-performing models, we calculated permutation feature importance to robustly quantify the global contribution of ea32ch variable to overall model performance. Subsequently, SHapley Additive exPlanations (SHAP) values were utilized to provide individual-level interpretability and elucidate the directionality of feature effects, complemented by visual representations (28).

For standard epidemiological survival analyses, the methodology was rigorously tailored to the specific nature of the outcomes. For all-cause mortality, standard Kaplan-Meier survival curves were plotted, and differences across quartiles were evaluated using the log-rank test. For cause-specific endpoints (cancer-related and CVD-related mortality), the presence of competing risks necessitated the application of the Aalen-Johansen cumulative incidence function (CIF). Accordingly, we performed cause-specific CoxPH regression to estimate the hazard ratios (HRs) for these endpoints, wherein individuals dying from competing causes were treated as right-censored. These classical statistical frameworks served as a robust baseline to corroborate the findings derived from the ML algorithms. All statistical and computational analyses were executed using R (version 4.3.3) and Python (version 3.11).

Results

Baseline characteristics of study participants

A total of 2,151 participants were included in this study, comprising 1,382 survivors and 769 deceased individuals. The median age of the cohort was 62.73 years, and deceased participants were significantly older than survivors. Women accounted for the majority of the study population (58.08%), whereas men represented a higher proportion among the deceased (50.85%). The racial composition was predominantly non-Hispanic White (86.63%), with no significant difference observed between survivors and deceased participants (P=0.320). Regarding education, 61.08% of participants had attained college or higher degrees. Marital status was mainly married or cohabiting, with a significantly higher proportion among survivors compared to the deceased. In terms of lifestyle factors, 74.33% reported alcohol consumption and 54.10% had a history of smoking, the latter being more common among deceased participants. With respect to comorbidities, 7.76% had coronary heart disease, which was significantly more prevalent in the deceased group (P<0.001). The prevalence of diabetes was 15.77%, also higher among deceased participants (P=0.002). The median follow-up duration was 89 months, during which 769 all-cause deaths were documented, including 235 cancer-specific deaths and 269 cardiovascular deaths (Table 1).

Associations between TyG-related indices and all-cause mortality

In the unadjusted model, each unit increment in TyG was associated with a 20.8% higher risk of all-cause mortality (HR =1.208; 95% CI, 1.054–1.385; P=0.0068). Participants in the highest quartile (Q4) exhibited a 49.9% elevated risk compared with the lowest quartile (Q1) (HR =1.499; 95% CI, 1.141–1.970; P=0.0037), and a significant dose–response trend was observed (P for trend =0.0037). Similarly, each unit increase in TyG-WC conferred a 0.1% higher mortality risk (HR =1.001; 95% CI, 1.000–1.001; P=0.0024), with Q4 versus Q1 yielding an HR of 1.442 (95% CI, 1.114–1.867; P=0.0055) and a significant linear trend (P for trend =0.0023). For TyG-WHtR, every unit increment corresponded to a 14.2% increase in mortality (HR =1.142; 95% CI, 1.052–1.238; P=0.0014). The Q4 versus Q1 comparison yielded an HR of 1.523 (95% CI, 1.162–1.996; P=0.0023), and the dose–response relationship remained statistically significant (P for trend =0.0022).

In our primary survey-weighted multivariable Cox analysis, neither TyG, TyG-WC, nor TyG-WHtR retained any significant independent association with all-cause mortality after exhaustive adjustment. In contrast, a distinctive pattern emerged for TyG-BMI: in the fully adjusted model (Model 3), this index exhibited a robust association with mortality among patients with malignancy. Using Q1 as the reference, participants in Q2 and Q3 experienced a statistically significant reduction in the risk of death (Q2: HR =0.753; 95% CI, 0.595–0.954; P=0.0188; Q3: HR =0.729; 95% CI, 0.553–0.961; P=0.0251). However, when TyG-BMI increased into Q4 this protective effect attenuated and became non-significant (HR =0.827; 95% CI, 0.618–1.108; P=0.204). Although the linear trend test across the continuous TyG-BMI scale did not reach statistical significance (P=0.167), the quartile analysis revealed a U-shaped risk distribution consistent with the previously described “obesity paradox.” This pattern suggests that while a moderate increase in TyG-BMI may confer a survival benefit—potentially reflecting a metabolic reserve—this protective association attenuates at the highest levels, without evidence of exerting a confirmed detrimental effect. A similar threshold pattern was observed for TyG-WC, with point estimates suggesting protection within the low-to-moderate abdominal obesity range (Q3 vs. Q1: HR =0.735; P=0.030), underscoring a non-monotonic rather than linear association (Figure 2).

Figure 2 Forest plot of the association between TyG-related indices and all-cause mortality. TyG, triglyceride-glucose; TyG-BMI, triglyceride-glucose index multiplied by body mass index; TyG-WC, triglyceride-glucose index multiplied by waist circumference; TyG-WHtR, triglyceride-glucose index multiplied by waist-to-height ratio.

Collectively, TyG, TyG-WC, and TyG-WHtR were significantly associated with all-cause mortality in unadjusted analyses, whereas TyG-BMI was not. After comprehensive covariate adjustment, only TyG-BMI and, to a lesser extent, TyG-WC demonstrated a threshold-dependent protective effect within low-to-moderate obesity strata, whereas the associations of TyG and TyG-WHtR were fully extinguished.

Non-linear dose-response relationships between TyG-related indices and all-cause mortality

RCS analyses revealed significant non-linear associations of both TyG-BMI and TyG-WC with all-cause mortality among patients with malignancy (P<0.05). RCS analyses revealed significant non-linear associations of both TyG-BMI and TyG-WC with all-cause mortality among patients with malignancy (P for non-linearity <0.05). Visually, the TyG-BMI curve exhibited a non-linear trend, reaching an approximate nadir around the value of 259.4. Below this descriptive inflection point, mortality risk declined steeply with increasing index values, whereas beyond it, the curve largely flattened, demonstrating a non-linear plateau effect. This pattern supports the concept that moderate TyG-BMI may confer protection via metabolic reserve, whereas extreme elevations re-expose patients to risks linked to metabolic dysregulation. The TyG-WC curve displayed an inverted-L threshold at 850.2. Mortality risk decreased progressively with rising TyG-WC below this point; however, once the threshold was exceeded, the risk plateaued, indicating that moderate abdominal adiposity-related insulin resistance within this range does not further worsen prognosis. By contrast, TyG-WHtR exhibited a monotonic downward trend, yet the non-linearity test did not achieve statistical significance (P non-linear =0.067). TyG itself showed neither linear nor non-linear associations with mortality (P overall =0.246). To further ascertain whether the results observed for TyG-BMI were driven exclusively by BMI, we performed a sensitivity analysis (Figures S1,S2). The analysis revealed that BMI alone also demonstrated a significant non-linear association with all-cause mortality (P non-linear <0.05). The trajectory of this curve closely mirrored that of TyG-BMI, with both patterns corroborating the ‘obesity paradox.’ This suggests that the apparent protective effect of TyG-BMI partially reflects the survival advantage provided by the metabolic reserves associated with a higher BMI in oncology patients. These overarching findings are illustrated in Figure 3.

Figure 3 Restricted cubic spline curves depicting the non-linear relationship between TyG-related indices and all-cause mortality. TyG, triglyceride-glucose; TyG-BMI, triglyceride-glucose index multiplied by body mass index; TyG-WC, triglyceride-glucose index multiplied by waist circumference; TyG-WHtR, triglyceride-glucose index multiplied by waist-to-height ratio.

Associations between TyG -related indices and cancer-related mortality

Following multivariable adjustment using a cause-specific CoxPH model (accounting for the competing risk of non-cancer mortality), we found that none of the TyG-related indices—including the standalone TyG index and its combinations with WC, WHtR, and BMI—showed a statistically significant association with cancer-specific mortality, and tests for trend were non-significant. Given the substantial heterogeneity in the participants’ self-reported cancer histories, we interpret these findings with caution. While this study did not identify a clear association between TyG indices and malignancy-specific mortality under the currently available covariates and the broad definition of cancer used, this does not imply that the TyG index completely lacks predictive value for tumor-related prognosis (Figure 4).

Figure 4 Forest plot of regression analyses between TyG-related indices and cancer- related mortality. TyG, triglyceride-glucose; TyG-BMI, triglyceride-glucose index multiplied by body mass index; TyG-WC, triglyceride-glucose index multiplied by waist circumference; TyG-WHtR, triglyceride-glucose index multiplied by waist-to-height ratio.

RCS analysis for cancer- related mortality

In the RCS analysis with cancer-specific mortality as the endpoint, none of the TyG indices exhibited significant nonlinear associations. The curves remained relatively flat around the reference line across the entire observed range, without apparent inflection points or U/J-shaped patterns, indicating the absence of nonlinear threshold effects for malignancy-related death and further supporting the aforementioned null findings (Figure 5). Furthermore, we performed a parallel analysis examining the association between BMI alone and cancer-specific mortality (Figures S3,S4). The results revealed that, following adjustment for confounding variables, BMI similarly demonstrated no independent association with cancer-specific mortality (P>0.05). These findings suggest that, within this cohort of cancer survivors, neither isolated BMI nor the composite TyG-BMI index possesses substantial predictive value for cancer-specific mortality risk. Accurate prognostic assessment for this specific outcome likely necessitates the incorporation of more granular clinical data, such as precise tumor staging or specific treatment modalities.

Figure 5 Restricted cubic spline curves for the non-linear relationship between TyG-related indices and cancer-specific mortality. TyG, triglyceride-glucose; TyG-BMI, triglyceride-glucose index multiplied by body mass index; TyG-WC, triglyceride-glucose index multiplied by waist circumference; TyG-WHtR, triglyceride-glucose index multiplied by waist-to-height ratio.

Associations of TyG indices with cardiovascular mortality

In our primary survey-weighted unadjusted models, TyG and its composite indices were significantly positively associated with cardiovascular mortality among cancer patients. For continuous TyG, each 1-unit increase corresponded to a 33% higher risk of cardiovascular death (HR =1.328, P=0.0059), and the Q4 vs. Q1 comparison yielded an HR of 1.66 (P=0.017), with a significant trend test (P=0.008). TyG-WC also demonstrated a dose–response relationship (continuous HR =1.001, P=0.013). The strongest initial effect was observed for TyG-WHtR: continuous HR =1.224 (P=0.001), with Q4 nearly doubling the risk (HR =1.984, P=0.002), and a significant trend (P=0.003).

“However, within this fully weighted, nationally representative framework, these associations attenuated markedly after sequential adjustment for demographic, lifestyle, and clinical covariates (Model 3), resulting in no statistically significant independent differences or trends remaining. TyG-BMI consistently showed no significant association throughout the hierarchical analyses. Correspondingly, when evaluated in isolation, BMI also demonstrated no significant independent association with cardiovascular mortality (Figures S5,S6).

These findings robustly suggest that the early positive signals of TyG-related indices for cardiovascular mortality are largely driven by confounding factors and do not independently predict outcomes when complex survey design and exhaustive covariates are accounted for. Risk stratification using these indices must therefore incorporate conventional cardiovascular risk factors to prevent spurious associations (Figure 6).

Figure 6 Forest plot of regression analyses between TyG-related indices and cardiovascular disease-specific mortality. TyG, triglyceride-glucose; TyG-BMI, triglyceride-glucose index multiplied by body mass index; TyG-WC, triglyceride-glucose index multiplied by waist circumference; TyG-WHtR, triglyceride-glucose index multiplied by waist-to-height ratio.

Nonlinear trends of TyG indices in cardiovascular mortality

RCS analysis for cardiovascular death revealed no significant linear or nonlinear trends for TyG or its composites with WC, WHtR, or BMI. All curves remained relatively flat, with no statistically meaningful inflection points (Figure 7).

Figure 7 Restricted cubic spline curves depicting the non-linear relationship between TyG-related indices and cardiovascular disease–specific mortality. TyG, triglyceride-glucose; TyG-BMI, triglyceride-glucose index multiplied by body mass index; TyG-WC, triglyceride-glucose index multiplied by waist circumference; TyG-WHtR, triglyceride-glucose index multiplied by waist-to-height ratio.

Survival and cumulative incidence analysis of TyG indices

To evaluate the prognostic value of TyG-related indices over time, we performed survival analyses tailored to the specific mortality outcomes. For all-cause mortality, standard Kaplan-Meier survival curves were utilized. The analysis indicated that among the indices, TyG-BMI quartiles demonstrated a significant association with all-cause mortality (P<0.05, log-rank test), with the quartile curves clearly separated, suggesting a robust predictive capacity for overall survival.

Importantly, to account for competing risks in cause-specific mortality (cardiovascular and cancer-specific deaths), we replaced the standard Kaplan-Meier estimates with Aalen-Johansen CIFs. The CIF curves visually depicted the cumulative incidence of events over the follow-up period. While certain indices (such as TyG-WHtR for cardiovascular death) exhibited a pronounced early rise in cumulative incidence among the high-value groups—indicating potential early risk—other combinations, including TyG and TyG-WC, showed limited prognostic differentiation for cancer-specific mortality. All survival and cumulative incidence curves were plotted across the quartiles of the four indices (Figure 8).

Figure 8 Kaplan-Meier survival curves and Aalen-Johansen cumulative incidence functions for mortality outcomes according to quartiles of TyG-related indices. Q2, Q3, and Q4 refer to the second, third, and fourth quartiles of the TyGrelated indices (TyG, TyGBMI, TyGWC, and TyG‑WHtR). CVD, cardiovascular disease; TyG, triglyceride-glucose; TyG-BMI, triglyceride-glucose index multiplied by body mass index; TyG-WC, triglyceride-glucose index multiplied by waist circumference; TyG-WHtR, triglyceride-glucose index multiplied by waist-to-height ratio.

Interaction analysis

To rigorously evaluate the prognostic stability of the TyG-related indices, we conducted systematic subgroup analyses and interaction tests for all-cause, cancer-specific, and cardiovascular mortality. These analyses incorporated ten stratification factors: sex, age, race/ethnicity, educational level, marital status, poverty-to-income ratio (PIR), alcohol consumption status, smoking status, and history of diabetes and coronary heart disease.

As shown in Figures S7-S18, the associations between the TyG-related indices and adverse outcomes remained largely consistent across subgroups. Notably, interaction tests for key lifestyle and clinical factors—including our detailed categorizations of smoking and alcohol consumption statuses—all yielded P≥0.05. This indicates that these traditional behavioral risk factors did not significantly modify the prognostic effects of the TyG indices. Although significant quantitative interactions (P for interaction <0.05) were observed for certain demographic characteristics (e.g., age and race), suggesting a more pronounced magnitude of association in individuals over 68 years of age and in specific racial groups, the overall direction of risk remained stable. Collectively, these findings demonstrate that the predictive performance of the TyG-related indices is highly robust and independent of individuals’ baseline lifestyle habits (Figures S7-S18).

ML results

Based on the C-index, the survival ML analysis identified the optimal TyG-related index and model combination for each specific outcome: TyG-BMI combined with the RSF model (C-index =0.840) for cardiovascular mortality, TyG combined with the GBSA model (C-index =0.771) for all-cause mortality, and TyG-BMI combined with the CoxPH benchmark model (C-index =0.704) for cancer-specific mortality. To strictly evaluate the predictive performance of these advanced algorithms, a standard unweighted CoxPH model was constructed solely as a methodological benchmark. Within this unweighted baseline setting, Cox regression yielded nominal associations. However, as established in our primary survey-weighted analysis, these specific associations were largely attenuated or altered when appropriately adjusting for the complex NHANES sampling design, underscoring the distinction between ML predictive benchmarking and epidemiological clinical inference. In contrast, TyG-related indices demonstrated limited predictive value for cancer-specific mortality (maximum C-index only 0.704, and no significant association was found in the benchmark Cox regression), a finding consistent across both ML and Cox regression comparative analyses (Figure 9).

Figure 9 Comparative analysis of predictive performance. (A) All-cause mortality, (B) cardiovascular mortality, and (C) cancer-specific mortality. The left panels display the C-index of the optimal machine learning models, and the right panels present the HR from the Cox regression benchmark. CI, confidence interval; CoxPH, Cox proportional hazards; GBSA, gradient boosting survival analysis; HR, hazard ratio; ML, machine learning; RSF, random survival forest; survivalSVM, survival support vector machin; TyG, triglyceride-glucose; TyG-BMI, triglyceride-glucose index multiplied by body mass index; TyG-WC, triglyceride-glucose index multiplied by waist circumference; TyG-WHtR, triglyceride-glucose index multiplied by waist-to-height ratio.

Prediction of all-cause mortality

Feature selection based on the permutation method highlighted age as a core predictor of all-cause mortality (permutation importance >0.05). SHAP visualization demonstrated that age consistently contributed positively to the predicted mortality risk, indicating that higher age values were associated with increased predicted risk within the fitted survival models. Kaplan-Meier analysis revealed a lower survival probability in the highest quartile (Q4) of TyG. In the unweighted benchmark Cox model utilized for ML comparison, TyG yielded an HR of 0.926 (95% CI, 0.816–1.051; P=0.234). Interestingly, higher TyG-BMI showed a minor but statistically significant inverse trend (HR =0.998; 95% CI, 0.997–0.999; P=0.005), whereas TyG-WC and TyG-WHtR showed no significant associations in this baseline setting. We reiterate that these unweighted estimates serve strictly to generate a baseline C-index; definitive clinical epidemiological conclusions rely exclusively on the primary survey-weighted models presented earlier. Among all predictive models, TyG combined with GBSA achieved the highest C-index (0.771), followed by TyG-BMI + CoxPH (0.753), TyG-WHtR + RSF (0.739), and TyG-WC + survivalSVM (0.725). Both ML and traditional Cox models possess clinical utility for all-cause mortality risk stratification (Table 2, Figures 10,11). It is noteworthy that the variable rankings derived from permutation importance and SHAP analysis were not entirely consistent. This is expected, as permutation importance quantifies each feature’s impact on overall model performance, while SHAP reflects a feature’s contribution to individual predictions. In the presence of correlated metabolic and anthropometric variables, SHAP may distribute predictive contributions across related features. (Figures S19-S22, Table S1-S4).

Figure 10 Interpretability analysis for all-cause mortality (TyG-WC + RSF). (A) The left panel shows the permutation feature importance, and (B) the right panel displays the SHAP summary plot. CHD, coronary heart disease; PIR, poverty income ratio; RSF, random survival forest; SHAP, SHapley Additive exPlanations; TyG-WC, triglyceride-glucose-waist circumference.

Figure 11 Model discrimination and Kaplan-Meier analysis for all-cause mortality. (A) The left panel shows the C-index comparison across survival models, and (B) the right panel presents the Kaplan-Meier survival curves based on TyG-WC quartiles. CI, confidence interval; CoxNet, LASSO-Cox regression; CoxPH, Cox proportional hazards; GBSA, gradient boosting survival analysis; RSF, random survival forest; survivalSVM, survival support vector machin; TyG-WC, triglyceride-glucose index multiplied by waist circumference.

Prediction of cardiovascular mortality

Feature selection identified age as a key predictor of CVD mortality (permutation importance >0.05). SHAP analysis indicated a positive association between age and cardiovascular mortality. Kaplan-Meier curves demonstrated that patients in the highest quartile (Q4) of TyG-BMI had a significantly lower survival probability compared to those in the lowest quartile (Q1). Evaluated strictly within the unweighted benchmark Cox model, TyG-related indices showed non-significant associations (e.g., TyG-BMI yielded HR =0.999; 95% CI: 0.996–1.001; P=0.194), which aligns with the findings in our primary survey-weighted analysis where specific associations were largely attenuated in the nationally representative framework. Cardiovascular mortality was the outcome predicted with the highest accuracy. The TyG-BMI + RSF model achieved the highest C-index (0.840), followed by the TyG-WHtR + CoxNet (0.817), TyG + CoxPH (0.815), and TyG-WC + survivalSVM (0.794) models. The discriminatory ability of the top-performing ML model (TyG-BMI + RSF) surpassed that of the traditional Cox models, highlighting the value of capturing non-linear relationships. The TyG-BMI + RSF model demonstrated robust generalization, supporting its potential as a reliable tool for cardiovascular risk stratification in cancer patients Table 3, Table S5-S8, Figures 12,13, Figures S23-S26. The analysis of TyGrelated indices and cancerrelated mortality is shown in Supplementary Figures S27–S30 and Supplementary Tables S9–S12.

Figure 12 Interpretability analysis for cardiovascular mortality (TyG-BMI + RSF). (A) The left panel shows the permutation feature importance, and (B) the right panel displays the SHAP summary plot. CHD, coronary heart disease; PIR, poverty income ratio; RSF, random survival forest; SHAP, SHapley Additive exPlanations; TyG-BMI, triglyceride-glucose-body mass index.

Figure 13 Model discrimination and Kaplan-Meier analysis for cardiovascular mortality. (A) The left panel shows the C-index comparison across survival models, and (B) the right panel presents the Kaplan-Meier survival curves based on TyG-BMI quartiles. CI, confidence interval; CoxNet, LASSO-Cox regression; CoxPH, Cox proportional hazards; GBSA, gradient boosting survival analysis; RSF, random survival forest; survivalSVM, survival support vector machin; TyG-BMI, triglyceride-glucose-body mass index.

Discussion

While recent studies have consistently reported positive associations between TyG-related indices and mortality in patients with diabetes, data concerning oncology populations remain sparse (29). Specifically, Liu et al. (30) analyzed 10,190 individuals with type 2 diabetes from the ACCORD trial and its follow-up cohorts, finding that those in the highest TyG-BMI quartile faced 74%, 165%, and 42% increased risks of cardiovascular mortality, congestive heart failure, and all-cause mortality, respectively, compared to those in the lowest quartile. These findings established TyG-BMI as a robust predictor of cardiovascular death in diabetic populations. Notably, malignancy and diabetes share a spectrum of overlapping metabolic comorbidities, most notably insulin resistance and chronic inflammation (31,32). Given that cancer patients often undergo further glucose and lipid metabolic dysregulation driven by tumor progression and various treatments—including chemotherapy, radiotherapy, or targeted therapies—it stands to reason that TyG-related indices may harbor significant prognostic utility in this population (33,34). However, because the metabolic landscape of cancer is distinct from that of diabetes, conclusions drawn from diabetic cohorts cannot be directly extrapolated to oncology patients (35,36). Consequently, the present study was designed to systematically evaluate the associations between TyG-related indices and mortality outcomes specifically in patients with malignant tumors.

This study enrolled 2,151 cancer patients with a median follow-up of 89 months, applying conventional survey-weighted CoxPH models alongside five survival ML algorithms to comprehensively evaluate the associations between the TyG index, its adiposity-related derivatives, and the risks of all-cause, cancer-specific, and cardiovascular mortality. In our preliminary unadjusted models, TyG, TyG-WC, and TyG-WHtR were positively associated with mortality. However, these associations were largely attenuated after adjusting for confounders such as age and sex, indicating that the initial correlations were primarily driven by demographic variables. Following multivariable adjustment within a nationally representative framework, only TyG-BMI exhibited a robust non-linear plateau relationship with all-cause mortality. Furthermore, ML analyses revealed differential predictive values: based on predictive discrimination (C-index), the TyG-BMI index paired with a RSF model was optimal for predicting cardiovascular mortality, whereas TyG combined with GBSA achieved the best performance for all-cause mortality.

Consistent with the findings of Zheng et al. (37), unadjusted TyG, TyG-WC, and TyG-WHtR indices were positively associated with all-cause mortality, exhibiting a clear dose-response relationship. However, upon adjustment for age, sex, lifestyle factors, and comorbidities, these associations were markedly attenuated, highlighting a substantial confounding influence. Notably, even after full multivariable adjustment within our survey-weighted design, TyG-BMI maintained a significant non-linear relationship with all-cause mortality: compared with the lowest quartile (Q1), the second and third quartiles (Q2 and Q3) demonstrated significantly reduced risks of death, whereas the highest quartile (Q4) showed no statistically significant difference from Q1.

While this trend mirrors the ‘obesity paradox’ reported in prior literature (38-42), we adopt a cautious interpretation. Specifically, the elevated risk observed at the lower end of the TyG-BMI spectrum may not signify a true ‘protective effect’ of moderate adiposity; instead, it likely stems from the inherent characteristics of the index construction and selection bias—such as occult cancer cachexia, inflammation, or end-of-life metabolic decline—that were not explicitly accounted for during the baseline assessment. Furthermore, as TyG-BMI inherently integrates anthropometric adiposity with insulin resistance, its predictive capacity may arise from the complex interplay between these factors rather than a singular protective metabolic phenotype. To this end, our independent analysis evaluating BMI in isolation revealed no statistically significant association with cardiovascular mortality. This further suggests that the prognostic value of TyG-BMI is likely derived from the integrative effect of metabolic dysregulation and body mass, underscoring the need to avoid over-interpreting such patterns as causal evidence of ‘metabolic protection’.

Our ML analyses revealed a distinct gradient in the predictive performance of TyG-related indices across different clinical outcomes: cardiovascular mortality exhibited the highest predictability, followed by all-cause mortality, while the predictive value for cancer-specific mortality remained relatively limited. Given the substantial heterogeneity inherent in self-reported cancer histories, tumor progression is more directly governed by complex factors such as treatment response—a notion consistent with Lee et al. (43). Similarly, Liu et al. (44) primarily validated TyG as a predictor in non-cancer cohorts, further underscoring the context-dependency of our findings. Crucially, our primary multivariable models explicitly confirmed that TyG-related indices demonstrate no independent association with cancer-specific mortality. This core finding underscores that the prognostic utility of TyG-derived indices in oncology patients is strictly confined to non-cancer competing outcomes and should not be misinterpreted as implying a direct relevance to tumor progression itself.

We acknowledge that the incremental gain in overall discrimination achieved by the ML models was modest compared with traditional penalized CoxNet or CoxPH regression. However, the value of these advanced algorithms transcends aggregate statistical performance; they fundamentally enhance clinical risk stratification through non-linear pattern recognition and individual-level interpretability. While traditional Cox models primarily reflect linear associations, tree-based ML architectures natively capture complex, multidimensional interactions. In particular, the non-linear relationship observed between TyG-BMI and all-cause mortality—with a visual risk nadir at approximately 259.4—reinforces the value of this index as a composite marker of insulin resistance and body mass, capable of capturing chronic inflammation and metabolic dysregulation (45,46). Furthermore, SHAP analysis not only confirmed the persistently dominant role of age in driving baseline risk but also illuminated significant inter-individual heterogeneity in the prognostic contributions of TyG indices. In a clinical context, this means that if SHAP visualizations attribute a patient’s elevated risk primarily to metabolic aberrations rather than chronological age alone, interventions can be aggressively tailored toward optimizing metabolic homeostasis, rather than dismissing the risk as an inevitable consequence of aging.

Strengths and limitations

To our knowledge, this study is the first to systematically evaluate the prognostic impact of TyG-related indices on all-cause mortality among patients with cancer, thereby bridging a critical gap left by prior research that predominantly focused on diabetic or cardiovascular cohorts. By integrating glycolipid metabolism with body fat distribution, these composite indices inherently reflect the complex metabolic and nutritional imbalances characteristic of malignancies. Methodologically, the application of RCS robustly elucidated the non-linear trajectory of TyG-BMI, while the integration of SHAP values and ML algorithms validated these predictive combinations using routinely available clinical data, thereby facilitating their potential translation into primary care settings.

Nonetheless, several limitations of this study must be acknowledged. First, all exposure variables were assessed at a single baseline time point. Given the median follow-up of 89 months in this oncology cohort, we were unable to capture the dynamic longitudinal shifts in metabolic parameters, which are profoundly influenced by disease progression and end-of-life decline (e.g., cancer cachexia). This inherent limitation heightens the susceptibility to reverse causation or survivor bias. Second, the ascertainment of cancer status relied on self-reported data, which lacks granular clinical details such as tumor histological subtypes, TNM staging, specific therapeutic modalities, and the exact time elapsed since diagnosis. Moreover, the potential for substantial residual confounding remains a critical limitation. Specifically, our current models could not be adjusted for unmeasured, highly dynamic clinical variables, including the degree of patient frailty, comprehensive nutritional status, and the specific efficacy of ongoing antineoplastic treatments. Consequently, the observed non-linear plateau relationship associated with BMI-based indices must be interpreted with extreme caution, as a lower TyG-BMI might merely serve as a surrogate marker for severe baseline frailty and treatment-refractory nutritional depletion.

Third, because the exact statistical uncertainty surrounding the identified mathematical nadir of 259.4 was not formally derived via resampling techniques, this threshold must be strictly interpreted as a descriptive approximation rather than a definitive clinical cutoff. Fourth, despite the protracted follow-up period, the limited number of cause-specific mortality events constrained our statistical power to perform extensive stratified analyses. Fifth, the exclusion of log-transformed C-reactive protein (lnCRP) due to missing data across several NHANES cycles may have introduced residual confounding related to systemic inflammation. Finally, our study population was restricted to U.S. residents, which may limit the generalizability of these findings to other diverse ethnic and demographic groups globally.

Future research should prioritize large-scale, multicenter prospective cohort studies that incorporate longitudinal monitoring of TyG index trajectories. Integrating robust stratification by tumor subtype and accounting for treatment-induced metabolic shifts will be essential to elucidate true causal relationships and minimize biases arising from population heterogeneity.

Conclusions

Survey-weighted Cox regression analyses based on a nationally representative cohort of cancer patients revealed that the initial linear associations between TyG-related indices and mortality were largely attenuated after adjusting for conventional demographic and clinical confounders. However, RCS analysis confirmed a robust non-linear plateau relationship between TyG-BMI and all-cause mortality. Furthermore, by integrating survival ML algorithms, we observed that prognostic discrimination was substantially improved through capturing complex interactions among variables. Specifically, TyG-BMI combined with the RSF algorithm demonstrated the highest predictive performance for cardiovascular mortality, whereas the TyG index combined with GBSA performed best in predicting all-cause mortality. Of note, neither conventional models nor ML frameworks revealed an independent association between these indices and cancer-specific mortality. In conclusion, TyG and its derived indices do not serve as causal drivers of tumor progression, but rather function as non-linear prognostic biomarkers for assessing systemic metabolic status and cardiovascular competing risk. By integrating traditional epidemiological approaches with ML, this study provides a novel analytical paradigm for individualized comprehensive risk stratification in oncology.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-aw-2385/rc

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-aw-2385/prf

Funding: This work was supported by Shanghai Health System Key Discipline Construction Project (No. 2024ZDXK0026); National Administration of Traditional Chinese Medicine, Comprehensive Office of Integrative Medicine “Flagship” Department Construction Project, 2024; Zhang Xiulan Charity Fund, Shandong Rural Revitalization Foundation; and Ren Nianrong Charity Fund, Shandong Rural Revitalization Foundation.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-aw-2385/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This research relies on information from the NHANES, which is a public database, thus eliminating the need for specific ethical approval.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Filho AM, Laversanne M, Ferlay J, et al. The GLOBOCAN 2022 cancer estimates: Data sources, methods, and a snapshot of the cancer burden worldwide. Int J Cancer 2025;156:1336-46. [Crossref] [PubMed]
2. Bray F, Laversanne M, Weiderpass E, et al. The ever-increasing importance of cancer as a leading cause of premature death worldwide. Cancer 2021;127:3029-30. [Crossref] [PubMed]
3. Siegel RL, Giaquinto AN, Jemal A. Cancer statistics, 2024. CA Cancer J Clin 2024;74:12-49. [Crossref] [PubMed]
4. The global burden of adolescent and young adult cancer in 2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Oncol 2022;23:27-52.
5. Werutsky G, Gössling G, Pellegrini RA, et al. Socioeconomic Impact of Cancer in Latin America and The Caribbean. Arch Med Res 2022;53:818-25. [Crossref] [PubMed]
6. Chen S, Cao Z, Prettner K, et al. Estimates and Projections of the Global Economic Cost of 29 Cancers in 204 Countries and Territories From 2020 to 2050. JAMA Oncol 2023;9:465-72. [Crossref] [PubMed]
7. Deo SVS, Sharma J, Kumar S. GLOBOCAN 2020 Report on Global Cancer Burden: Challenges and Opportunities for Surgical Oncologists. Ann Surg Oncol 2022;29:6497-500. [Crossref] [PubMed]
8. Nabipoorashrafi SA, Seyedi SA, Rabizadeh S, et al. The accuracy of triglyceride-glucose (TyG) index for the screening of metabolic syndrome in adults: A systematic review and meta-analysis. Nutr Metab Cardiovasc Dis 2022;32:2677-88. [Crossref] [PubMed]
9. Han M, Wang H, Yang S, et al. Triglyceride glucose index and Atherogenic index of plasma for predicting colorectal neoplasms in patients without cardiovascular diseases. Front Oncol 2022;12:1031259. [Crossref] [PubMed]
10. Pomares-Millan H, Saxby SM, Al-Mashadi Dahl S, et al. Dietary Glycemic Index, Glycemic Load, Sugar, and Fiber Intake in Association With Breast Cancer Risk: An Updated Meta-analysis. Nutr Rev 2025;83:1171-82. [Crossref] [PubMed]
11. Adams-Huet B, Jialal I. An Increasing Triglyceride-Glucose Index Is Associated with a Pro-Inflammatory and Pro-Oxidant Phenotype. J Clin Med 2024;13:3941. [Crossref] [PubMed]
12. Li J, Fan L, Nan Y, et al. Focus on Cell Apoptosis, Pyroptosis and Ferroptosis to Explore Strategic Breakthrough for GDM. J Inflamm Res 2025;18:10355-73. [Crossref] [PubMed]
13. Cui C, Liu L, Qi Y, et al. Joint association of TyG index and high sensitivity C-reactive protein with cardiovascular disease: a national cohort study. Cardiovasc Diabetol 2024;23:156. [Crossref] [PubMed]
14. Zhou L, Wang J, Zhou Z, et al. Long-term dual-trajectories of TyG and LAP and their association with cardiometabolic multimorbidity in midlife: the CARDIA study. Cardiovasc Diabetol 2025;24:198. [Crossref] [PubMed]
15. Oh R, Kim S, Park SH, et al. Elevated triglyceride-glucose index is a risk factor for cardiovascular events in adults with type 1 diabetes: a cohort study. Cardiovasc Diabetol 2025;24:150. [Crossref] [PubMed]
16. Zhao X, Song L, Li J, et al. Effect of Triglyceride-Glucose Indices and Circulating PCSK9-Associated Cardiovascular Risk in STEMI Patients with Primary Percutaneous Coronary Artery Disease: A Prospective Cohort Study. J Inflamm Res 2023;16:269-82. [Crossref] [PubMed]
17. Lyu L, Wang X, Xu J, et al. Association between triglyceride glucose-body mass index and long-term adverse outcomes of heart failure patients with coronary heart disease. Cardiovasc Diabetol 2024;23:162. [Crossref] [PubMed]
18. Zhong J, Liu D, Huang X, et al. Triglyceride glucose-waist circumference dynamics and cardiovascular risk: a national longitudinal study. Lipids Health Dis 2025;24:354. [Crossref] [PubMed]
19. Zhang L, Zhao W, Zheng P, et al. Association of triglyceride-glucose-related indices with all-cause and cause-specific mortality in individuals with prediabetesss. Cardiovasc Diabetol 2025;24:330. [Crossref] [PubMed]
20. Niu S, Jiang C, Miao X, et al. Association of triglyceride-glucose related indices with colorectal cancer risk among the US population: a cross-sectional study. BMC Cancer 2025;25:1186. [Crossref] [PubMed]
21. Deng Y, Wang Y, Yang J, et al. Global, Regional, and National Disease Burden and Prediction Analysis of Colorectal Cancer Attributable to Tobacco, Alcohol, and Obesity From 1990 to 2030. Front Oncol 2025;15:1524308. [Crossref] [PubMed]
22. Yi T, Lin Z, Mai Z, et al. The triglyceride-glucose index associated with reduced risk of liver metastasis in pancreatic cancer. Front Endocrinol (Lausanne) 2025;16:1592788. [Crossref] [PubMed]
23. Sun Y, Ji H, Sun W, et al. Triglyceride glucose (TyG) index: A promising biomarker for diagnosis and treatment of different diseases. Eur J Intern Med 2025;131:3-14. [Crossref] [PubMed]
24. Harrell, Frank E Jr. Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Analysis. 2nd ed., Springer Series in Statistics, Springer, 2015.
25. O'Donnell A, Cronin M, Moghaddam S, et al. A Systematic Review on Machine Learning Techniques for Survival Analysis in Cancer. Cancer Med 2025;14:e71375. [Crossref] [PubMed]
26. Ishwaran H, Kogalur UB. Consistency of Random Survival Forests. Stat Probab Lett 2010;80:1056-64. [Crossref] [PubMed]
27. Lynch CM, Abdollahi B, Fuqua JD, et al. Prediction of lung cancer patient survival via supervised machine learning classification techniques. Int J Med Inform 2017;108:1-8. [Crossref] [PubMed]
28. Alabdallah A, Pashami S, Rögnvaldsson T, et al. SurvSHAP: a proxy-based algorithm for explaining survival models with SHAP. In: 2022 IEEE 9th International Conference on Data Science and Advanced Analytics (DSAA). IEEE, 2022:1-10.
29. D'Elia L, Rendina D, Iacone R, et al. Triglyceride-Glucose Index and New-Onset Type 2 Diabetes Mellitus in Middle-Aged Men. Metabolites 2025;15:537. [Crossref] [PubMed]
30. Liu MJ, Xiang SM, Hu XQ. Triglyceride glucose-body mass index is associated with cardiovascular outcomes and overall mortality in type-2 diabetes mellitus patients. World J Diabetes 2025;16:108839. [Crossref] [PubMed]
31. Ennis CS, Seen M, Chen A, et al. Plasma exosomes from individuals with type 2 diabetes drive breast cancer aggression in patient-derived organoids. Commun Biol 2025;8:1276. [Crossref] [PubMed]
32. Hu JJ, Zhang QY, Yang ZC. The correlation between obesity and the occurrence and development of breast cancer. Eur J Med Res 2025;30:419. [Crossref] [PubMed]
33. Deepak K, Roy PK, Das A, et al. Glucose-6-phosphate dehydrogenase (G6PD) shields pancreatic cancer from autophagy-dependent ferroptosis by suppressing redox imbalance induced AMPK/mTOR signaling. Free Radic Biol Med 2025;237:195-209. [Crossref] [PubMed]
34. Yeom S, Lee DH, Song J. Therapeutic Potential of Anti-Diabetes Drugs and Anti-Dyslipidemia Drugs to Mitigate Head and Neck Cancer Risk in Metabolic Syndrome. CNS Neurosci Ther 2025;31:e70446. [Crossref] [PubMed]
35. Lyu X, Wang Y, Xu Y, et al. Metabolomic Profiling of Tumor Tissues Unveils Metabolic Shifts in Non-Small Cell Lung Cancer Patients with Concurrent Diabetes Mellitus. J Proteome Res 2024;23:3746-53. [Crossref] [PubMed]
36. Bosso M, Haddad D, Al Madhoun A, et al. Targeting the Metabolic Paradigms in Cancer and Diabetes. Biomedicines 2024;12:211. [Crossref] [PubMed]
37. Zheng X, Zhang W, Yang F, et al. Evaluative performance of TyG-ABSI versus traditional indices in relation to cardiovascular disease and mortality: evidence from the U.S. NHANES. Cardiovasc Diabetol 2025;24:344.
38. Jang H, Kim R, Lee JT, et al. Overall and abdominal obesity and risks of all-cause and cause-specific mortality in Korean adults: a pooled analysis of three population-based prospective cohorts. Int J Epidemiol 2023;52:1060-73. [Crossref] [PubMed]
39. Iliodromiti S, Celis-Morales CA, Lyall DM, et al. The impact of confounding on the associations of different adiposity measures with the incidence of cardiovascular disease: a cohort study of 296 535 adults of white European descent. Eur Heart J 2018;39:1514-20. [Crossref] [PubMed]
40. Zhang C, Quinones A, Le A. Metabolic reservoir cycles in cancer. Semin Cancer Biol 2022;86:180-8. [Crossref] [PubMed]
41. Dotan E, Tew WP, Mohile SG, et al. Associations between nutritional factors and chemotherapy toxicity in older adults with solid tumors. Cancer 2020;126:1708-16. [Crossref] [PubMed]
42. Georgakopoulou VE, Lempesis IG, Trakas N, et al. Lung cancer and obesity: A contentious relationship Oncol Rep 2024;52:158. (Review). [Crossref] [PubMed]
43. Lee YH, Han K, Yoon HE, et al. Triglyceride-Glucose Index and Risks of All-Cause and Cause-Specific Mortality in Young Adults. J Clin Endocrinol Metab 2025;110:e3607-16. [Crossref] [PubMed]
44. Liu M, Yan Z, Zhang Y, et al. Association Between the Triglyceride-Glucose Index and All-Cause Mortality Among Patients with Diabetes and Chronic Kidney Disease: A Retrospective Cohort Study. Diabetes Metab Syndr Obes 2025;18:2923-33. [Crossref] [PubMed]
45. Košuta D, Novaković M, Božič Mijovski M, et al. The Correlation between the Triglyceride-Glucose Index and Coagulation Markers in Patients with Recent Acute Myocardial Infarction. Dis Markers 2022;2022:6206802. [Crossref] [PubMed]
46. Baydar O, Kilic A, Okcuoglu J, et al. The Triglyceride-Glucose Index, a Predictor of Insulin Resistance, Is Associated With Subclinical Atherosclerosis. Angiology 2021;72:994-1000. [Crossref] [PubMed]