In this study, we analyzed data collected from 1999 to 2018. The criteria for excluding samples included the following (1): lack of necessary parameters for assessing obesity and lipid-related indices; (2) absence of definitional information for diabetes and prediabetes; (3) missing covariate data; (4) absence of survival data ( Figure 1 ). The NHANES study protocol was approved by the Institutional Review Board of the National Center for Health Statistics (NCHS), and all participants provided written informed consent. For more detailed information about this study, please visit: www.cdc.gov/nchs/nhanes/irba98.htm.

The diagnosis of diabetes was based on one or more of the following criteria: (1) a medical diagnosis confirmed by the patient’s healthcare provider with self-reporting; (2) glycated hemoglobin (HbA1c) level ≥ 6.5%; (3) FPG level ≥ 7.0 mmol/L; (4) questionnaire results indicating that the patient is using diabetes medications. Prediabetes was defined by the following criteria: (1) a diagnosis confirmed by a healthcare professional through self-reporting; (2) HbA1c levels between 5.7% and less than 6.5%; (3) FPG levels between 100 mg/dL and 125 mg/dL (19).

In this study, the obesity and lipid-related indices included TyG, TyG-BMI, TyG-WWI, TyG-WHtR, TyG-WC, ABSI, LAP, and VAI. The calculation methods for TyG and its related obesity indicators are as follows. Since ABSI values are typically low, the values presented in this study are shown as 10-fold multiples (18):

VAI (13) is calculated as follows:

LAP (13) is calculated as follows:

The ABSI (12) is calculated as follows:

FPG: Fasting blood glucose; HDL-C: High density lipoprotein cholesterol; WWI: Weight-adjusted waist circumference index; WC: Waist circumference.

In this study, mortality data for all NHANES participants were matched with the National Death Index (NDI) using probabilistic matching methods, with a cutoff date of December 31, 2019. This process was used to calculate all-cause mortality rates. Additionally, considering the close relationship between obesity, lipid-related indices, and diabetes in relation to cardiovascular mortality risk, we used the death codes provided by the NDI to ascertain the cause of death. Causes of death were classified as cardiovascular-related according to the International Classification of Diseases, Tenth Revision (ICD-10), using relevant codes I00-I09, I11, I13, and I20-I51.

Covariate information for this study was collected from NHANES demographic data, questionnaires, and laboratory tests. We categorized the covariates into three main groups: baseline demographic data (age, gender, race), lifestyle factors (marital status, education level, BMI, alcohol consumption, and sedentary behavior), and comorbidities (hyperlipidemia, and CVD). The specific definitions are as follows:

The original five racial classifications from NHANES were condensed into three categories: Hispanic, Non-Hispanic Black, and Non-Hispanic White and Other; Marital status was categorized as divorced, married, or unmarried; Education level was classified as Below high school (Less Than 9th Grade), High school graduate or GED (9-11th Grade, including 12th grade with no diploma), and Some college or above (Some College or AA degree/College Graduate or above); BMI classifications were defined as normal (<25 kg/m²), obese (≥30 kg/m²), and overweight (≥25 kg/m² but <30 kg/m²); Sedentary behavior was defined as sitting or reclining for more than 480 minutes per day, or responding the questionnaire with an emphasis on a sedentary typical day; Alcohol consumption status was classified into three categories: current drinker (defined as having consumed more than 12 types of alcoholic beverages in their lifetime and currently consuming), former drinker (defined as having consumed more than 12 types of alcoholic beverages at any time during their lifetime but not in the past year), and never drinker (defined as having consumed no more than 12 types of alcoholic beverages in their lifetime); Hyperlipidemia was defined by any of the following criteria: total cholesterol levels equal to or exceeding 200 mg/dL, triglyceride levels equal to or exceeding 150 mg/dL, male HDL-C levels below 40 mg/dL, female HDL-C levels below 50 mg/dL, or low-density lipoprotein cholesterol (LDL-C) levels equal to or exceeding 130 mg/dL; CVD was defined as a positive response to any of the following questions: “Has a doctor or other health professional ever told you that you have congestive heart failure (CHF), coronary heart disease (CHD), angina, a heart attack, or a stroke?”

We employed a complex sampling design to ensure nationally representative estimates, and all analyses were adjusted for survey design and weighting variables. The new sample weights were calculated by dividing the original two-year sample weights by 20. Continuous variables are presented as means ± standard deviation (SD), while categorical variables are expressed as counts (N) and percentages (%). The obesity and lipid-related indices were categorized into four groups using quartiles. We used weighted t-tests (for continuous variables) or weighted chi-square tests (for categorical variables) to assess differences between survival and mortality group. The survival probabilities of diabetes/prediabetes patients under different obesity and lipid-related indices were compared using Kaplan-Meier (KM) curves and log-rank tests. The Cox regression model was used to analyze the mortality risk in diabetes/prediabetes patients, with model construction undergoing multiple adjustments: Model 1 adjusted for baseline demographic data; Model 2 further adjusted for lifestyle factors; and Model 3 adjusted for comorbidities on top of Model 2. A p-value of less than 0.05 was considered statistically significant for all two-sided tests. Furthermore, we employed a restricted cubic spline (RCS) model to treat obesity and lipid-related indices as continuous variables, investigating the linear and non-linear associations between these indices and mortality risk in diabetes/prediabetes patients by setting the 10th, 50th, and 90th percentiles as nodes of the RCS (20).

This study also employed machine learning (ML) modeling strategies to compare predictive abilities of obesity and lipid-related indices for mortality risk based on the Cox regression model. We used supervised MLs, integrating various obesity and lipid-related indices, components of each index, all covariates, and survival data into the machine learning dataset (21): extreme gradient boosting (XGBoost), decision tree (DT), robust support vector machine (RSVM), elastic net regression (Enet), multi-layer perceptron (MLP), logistic regression, random forest (RF), and k-nearest neighbors (KNN). The dataset was divided into two non-overlapping parts: a training set (60%) and a testing set (40%). In the training dataset, each model underwent automatic hyperparameter tuning using Bayesian optimization and five-fold cross-validation. When comparing the eight machine learning algorithms, we synthesized the assessment of the best algorithm using the receiver operating characteristic - area under the curve (ROC-AUC), accuracy, precision, recall, and calibration curves. We subsequently applied SHapley Additive Explanations (SHAP) to interpret the machine learning models, aiming to address the black box issue associated with these models. The Shapley value, derived from cooperative game theory, quantifies the importance of each feature in the model by calculating marginal contributions (22). We used the “fastshap” package to generate SHAP beeswarm plots to visualize each variable’s contribution to individual predictions. This clearly illustrates the significance of obesity and lipid-related indices and analyzes how the components of different indices contribute to and influence mortality risk.

This study included 4,058 participants with diabetes/prediabetes, of whom 640 (12%) died before December 31, 2019. Significant differences were observed between the mortality group and the survival group across multiple variables. First, the mortality group had a higher average age of 70.5 years and relatively fewer male survivors. Additionally, the mortality group had a higher proportion of Non-Hispanic White and other racial groups, as well as a greater proportion of married individuals. Notably, the mortality group exhibited higher rates of sedentary behavior, poverty, and low educational attainment, along with higher proportions of non-drinkers, individuals with hyperlipidemia, and CVD. Furthermore, some anthropometric measures, such as height, weight, and waist circumference, were slightly lower in the mortality group compared to the survival group. However, new anthropometric measurements derived from these indicators, such as the WWI and ABSI, were higher in the mortality group. Blood lipid and glucose levels were also significantly elevated in the mortality group. Additionally, the TyG was higher in the mortality group. Other derived anthropometric indices, such as TyG-WHtR and TyG-WWI, were also elevated compared to the survival group, with the VAI significantly higher as well. However, no significant difference was found in the LAP between the two groups ( Table 1 ).

We conducted a survival analysis on the indices that demonstrated statistical differences between the survival and mortality groups in Table 1 . The KM curves revealed that diabetes/prediabetes patients in the lowest quartile of TyG, TyG-WWI, and ABSI had significantly higher overall survival probabilities compared to those in the highest quartile (P = 5e-05, P < 2e-16, and P < 2e-16, respectively) ( Figures 2A–C ). Additionally, the TyG-BMI in the Q2 group demonstrated the highest survival probability (P = 4e-06) ( Figure 2D ), while TyG-WC, TyG-WHtR, LAP, and VAI did not show significant differences (P = 0.3, P = 0.2, P = 0.05, and P = 0.3, respectively) ( Figures 2E–H ).

We performed a quartile-based analysis of obesity and various lipid-related indices, including TyG, TyG-BMI, TyG-WHtR, TyG-WWI, TyG-WC, LAP, VAI, and ABSI. The results from the Cox regression analysis indicated significant associations between TyG, TyG-WWI, and ABSI and all-cause mortality in diabetes/prediabetes patients. After adjustment in Model 3, compared to baseline levels (Q1), the highest quartile of TyG (Q4: 9.18, 11.03) and TyG-WWI (Q4: 106.26, 144.15) respectively increased the risk of all-cause mortality, with a HR of 1.49 (95% CI: 1.09-2.03) and 1.52 (95% CI: 1.02-2.26). Furthermore, ABSI in Q3 (0.82, 0.85) and Q4 (0.86, 0.99) also indicated increased all-cause mortality risk, with HRs of 1.80 (95% CI: 1.23-2.64) and 1.68 (95% CI: 1.17-2.41), respectively. Additionally, the analysis of ungrouped continuous variables revealed that for each one-unit increase in TyG, TyG-WWI, and ABSI, the all-cause mortality risk increased by 1.4 times, 1.02 times, and 48.6 times, respectively ( Figure 3 , Supplementary Table S2 ).

Moreover, focusing solely on patients who died from cardiovascular causes ( Supplementary Table S1 ), the Cox regression analysis indicated a significant association between TyG and cardiovascular mortality in diabetes/prediabetes patients. After adjustment in Model 3, the highest quartile of TyG (Q4: 8.75, 11.03) was associated with an increased risk of cardiovascular mortality, with a HR of 1.98 (95% CI: 1.04-2.35). Additionally, the analysis of ungrouped continuous variables revealed that for each one-unit increase in TyG, the risk of cardiovascular mortality increased by 1.57 times ( Figure 3 , Supplementary Table S3 ). Additionally, TyG-WC in the Q4 was associated with elevated cancer-related mortality, with a HR of 3.09 (95% CI: 1.11–8.58) ( Figure 3 , Supplementary Table S4 ).

Cox regression analysis of race revealed that race-specific associations between obesity/lipid-related indicators and mortality risks. For Hispanic populations, elevated TyG quartiles (Q4) significantly increased all-cause mortality (HR = 2.35, 95% CI: 1.18–4.68) and cardiovascular mortality (HR = 2.42, 95% CI: 1.12–5.24), with per-unit TyG increases further amplifying risks (all-cause: HR = 1.94; Cardiovascular mortality: HR = 1.69). Non-Hispanic Black groups exhibited extreme obesity-driven risks, particularly with higher ABSI quartiles (all-cause mortality Q2–Q4 HRs = 3.79–3.00, all *P < 0.05; cardiovascular mortality per-unit HR = 49.3, 95% CI: 6.19–392, ***P < 0.001). Additionally, ABSI at the Q4 level demonstrated an elevated risk of all-cause mortality in both the Hispanic and Non-Hispanic White and Other groups. Non-Hispanic White/Other populations only showed significant cancer mortality risks with TyG-WC Q4 (HR = 3.26, 95% CI: 1.03–10.34). LAP/VAI showed no significant associations across races ( Tables 2 – 4 ).

Given that VAI and LAP were calculated based on gender, we further explored their predictive capacity for mortality through subgroup analysis in diabetes/prediabetes patients. The Cox regression analysis demonstrated a significant association between VAI and cardiovascular mortality specifically among males. Notably, after adjustment in Model 3, individuals in the Q3 (1.69, 2.75) and Q4 (2.76, 11.39) groups showed significantly increased cardiovascular mortality risks, with HR of 3.70 (95% CI: 1.21-11.3) and 3.43 (95% CI: 1.36-8.65), respectively. Furthermore, the analysis of ungrouped continuous variables indicated that for each one-unit increase in VAI, the cardiovascular mortality risk increased by 1.29 times. Finally, neither VAI nor LAP demonstrated significant differences in cancer-related mortality, with both showing no statistical significance ( Table 5 ).

Using multivariable-adjusted RCS analysis, we visualized the associations of various indices with all-cause and cardiovascular mortality in diabetes/prediabetes patients. The analysis of all-cause mortality revealed that both TyG and TyG-WWI displayed a linear relationship with all-cause mortality (overall P-values < 0.0001), with cutoff points where the HR exceeded 1 at 9.21 and 103.03, respectively ( Figures 4A, B ). In contrast, ABSI exhibited a non-linear relationship with all-cause mortality, with a non-linear P-value of 0.0391, showing cutoff points at 0.80 and 0.83 ( Figure 4C ). Additionally, among male diabetes/prediabetes patients, moderate to high levels of VAI (Q3: 1.69, 2.75; Q4: 2.76, 11.39) were associated with a clear linear relationship to cardiovascular mortality (overall P-value < 0.0001), with a cutoff point where HR exceeded 1 at 1.47 ( Figure 4D ).

After determining that TyG, TyG-WWI, and ABSI are significant risk factors for all-cause mortality in diabetes/prediabetes patients, we aimed to assess the importance of these indices and their contributions to mortality risk by comparing eight ML algorithms: XGBoost, DT, SVM, Enet, MLP, RF, and KNN. The results indicated that the XGBoost model achieved the highest area under the curve (AUC) value of 0.85, with an accuracy of 0.79, precision of 0.94, and recall of 0.81, indicating strong performance across various metrics ( Figures 5A, B , Supplementary Table S5 ). Furthermore, the calibration curve showed good consistency between the predicted probabilities from the XGBoost model and the actual probabilities ( Figure 5C ). Thus, we selected XGBoost as the optimal ML model for predicting mortality risk in this study. Subsequently, we used the SHAP model to interpret and visualize feature importance. The beeswarm plot illustrates the cumulative impact of each feature on mortality risk, arranged in descending order of importance. Positive SHAP values indicate that increasing feature values are directly correlated with higher mortality risk, with larger SHAP values contributing more significantly to the predictions.

In the global interpretability of the optimal XGBoost model, age emerged as the most significant contributor to mortality risk, consistent with clinical expectations. In the overall model that included covariates, the importance ranking of obesity and lipid-related indices was as follows: TyG > ABSI > TyG-WWI ( Figure 5D ). In the TyG-related model, apart from age, fasting blood glucose and triglycerides ranked second and third, respectively. Notably, very low fasting blood glucose also increased mortality risk ( Figure 6A ). A similar trend was observed in the TyG-WWI model, with fasting blood glucose and triglycerides carrying substantial weight, followed by weight and waist circumference ( Figure 6B ). In the ABSI model, waist circumference was the most significant factor, followed by weight and height, with very low weight significantly increasing the risk of mortality ( Figure 6C ).

In summary, both blood glucose-related indices and anthropometric measures substantially influence mortality risk in patients with diabetes/prediabetes. For cardiovascular mortality risk in male patients, the feature importance for VAI was ranked as follows: triglycerides, HDL-C, height, waist circumference, and weight, with triglycerides, waist circumference, and weight positively contributing to mortality risk, while other indicators showed a negative impact ( Figure 6D ).