Systematic comparison of triglyceride-glucose (TYG)-index with other mortality predictive models in obese patients with different mortality rates

Abstract

Background

The Triglyceride-Glucose (TYG)-Index has been increasingly used as a simple surrogate marker of insulin resistance and has been associated with adverse cardiometabolic outcomes in several populations. TYG has gained great attention as a predictive index for mortality, but comparisons with other predictive indexes are unexplored, except for TYG-derived indexes such as TYG-body mass index (TYG-BMI).

Methods

We conducted a comparative prognostic study in two independent cohorts of adults with obesity: a general obesity cohort (n = 1,359) and a bariatric surgery cohort (n = 854), both with long-term follow-up approaching 14 years and with different mortality rates (11.5 vs. 5.5%, respectively). We compared TYG index, TYG-BMI index, blood glucose, age, Charlson Index, metabolic syndrome, glucose tolerance, diabetes mellitus, through Cox proportional hazard models with Harrell’C index, and through ROC analysis. We also evaluated the possible incremental predictive value of the above prognostic indexes when combined with blood glucose, the TYG-index, and TYG-BMI index.

Results

Across both cohorts, several metabolic and clinical indices were significantly associated with all-cause mortality in univariable analyses. However, age and Charlson Comorbidity Index consistently showed the strongest discrimination and prognostic performance. The various indexes significantly predicted mortality at Cox proportional hazard models (p always < 0.001). Harrell’C index correlated with ROC area under the curves of each index (p < 0.001), and both Harrell and ROC correlated with quality indexes of Cox analysis (LR, p < 0.001) and with quality indexes of linear regression (F, p < 0.001). Findings were directionally consistent in the bariatric surgery cohort, although lower event rates attenuated overall discrimination. The combined use of more indices together was not uniformly useful to increase the predictive value of the above indices.

Conclusion

In obesity, TyG-based indices are associated with long-term mortality risk but add limited prognostic value beyond age and multimorbidity burden. These markers may be considered complementary tools for metabolic characterization rather than primary instruments for mortality risk stratification. This study reinforces the concept that various mortality indexes are as valid as, or even more predictive than, TYG index.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12933-026-03217-w.

Introduction

Obesity has become one of the most significant public health challenges worldwide, with prevalence rising steadily over the past decades. According to the World Health Organization, more than 650 million adults are obese, and obesity is a leading cause of morbidity and mortality due to its strong association with type 2 diabetes (diabetes), cardiovascular disease (CVD), certain cancers, and premature death [1]. The burden of obesity-related complications underscores the need for reliable, cost-effective tools to stratify risk and guide preventive strategies. Despite the widespread use of body mass index (BMI) to classify obesity, this metric has well-known limitations: it cannot distinguish between lean and fat mass, nor does it capture fat distribution (e.g., visceral versus subcutaneous adiposity), which is crucial in determining metabolic and cardiovascular risk [2, 3]. A key pathophysiological mechanism linking obesity to adverse outcomes is insulin resistance (IR). IR develops when insulin-sensitive tissues such as skeletal muscle, liver, and adipose tissue fail to respond adequately to normal insulin concentrations, resulting in impaired glucose uptake and increased hepatic glucose production [4]. The gold standard method to quantify IR is the hyper insulinemic–euglycemic clamp, which provides a direct measurement of glucose disposal under controlled conditions. However, this approach is invasive, costly, time-consuming, and impractical for large-scale clinical or epidemiological research. Consequently, surrogate markers of IR that are simple, inexpensive, and reproducible are of considerable interest. One of the most widely studied surrogate markers is the TYG-index; originally introduced as a low-cost surrogate index of insulin resistance [5, 6], the TYG-index is a risk factor for mortality and co-morbidities in the general population, in diabetes, and in patients with CVD [7, 8]. Several studies have investigated the predictive capacity of TYG-index, and 2,349 papers have appeared in PubMed dealing with TYG-index since 2022 [9]. The TYG-index is calculated using the formula ln [fasting triglycerides (mg/dL) x fasting glucoses (mg/dL)]/2 [5, 6]. Originally proposed by Guerrero-Romero and colleagues in 2008 as a practical indicator of IR [10], the TYG-index allows simultaneous tracking of changes in triglyceride and glucose levels. Its predictive validity was subsequently confirmed in studies demonstrating strong correlations with insulin sensitivity assessed both by the euglycemic clamp and by the homeostasis model assessment of insulin resistance (HOMA-IR) [10]. Khan and colleagues further validated the TYG-index as a reliable and low-cost alternative to more complex metabolic measurements [5]. The advantage of the TYG-index is that it relies exclusively on fasting triglycerides and glucose, two laboratory parameters that are routinely measured in virtually all healthcare settings, making the index widely applicable. Beyond its role as a surrogate of IR, TYG-index has been extensively investigated as a predictor of clinical outcomes. Several meta-analyses and large-scale cohort studies have shown that higher TYG levels are associated with increased risk of incident diabetes, metabolic syndrome, hypertension, and non-alcoholic fatty liver disease (NAFLD). More recently, growing attention has focused on the role of TYG-index as a prognostic marker for cardiovascular events and mortality. In the Prospective Urban Rural Epidemiology (PURE) cohort spanning > 140,000 adults across five continents and 22 countries, followed for over a decade, higher baseline TYG-index was prospectively associated with greater risks of major cardiovascular events (myocardial infarction and stroke), cardiovascular mortality, and incident diabetes, independent of conventional risk factors [8]. To enhance predictive capacity, several TYG-derived indices have been developed. Studies on the prognostic value of the TYG-index usually did not compare the results obtained with other predictive indexes, and the majority only compared TYG with TYG-derived indexes like TYG-BMI index or TYG-waist circumference (WC) or TYG-waist-to-height ratio (WHtR) [11, 12], showing a similar predictive capacity [13, 14]. In obesity, it was shown quite recently that TYG-index, just like blood glucose (BG) and metabolic syndrome [15], is associated with all-cause mortality [16]; a recent paper reported a negligible effect of either TYG-index or BG on the predictive effect of models including age and sex plus Charlson Index [17], diabetes, metabolic syndrome or glucose tolerance, as judged by Cox-proportional hazard model with Harrell’C index [18, 19]. In contrast, the predictive value of TYG-index in obese patients undergoing bariatric surgery has never been evaluated, and the available reports have shown a reduction of TYG-index associated with the surgery-linked slimming [20], while TYG-index seems to be associated with failure of diabetes to revert to normal after bariatric surgery [21]. The aim of this study was to systematically analyze and compare the value of several prognostic mortality indexes through two complementary analytic strategies, the ROC curve analysis and the Cox-proportional hazard model with Harrell’C index in a cohort of obese patients [16, 19]. To add value to this research, we also evaluated another population represented by obese patients undergoing bariatric surgery [22]. By combining these approaches, we aimed to rigorously assess the discriminative performance of different indexes and determine whether TYG-index or TYG-BMI index confer advantages over traditional predictors such as age, Charlson Index, diabetes, metabolic syndrome, glucose tolerance, and blood glucose. Finally, we evaluated the possible incremental predictive value of the above prognostic indexes when combined with blood glucose, the TYG-index, and TYG-BMI index.

Materials and methods

As previously reported [16, 19], we analyzed a cohort of 1,359 obese subjects: 371 men and 988 women, aged 44.1 ± 12.6 years, with a BMI 39.9 ± 5.2 kg/m², followed for a median period of 13.9 years. Clinical and laboratory characteristics of subjects of the cohort under study are reported in the publication [16, 19]. For this study we also analyzed a cohort of 854 obese patients, 210 men and 644 women, 41.4 ± 10.6 years, with a BMI 44.7 ± 7.2 kg/m², followed for a median period of 13.6 years [22].

Statistical analysis

(A) Analysis of single predictive factors

For both cohorts, the prognostic value for mortality of age and age quartiles (AgeQ), blood glucose (BG) and blood glucose quartiles (BGQ), TYG (TYG) and TYG quartiles (TYGQ), TYG-BMI (TYGBMI) and TYGBMI quartiles (TYG-BMIQ), diabetes, Charlson Comorbidity Index, Metabolic Syndrome quartiles (MSQ), Glucose Tolerance (GT) was investigated using Cox proportional models. In Cox analyses, data were expressed as the hazard ratio (HR), 95% confidence interval (CI), and p-value. The predictive accuracy of each prognostic model was assessed by calculating the Harrell C-index, ranging from 50 to 100%. A Harrell’s C index 50% indicates no discrimination ability of the model being tested, whereas a Harrell’s C of 100% indicates perfect discrimination [18]. Basic models were also analyzed through ROC curves, using the Stata software that provides a test for the equality of the area under the curves, using an algorithm suggested by DeLong, DeLong, and Clarke-Pearson [23]. The results were also correlated, i.e. the results of Harrell’s C index were correlated with ROC curves, and with indexes of quality of Cox analyses (Likelihood Ratio statistic, quality index of Cox analysis, LR) and of regressions (F test, quality index of linear regression, F). In detail, for each index, the correlation between Harrell’s C index and ROC curves, between Harrell’s C index and LR and F, between ROC curves and LR and F, and between LR and F was calculated and analyzed through Spearman correlation.

(B) Analysis of combined predictive factors

The second part of this analysis was centered on evaluation of the predictive value of models of increasing complexity, through various methods, for both Cox analysis and ROC curves. In addition, the New Reclassification Index (NRI) was employed, together with multivariable analysis, i.e. logistic regression analysis and stepwise regression analysis. Sensitivity analysis was also performed, taking into consideration single indexes and indexes combined with age; the sensitivity analysis considered age, sex, presence of diabetes and arterial hypertension, and serum creatinine. p values < 0.05 were considered statistically significant, and all statistical analyses were performed by Stata, version MP18.5, for MacIntosh.

Results

Table 1 shows baseline details of the two cohorts. Even though a direct comparison of the two cohorts was not among our objectives, patients of both cohorts were similar under several aspects, except for age and BMI; in addition, co-morbidities developing in the long-term were much less frequent in the cohort undergoing BS, as expected, and death rate was also lower in the cohort underdoing BS.

Table 1.

Clinical and metabolic details of subjects of the two cohorts

	First cohort (medical treatment)	Second cohort (bariatric surgery)
Number	1359	854
(Sex M/F)	(371/988)	(210/644)
% men	(27.3%)	(24.6%)
Age (y)	44.1 ± 12.6	41.4 ± 10.6
BMI (kg/m2)	39.9 ± 5.2	44.7 ± 7.2
Median duration follow-up (y)	13.9	13.3
BG (mg/dL)	118.1 ± 47.2	114.5 ± 47.4
TYG	8.9 ± 0.69	8.9 ± 0.69
TYGBMI	356.3 ± 54.58	399.6 ± 76.57
Charlson C.I	1.5 ± 1.25	1.2 ± 1.22
Charlson C.I. (without age)	0.98 ± 0.87	0.96 ± 0.99
Total-cholesterol (mg/dL)	212.8 ± 66.4	207.1 ± 44.5
LDL-cholesterol (mg/dL)	136.5 ± 64.3	143.2 ± 40.9
HDL-cholesterol (mg/dL)	50.0 ± 13.6	49.3 ± 13.1
Triglycerides (mg/dL)	159.5 ± 133.2	146.8 ± 74.5
ALT (U/L)	25.9 ± 13.8	24.0 ± 14.5
AST (U/L)	35.2 ± 24.0	33.4 ± 23.9
Creatinine (mg/dL)	0.8 ± 0.2	0.8 ± 0.2
AH (%)	425 (31.3%)	179 (21.0%)
T2DM (%)	131 (9.6%)	163 (19.1%)
Metabolic syndrome (%)	717 (52.8%)	191 (22.4%)
CVD (%)	51 (3.8%)	19 (2.3%)
Incident T2DM	196 (14.4%)	24 (2.8%)
Incident AH	289 (21.3%)	67 (7.9%)
Incident CVD	273 (20.1%)	62 (7.3%)
All-cause mortality
(All-cause) %	154	46
Sex-related mortality	(11.3%)	(5.4%)
Men (%)	59 (15.9%)	18 (8.6%)
Women (%)	95 (9.6%)	28 (4.3%)

BG = blood glucose; TYG = triglyceride–glucose-index [ln triglycerides (mg/dL) × blood glucose (mg/dL)/2]; TYGBMI = TYG x BMI; Charlson C. I. = Charlson Comorbidity Index; ALT = alanine transaminase; AST = aspartate transaminase; AH = arterial hypertension; DM = diabetes mellitus; CVD = cardiovascular disease;

Table 2 shows the main features of the various predictive indexes used in the study, including Cox regressions analysis and linear regression, with death as the dependent variable. In the first cohort the predictive value of all indexes was < 0.001, both at Cox analysis and at linear regression. In the second cohort the predictive value of most indexes was again < 0.001, with the noticeable exception for TYG-BMIQ (NS). ROC analysis curves, and Harrell’C index are also shown in Table 2.

Table 2.

Main details of the various prognostic indexes, at Cox analysis, at linear regression, at ROC analysis, and at Harrell’C index in the two cohorts

(a) People obese (1359) followed for about 13.9 ± 4.34, dead patients 154 [16]
Cox analysis					Linear regression				ROC	Harrell’C
stcox	HR	95% C.I	LR	p	r	95% C.I	F	p	AUC	Index
Age	1.082	1.065–1.099	114.13	< .001	.006	.004-.008	102.20	< .001	.7475	0.7367
AgeQ	2.270	1.903–2.709	103.14	< .001	.073	.059-.088	96.8	< .001	.7274	.7164
Sex (W vs. M)	0.562	0.406–0.778	11.42	.001	− .062	− .101-.025	10.68	.001	.4379	.5621
Charlson C.I	1.830	1.639–2.045	106.19	< .001	.073	.061-.086	121.31	< .001	.7349	.7226
GT	1.956	1.619–2.365	51.08	< 001	.069	.049-.088	48.4	< .001	.6533	.6543
DM	3.027	2.205–4.156	45.56	< .001	.128	.092-.165	47.47	< .001	.6316	.6283
MS	2.568	1.801–3.661	30.40	< .001	.091	.057-.124	28.32	< .001	.6126	.6121
MSq	1.692	1.429–2.004	40.31	< .001	.052	.035-.068	38.10	< .001	.6451	.6449
BG	1.009	1.006–1.011	46.22	< .001	.001	.001-.001	64.21	< .001	.6857	.6825
BGQ	1.808	1.539–2.124	59.28	< .001	.697	.041-.071	55.05	< .001	.6746	.6712
TYG	1.965	1.624–2.377	41.94	< .001	.082	.058-.107	45.13	< .001	.6563	.6525
TYGQ	1.621	1.391–1.889	41.74	< .001	.048	.033-.062	39.35	< .001	.6477	.6436
TYG-BMI	1.005	1.003–1.008	16.35	< .001	.001	.000-.001	18.54	< .001	.6094	.6022
TYG-BMIQ	1.435	1.239–1.663	24.41	< .001	.037	.022-.052	23.69	.001	.6155	.6085

(b) People (854) undergoing BS followed for about 13.6 ± 4.14, dead patients 46 [19]
Cox analysis					Linear regression				ROC	Harrell’C
stcox	HR	95% C.I	LR	p	r	95% C.I	F	p	AUC	Index
Age	1.075	1.045–1.107	26.28	< .001	.004	.002-.005	26.37	< .001	,6912	.7115
ageQ	2.060	1.504–2.823	24.90	< .001	.033	.019-.046	23.90	< .001	.6797	.6988
Sex (W vs M)	0.494	0.273–0.893	5.09	.020	− .042	− .077-.007	5.57	.019	.4232	.5755
Charlson C.I	2.508	1.641–3.833	22.05	< .001	.042	.024-.060	20.77	< .001	.6686	.6796
GT	2.412	1.726–3.371	26.99	< .001	.053	.034-.072	30.93	< .001	.6696	.6815
DM	4.545	2.548–8.105	24.13	< .001	.108	.069-.146	31.11	< .001	.6634	.6532
MS	1.414	1.097–1.822	6.48	.007	.021	.006-.036	7.15	.008	.6327	.6212
MSq	1.554	1.175–2.054	10.30	.002	.022	.008 .035	9.96	.002	.6360	.6213
BG	1.005	1.002–1.009	5.44	.007	.001	.001-.001	7.72	.006	.6695	.6519
BGQ	1.678	1.237–2.277	12.19	.001	.032	.016-.047	16.10	< .001	.5718	.6362
TYG	1.856	1.249–2.755	7.76	.002	.044	.015-.074	8.64	.003	.6462	.6282
TYGQ	1.620	1.209–2.171	11.43	.001	.027	.011-.042	11.03	.001	.5880	.6300
TYG-BMI	1.005	1.002–1.009	9.90	.001	.004	.001-.006	10.01	.002	.6552	.6454
TYG-BMIQ	1.737	0.765–3.975	1.80	.186	.018	− .063-.052	1.87	.173	.6392	.6356

Initially, we tested the ability of various indexes to discriminate the risk of death at Cox regression (Fig. 1); age quartiles, Charlson comorbidity index, presence of diabetes and glucose tolerance were analyzed (Fig. 1a) as well as, BGQ and metabolic syndrome, TYGQ, and TYG-BMIQ (Fig. 1b). The risk of mortality is extremely lower in the second cohort of patients undergoing bariatric surgery, as opposed to the first cohort, and the predictive role of various indexes seems similarly affected.

Fig. 1.

Cox proportional models for mortality. Patients are subdivided according to the cohort. Subjects at risk are indicated for each analysis. a: Age Quartiles; Charlson = Charlson Comorbidity Index; DM = Diabetes mellitus; GT = glucose tolerance; b: BGQ = Blood glucose quartiles; TYGQ = TYG quartiles; SMQ = Metabolic syndrome; TYG-BMIQ = TYG-BMI quartiles

Next, we tested the predictive mortality indexes using ROC analysis. In the first cohort, the maximum AUC was greatest for Age, followed by BGQ, TYGQ, and TYG-BMI (Fig. 2); in the second cohort, the predictive value of all variables was not significantly different (Fig. 2), probably because of a far lower mortality risk.

Fig. 2.

ROC analysis of the various predictive indexes for mortality. Patients are subdivided according to the cohort. CHARLSON = Charlson Comorbidity Index; TYGQ = TYG quartiles; BGQ = Blood glucose quartiles; GT = Glucose tolerance; TYG-BMIQ = TYG-BMI quartiles

Finally, Harrell’s C index and ROC area under the curves (AUCs) were correlated. In the first cohort the two indexes showed a high correlation (p = 0.001) (Fig. 3a), as it was for the correlation of Harrell’C index wth LR (Cox analysis, p < 0.001); ROC AUCs with LR (r = p < 0.001) and F (p < 0.001), and LR with F (p < 0.001) (Fig. 3b). In the second cohort a similar correlation was found between Harrell’s C index and ROC AUCs (p = 0.002), between Harrell’C index and LR (p = 0.041), and between Harrell’C index and F (p = 0.001, Fig. 3a), while the correlation between ROC AUCs and LR and F was not significant, in spite of a significant correlation between LR and F (p = 0.001, Fig. 3b).

Fig. 3.

a: correlation between Harrell’C indexes and ROC AUCs, between Harrell’C indexes and quality index of Cox analysis, and between Harrell’C indexes and quality indices of Cox analysis (LR) and between Harrell’C index and quality index of linear regression (F). Correlations and their significance are shown. b: correlation between ROC AUCs and quality index of Cox analysis (LR), between ROC AUCs and quality index of linear regression (F), and between LR and F. Correlations and their significance are shown

The incremental predictive power of various indexes combined is shown in Table 3. The incremental value of Harrell’C index (Table 3a, b) and of NRI (Table 3c, d) are represented for both cohorts, plus NRI for ROC AUC (Table 3e, f), and results are summarized in Table 4. Finally, ROC curves for both cohorts are shown in Fig. 4; ROC analysis confirms that, at difference from the first cohort, the predictive value of all variables was not significantly different in the second cohort, probably because of a far lower mortality risk. Supplementary Fig. 1 shows the behavior of various indices when calculated with sex, and with sex plus BG, TYG, TYG-BMI in the first cohort, and indicates that Charlson C.I. was always the strongest predictive index.

Table 3.

Incremental prognostic value of Harrell’C index (A and B) and of New Re-classification Index (NRI) (C and D), calculated for the first and the second cohort of obese patients in the study, and on ROC AUC (E and F)

Harrell’C index
(A) Harrell’C index, First cohort
Index	basal	+ BG	+ TYG	+ TYGBMI	+ BGQ	+ TYGQ	+ TYGBMIQ
Basic	.7515	.7622	.7629	.7563	.7625	.7609	.7581
Charlson C.I	.7559	.7658	.766	.7655	.7646	.7632	.7658
DM	7600	.7626	.7643	.7626	.7622	.7626	.7626
GT	.7607	.7628	.7642	.7625	.7626	.7632	.7631
MS	.7609	.7648	.7634	.7592	.7649	.7624	.7604
Mean ± SD	.7578 ± .0041	.7636 ± .0016*	.7642 ± .0012*	.7612 ± .0035	.7634 ± .0013*	.7625 ± .0001*	.7620 ± .0029*
Comparisons (p)		NS versus TYG	0.0275 versus TYGBMI	NS versus BG	NS versus TYGQ	0.0272 versus TYGBMIQ	NS versus BGQ

(B) Harrell’C index, Second cohort
Index	basal	+ BG	+ TYG	+ TYGBMI	+ BGQ	+ TYGQ	+ TYGBMIQ
Basic	.7425	.7289	.7548	.7892	.7275	.7538	.7879
Charlson C.I	.7600	.7326	.7532	.7921	.7286	.7552	.8654
DM	.7785	.7609	.7609	.7889	.7589	.7664	.8141
GT	.7696	.7433	.7643	.7862	.7426	.7609	.8297
MS	.7471	.7280	.7535	.7929	.7284	.7557	.8046
Mean ± SD	.7595 ± .0151	.7387 ± .0138*	.7573 ± .0051	.7899 ± .0027*	.7372 ± .0137*	.7584 ± .0052	.8203 ± .0294*
Comparisons (p)		0.0088 versus TYG	0.0003 versus TYGBMI	0.0009 versus BG	0.0025 versus TYG	0.0048 versus BG	0.0023 versus BG

NRI on increment of indexes
(C) NRI, First Cohort
Main marker	+ TYG	+ BG	+ TYGBMI	+ TYGQ	+ BGQ	+ TYGBMIQ
Charlson C.I	.3413*	.232 *	.1238	.3091*	.3696*	.235*
Charlson C.I. WA	.2351*	.0891*	.1172	.2834*	.298*	.2051*
GT	.2145*	-.1036	.0953	.2649*	.3245*	.2511
DM	.2041*	.0487	.1628	.3978*	.3297*	.1595
MS	.2755*	.3172*	.1095	.2882*	.2721*	.0877
Basic	.4792	.3352*	.1647	.3491*	.379*	.2614*
Total	.29 ± .105	.15 ± .172	.13 ± .029	.31 ± .049	.33 ± .041	.20 ± .066
Comparisons (p)	.0159 versus BG	.0042 versus TYGBMI	NS versus BG	.0106 versus TYGBMIQ	NS versus TYGQ	.0004 versus BGQ

(D) NRI, Second Cohort
Main marker	+ TYG	+ BG	+ TYGBMI	+ TYGQ	+ BGQ	+ TYGBMIQ
Charlson C.I	.1291	− .2756	.2529	.0917	− .0825	.3176
Charlson C.I. WA	− .0423	− .3763	.2247	.0507	− .2711	.3802 *
GT	− .0412	− .3763	.1348	0.0867	− .1482	.2879
DM	− .083	− .3749	.1971	.0397	− .2165	.3027 *
MS	.0468	− .3329	.1957	.1263	− .0751	.33 *
Basic	.0546	− .2632	.1595	.1079	− .75	.3304
Total	.01 ± .081	− .33 ± .052	.19 ± .04	.08 ± .037	− .26 ± .253	.033 ± 0.032
Comparisons (p)	.0001 versus BG	.0001 versus TYGBMI	.0460 versus TYG	.00115 versus BGQ	.0001 versus TYGBMIQ	.0001 versus TYG

NRI on increment of ROC AUC
(E) First Cohort
Main marker	+ TYG	+ BG	+ TYGBMI	+ TYGQ	+ BGQ	+ TYGBMIQ
Charlson C.I	.5099	− .0946	− 0.2774	− 0.0246	− 0.0012	− 0.2204
Charlson C.I. WA	.4975	.0116	− 0.1141	0.0865	0.1970	− 0.0670
GT	.4533	-.2080	− 0.2671	− 0.0674	0.0984	− 0.2453
DM	.4697	.2770	− 0.2624	− 0.0484	0.5715	− 0.2530
MS	.4620	.4551	− 0.2024	− 0.0295	0.0253	− 0.1503
Basic	.4785	.558	0.2405	0.4031	0.4472	0.2597
Total	.48 ± .022	.17 ± .310	− 0.15 ± 0.199	0.05 ± 0.179	0.22 ± 0.235	− 0.13 ± 0.195
Comparisons (p)	.0300 versus BG	.0119 versus TYGBMI	0.0003 versus TYG	NS versus BGQ	.0104 versus TYGBMIQ	0.0000 versus TYGQ

(F) Second Cohort
main marker	+ TYG	+ BG	+ TYGBMI	+ TYGQ	+ BGQ	+ TYGBMIQ
Charlson C.I	.0522	− .0362	− 0.1278	− 0.0293	− 0.0362	− 0.0832
Charlson C.I. WA	.3289	− .0029	0.1667	0.0322	0.3289	0.0761
GT	− .0637	− .3105	− 0.1808	− 0.0985	0.0156	− 0.0646
DM	− .1363	− .2868	− 0.1515	− 0.0913	− 0.0158	− 0.1097
MS	.1692	− .0673	− 0.1883	− 0.0127	0.1416	− 0.0095
Basic	.2556	.2934	0.1843	0.1783	0.3422	0.1407
Total	.10 ± .182	− .07 ± .220	− .05 ± .176	− .00 ± .102	.13 ± .171	− .01 ± .098
Comparisons (p)	.0127 versus BG	NS versus TYGBMI	0.0130 versus TYG	0.0119 versus BGQ	0.0039 versus TYGBMI	NS versus TYG

^*Significant versus basal

Table 4.

Summary of increases of predictive role of indexes

	Harrell’C for indexes	NRI for indexes	NRI for ROC AUC
1st cohort
Numerical increase	TYG and BG	TYG	TYG
Quartile increase	TYGQ and BGQ	BGQ	BGQ
2nd cohort
Numerical increase	TYGBMI	TYG	TYG
Quartile increase	TYGBMIQ	TYGBMIQ	BGQ

Fig. 4.

Incremental ROC analysis of the various combined indices for mortality. Patients are subdivided according to the cohort. AGESEXCHARLSONTYG = age x sex x CHARLSON x TYG; AGESEXTYG = age x sex x TYG; AGESEXDMTYG = age x sex x TYG x DM; AGESEXMSTYG = age x sex x TYG x MS; AGESEXGTTYG = age x sex x TYG x MS; AGESEXMSBG = age x sex x BG x MS; AGESEXdmBG = age x sex x BG x DM; AGESEXGTBG = age x sex x BG x GT; AGESEXBG = age x sex x BG; AGESEXCHARLSONBG = age x sex x BG x CHARLSON; AGESEXCHARLSONTYGBMI = age x sex x TYGBMI x CHARLSON; AGESEXTYGBMI = age x sex x TYGBMI x; AGESEXDMTYGBMI = age x sex x TYGBMI x DM; AGESEXGTTYGBMI = age x sex x TYGBMI x GT

Sensitivity analysis is depicted in Table 5. The indexes considered alone were virtually always significant; instead, when the indexes were considered together with age, non-significant values emerged (1–5/10) for several indexes. Stepwise and logistic analysis of various models of increasing complexity are shown in the supplementary appendix, showing that in the first cohort TYG or TYGQ and BG or BGQ are significantly associated, either at cox proportional analysis or at stepwise or at logistic regression, while in the second cohort TYG-BMI is the only marker significantly associated at either cox proportional analysis or at stepwise or at logistic regression.

Table 5.

Sensitivity analysis of predictive indexes of mortality

1 factor
	Age < 50	Age > 50	F	M	DM yes	DM no	AH yes	AH no	Crea yes	Creano	NS
Age	.001	.000	.000	.000	.000	.000	.000	.000	.000	.000	0/10
ageQ	.009	.000	.000	.000	.000	.000	.000	.000	.000	.000	0/10
Charlson	.004	.000	.000	.000	.047	.000	.000	.000	.000	.000	0/10
Charlson WA	.002	.000	.000	.000	.005	.003	.000	.000	.000	.000	0/10
DM	.000	.000	.000	.001	–	–	.007	.000	.006	.000	0/10
MS	.002	.004	.000	.004	.035	.001	.015	.029	.000	.000	0/10
BG	.000	.000	.000	.003	.012	.004	.000	.000	.006	.000	0/10
TYG	.001	.000	.000	.007	.016	.018	.000	.000	.006	.000	0/10
TYG-BMI	.000	.044	.001	.015	.495	.030	.165	.000	.012	.002	2/10
TYGQ	.001	.000	.000	.005	.017	.022	.001	.000	.005	.000	0/10
BGQ	.000	.000	.000	.002	.031	.012	.000	.000	.001	.000	0/10
TYG-BMIQ	.001	.006	.001	.002	.159	.019	.011	.001	.003	.002	1/10

2 factors
	Age < 50	Age > 50	F	M	DM yes	DM no	AH yes	AH no	Crea yes	Creano	NS
Age, Charlson	0.028	0.001	0.001	0.043	0.013	0.267	0.004	0.004	0.362	0.000	2/10
Age, Charlson WA	0.028	0.000	0.001	0.017	0.009	0.353	0.004	0.006	0.300	0.000	2/10
Age, GT	0.001	0.007	0.000	0.479	–	0.607	0.009	0.019	0.730	0.000	3/9
Age, DM	0.001	0.004	0.000	0.233	–	–	0.009	0.008	0.000	0.822	2/8
Age, MS	0.011	0.044	0.028	0.078	0.710	0.024	0.033	0.156	0.780	0.001	4/10
Age, BG	0.006	0.000	0.000	0.297	0.021	0.482	0.000	0.014	0.823	0.000	3/10
Age, TYG	0.000	0.000	0.000	0.055	0.017	0.121	0.001	0.005	0.000	0.703	3/10
Age, TYG-BMI	0.001	0.434	0.075	0.123	0.621	0.253	0.411	0.019	0.055	0.047	5/10
Age, TYGQ	0.002	0.001	0.003	0.032	0.043	0.128	0.005	0.008	0.947	0.000	2/10
Age, BGQ	0.000	0.008	0.000	0.379	0.070	0.679	0.002	0.038	0.435	0.000	4/10
Age, TYGBMIQ	0.002	0.156	0.067	0.051	0.258	0.252	0.036	0.100	0.091	0.011	6/10

Data are expressed as index alone, and index combined effect with age

Discussion

In this study, we systematically compared the prognostic value of the TYG-index, TYG-BMI, and several conventional mortality predictors, including age, Charlson Comorbidity Index, blood glucose, glucose tolerance, diabetes mellitus, and metabolic syndrome. We utilized a large cohort of 1359 obese individuals followed for nearly 14 years, and a second cohort of 854 obese individuals, undergoing bariatric surgery, also followed for nearly 14 years. From the first cohort of obese individuals, our main finding is that although all investigated indexes significantly predicted mortality, age and Charlson Index consistently outperformed TYG index and TYG-BMI in terms of discriminative ability, as reflected by both ROC analysis and Harrell’s C index. Importantly, we observed strong concordance between the two statistical approaches, supporting the robustness of our results. Also, the two indexes correlated with quality index of Cox analysis and of linear regression. In agreement with a previous paper taking advantage of Cox proportional models and Harrell’s C index [19], use of ROC curve analysis in our study demonstrated that various predictive mortality indexes are as valid as TYG-index, or even superior to TYG-BMI. Our findings align with, and support, previous data indicating that TYG-index, while associated with mortality risk, does not add significant incremental predictive value once classical predictors are considered [19]. The data obtained from the second cohort are in agreement with the results of the first cohort, showing once again that TYG-index is not superior to other indexes in predicting mortality. Nevertheless, the second cohort has a far lower mortality risk compared to the first cohort (about 5.4% vs 11.3%, p < 0001). We hypothesize that differences between the first and the second cohort were mainly due to some “floor effect”, that does not contradict our findings. So far, only one study assessed both ROC curves and Harrell’s C indexes comparing TYG-index and TYG-derived indexes, but with no direct comparison between the two indexes [24]. In this study, we also revealed an absolute concordance between the results obtained using the two indexes, further supporting the validity of both statistical approaches in assessing the performance of disitnct mortality indexes.

In this study we also considered the predictive role of combined indexes on mortality; in particular, we evaluated the effect of the addition of BG and BGQ, of TYG and TYGQ, of TYG-BMI and TYG-BMIQ in both cohorts, and we made this through stepwise and logistic regression analysis, and through NRI for indexes and for ROC AUCs. Even though a significant increase of the predictive capacity of indexes was observed in some cases, the increase was not different among BG, TYG, TYG-BMI, or among BGQ, TYGQ, TYG-BMIQ; also, there was not necessarily an increase of the predictive role of indexes, as it could also be a decrease.

Large-scale studies have generally supported the predictive role of TYG-index. For instance, the PURE study found consistent associations of TYG-index with all-cause and cardiovascular mortality across diverse global populations [8]. Zhang et al. [12] demonstrated that TYG-related indices were associated with mortality among individuals with cardiovascular-kidney-metabolic syndrome stages 0–3. Chen and colleagues assessed young and middle-aged adults and reported that TYG-index predicted long-term cardiovascular events, with stronger effects among obese participants [11]. However, these studies often focus on broad or heterogeneous populations, and the strength of association varies depending on baseline risk, age distribution, and comorbidity burden. Our results suggest that in obese populations with substantial comorbidity, traditional predictors remain dominant, and TYG-index adds little beyond them [19]. Regarding clinical implications, the appeal of TYG-index lies in its simplicity, and its low cost is undoubted. TYG-index relies on two parameters that are routinely measured in clinical practice, making it a convenient marker for large-scale screening. In resource-limited settings, TYG index may serve as a useful tool to flag individuals at higher cardiometabolic risk, particularly when age or comorbidity data are unavailable or incomplete. However, in specialized care for obesity, where patients are already carefully assessed, our findings indicate that TYG-index and TYG-BMI do not provide superior prognostic information compared to age and Charlson Index. From a clinical perspective, this emphasizes the relevance of comprehensive evaluation of patients, including comorbidity assessment, rather than relying on a single biomarker. Moreover, given the strong intercorrelations observed among the various indexes, clinicians should be cautious about interpreting TYG index as independent of other risk measurements. Overall, TYG-index may reflect overlapping pathophysiological processes captured more directly by established indices. Our data reinforces the concept that TYG and TYG-derived indexes are not magic indexes, and that there is the need of a great volume of research to show whether TYG and TYG-derived indexes are at all different from other indexes.

Strengths and limitations

An important contribution of our study is the dual assessment of predictive accuracy using ROC curves and Harrell’s C index. ROC analysis provides an intuitive measure of discrimination ability, while Harrell’s C index is well-suited for survival data and accounts for censored observations. The high correlation observed between these two metrics in our study (r ≈ 0.95) underscores their concordance and reinforces confidence in the validity of our findings. To our knowledge, few studies have applied both methods simultaneously to compare TYG-index with other predictors, making our work a methodological reference for future research.

The other strengths of our study include the relatively large sample size of the two cohorts of 1,359 and of 854 obese individuals, the long median follow-up of nearly 14 years, and the use of complementary statistical approaches. The detailed clinical characterization of participants allowed us to evaluate multiple predictors simultaneously and to explore their correlations. Nevertheless, some limitations warrant consideration. First, our cohort consisted exclusively of obese individuals recruited in a single country, which may limit generalizability to lean populations or to different ethnic and geographic contexts. Second, our analysis was based on baseline measurements only; changes in metabolic parameters over time were not captured, and dynamic trajectories may provide additional prognostic information. Third, we did not include other emerging markers of cardiometabolic risk, such as inflammatory biomarkers, genetic variants, or imaging-based measures of adiposity, which could potentially interact with TYG-index. Finally, external validation in independent cohorts is necessary to confirm the robustness of our conclusions.

Further research is needed to clarify the contexts in which TYG-index offers incremental prognostic value. For example, TYG-index may be particularly useful in younger populations or in those without overt comorbidities, where age and Charlson Index are less discriminative. Studies integrating TYG-index with novel biomarkers of inflammation, liver function, or visceral adiposity could explore synergistic effects. Additionally, machine learning approaches may help integrate TYG-index into multiparametric risk models, potentially uncovering patterns not evident in traditional analyses. It will also be important to assess whether repeated measurements of TYG-index over time improve prognostic accuracy compared with single baseline values. Longitudinal trajectories of TYG-index may better capture the dynamic interplay of dyslipidaemia and glycemia, especially in populations undergoing lifestyle or pharmacological interventions, or bariatric surgery. Finally, given the global burden of obesity and diabetes, evaluating TYG-index across diverse ethnicities and socioeconomic settings remains crucial for determining its universal applicability.

Our study demonstrates that although the TYG-index and TYG-BMI index are valid predictors of mortality in obesity, but that they are not superior to traditional predictors such as age and Charlson Comorbidity Index. The strong concordance between ROC analysis and Harrell’s C index underscores the robustness of these findings. While TYG-index remains a useful and inexpensive marker, its role should be viewed as complementary rather than superior to established indices. Future studies should focus on identifying specific populations or clinical contexts where TYG-index provides independent prognostic value, and on integrating it into multiparametric models that reflect the multifactorial nature of cardiometabolic risk.

In conclusion, this study reinforces once again the concept that various mortality indexes are as valid or even more valid than TYG-index, highlighting the necessity to compare in future studies the value of TYG-index with other predictive indexes.

Author contributions

AEP, LLS and IZ conceptualized the study SM, VC, MR were responsible for investigation and analysis AEP, LLS and MR prepared all figures LLS and SM prepared tables AEP, LLS, and IZ. Wrote the main manuscript all authors reviewed the manuscript.

Funding

Fondazione Romeo ed Enrica Invernizzi, Ministero della Salute,ricerca corrente.

Data availability

The data sets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The study protocol [16] was approved by local Ethics Committees in 2015 (Coordinating Center: Ospedale San Paolo, Comitato Etico Interaziendale di Milano Area A, official approval SC: 2015 ST 125), and written informed consent was obtained from all adult participants. For the study under [19], the Ethics Committee of IRCCS MultiMedica notified the Authors that because of the nature of the study (non-interventional retrospective analysis of anonymized data), the study could be approved without further analysis by the Ethics Committee.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.