The triglyceride-glucose index (TyG) and Nonalcoholic fatty liver in the Japanese population: a retrospective cross-sectional study

Background: Evidences regarding the association between triglyceride-glucose index (TyG) and Nonalcoholic fatty liver (NAFLD) are controversial. Therefore, the goals of this research are to evaluate whether TyG is independently associated with NAFLD and the ability of TyG index to detect NAFLD in the Japanese population. Methods: The present study was a cross-sectional study. The data was downloaded from the DATADRYAD website. A total of 13178 participants was involved in a hospital in Japan from 2004 to 2015. The correlation between TyG and NAFLD was detected by using binary logistic regression and Generalized additive models. The likelihood ration test was used to examine the modification and interaction of subgroups. Furthermore, the ability of TyG to predict NAFLD was evaluated by using receiver operating characteristic (ROC) curves. The formula for the TyG index was ln [fasting triglyceride level (mg / dl) × fasting blood glucose level (mg / dl) / 2] Results: The average age of the selected participants was 43.36±8.89 years old, and about 51.02% of them were male. In fully-adjusted binary logistic regression model, TyG was positively related with the risk of NAFLD (Odds ratio (OR)=2.45, 95%CI 2.12-2.82). The relationship between TyG and NAFLD was a non-linear relationship, and its inflection point was 8.22. The effect sizes and the confidence intervals of the left and right sides of inflection point were 3.26(2.44 - 4.35) and 2.09 (1.72 - 2.54), respectively. By subgroup analysis, the stronger association was found in females, low GGT, non-obesity, non-visceral fat obesity (P for interaction <0.05). Among the total population, the AUC for TyG [0.810 (0.804 - 0.817)] was worse than ALT [ 0.829 (0.822 - 0.835)] but better than TG [ 0.799 (0.792 - 0.805)] and FPG [ 0.715 (0.707 - 0.722)]. The


Introduction
Non-alcoholic fatty liver disease (NAFLD) is characterized by ectopic triglycerides deposition in liver cells that includes steatosis, nonalcoholic steatohepatitis (NASH), fibrosis, and cirrhosis 1 . NAFLD is increasingly considered as a clinical syndrome, which is associated with type 2 diabetes mellitus (T2DM), cardiovascular diseases (CVD), and chronic kidney disease (CKD) 2 . The global prevalence of NAFLD is 25.24%, and it is rising rapidly 3 . It will cause huge health and economic burden for individuals, society, and nation. The gold standard of NAFLD is liver biopsy, which can accurately evaluate the important pathological manifestation of NAFLD, such as hepatic steatosis, inflammation, and fibrosis. Taking the test with high-risk and invasive into account, a non-invasive and convenient screening method is needed to identify patients with NAFLD in the large-scale epidemiological investigations.
The pathogenesis of NAFLD is considered to be insulin resistance(IR) and oxidative stress 4 . The product of fasting triglycerides and glucose levels (TyG), suggested by Simental-Mendía et al. 5 , is highly associated with IR 6-8 . Thus, TyG is a cheap and convenient clinical alternative indicator of IR.
In recent years, TyG has related to the risks of T2DM 9-12 , hypertension 13 Zheng et al. 25 suggested that there was no association between TyG and NAFLD in Brazil and China, respectively. The findings from previous studies regarding the relationship between TyG and NAFLD are controversial. In view of the differences in research design, target population, and data analysis of these studies, more research and exploration is needed to draw reliable conclusions. Therefore, the purpose of our research was to detect the relationship between TyG and the risk of NAFLD and to assess the predictor power of TyG for NAFLD in the Japanese population.

Data source and study population
Researchers can use the "DATADRYAD" database (www.Datadryad.org) to download raw data freely.

Data collection and measurements
Baseline information on all participants was obtained through questionnaires. Okamura et al. estimated the average weekly ethanol intake by investigating participants the type and amount of elcohol consumed each week in the previous month. They categorized participants into four categories: non-alcoholic or minimal drinking, <40 g per week; light drinking, 40-140 g per week; moderate drinking, 140-280 g per week; or heavy drinking, > 280 g per week. Participants were grouped into three categories based on smoking status: non-smokers were participants who had never smoked before, ex-smokers were defined as participants who had previously smoked but had quit smoking before the baseline survey, but current smokers were participants who continued to smoke during the baseline survey. For sports activities, they defined regular athletes as regularly participating in an activity every week. Those who met the following conditions are type 2 diabetes: HbA1c ≥ 6.5%, fasting plasma glucose ≥ 7 mmol / L or self-reported. BMI ≥ 25 kg / m2 was defined as obesity. Waist circumference (WC) ≥90 cm for men and waist circumference ≥80 cm for women were defined as visceral fat obesity. Hypertension was defined as systolic blood pressure (SBP) ≥140mmHg or diastolic blood pressure (DBP) ≥90mmHg. The triglyceride-glucose index (TyG) was calculated as Ln (TAG (mg/dL) × glucose (mg/dL))/2.

Definition of fatty liver
According to Asia-Pacific Working Party guidelines 27 , people with excessive drinking (> 140 g / week for men, > 70 g / week for women) or liver virus were excluded. The experienced operator performed abdominal ultrasound tests, and then, without knowing other data from the participants, the gastroenterologists diagnosed fatty liver which was based on liver contrast and liver brightness 28 .

Statistical analysis
The presentation of continuous variables was divided according to their distribution status: (1) normal distribution (mean ± standard deviation) and skewed distribution (median (quartile)).
Categorical variables were presented as frequency or as percentage. Whether the differences between groups were statistically different was tested by One-way analysis of variance (normal distribution), Kruskal Wallis H test (skewed distribution), and Chi-square test (categorical variable).
We used univariate and multivariate binary logistic regression to assess the correlation between TyG and NAFLD. According to the STROBE statement, three models were used to evaluate the relationship: unadjusted, minimally adjusted, and fully adjusted models. We selected potential covariates if they changed the estimates of TyG on NAFLD by at least 10% or were significantly correlated with NAFLD.
Thus, potential covariates involved in this study included age, BMI, WC, ALT, AST, GGT, HDL-C, TC, TG, HbA1C, FPG, SBP, DBP, and ethanol consumption, gender, regular exercise, Smoking status. For sensitivity analysis, we treated TyG as a continuous and categorical variable to evaluate its relationship to NAFLD. Besides, the nonlinear relationship between TyG and NAFLD was estimated using the Generalized Additive Model (GAM). If there was a non-linear relationship, we used the recursive method to calculate the inflection point, performed a two-piecewise linear regression, and used the maximum model likelihood test. To find modifications and interactions, we used a stratified linear regression model and a likelihood ratio test in the different subgroups. For the evaluation of diagnostic ability, we performed the receiver operating characteristic (ROC) curve analysis and compared the areas under the ROC curve (AUC) to find the strongest predictor of NAFLD. We

Baseline characteristics of selected participants
13,178 participants in total were selected for the present research through strict selection criteria. Table 1, grouped by TyG quartiles, shows the baseline characteristics of these selected participants.
In general, the average age of those selected was 43.36 ± 8.89 years old, and 51.02% of those were male. A total of 2361 had NAFLD and the prevalence was 17.92%. No statistically significant differences were detected among different TyG groups (all p values < 0.05). Compared to that of other three groups (Q 1-3), Participants in the highest group of TyG (Q 4) had the higher values in age, BMI, WC, ALT, AST, GGT, TC, TG, HbA1C, FPG, SBP, DBP and consisted of more male, past and current smoker, NAFLD than those of the other groups. The contrary results were detectd in HDL-C, female, regular exercise, non -smoker, non-NAFLD.

Univariate analysis
The results of the univariate analysis were showed in Table 2. We performed univariate binary logistic regression and found that HDL-C and regular exercise were negatively correlated with NAFLD.

Results of unadjusted and adjusted binary logistic regression
In this study, we used univariate and multivariate binary logistic regression to construct three models through to analyze to assess the independent effect of TyG on NAFLD. Table 3 shows the effect sizes (Odds ratio (OR)) and 95% confidence intervals. In the unadjusted model, TyG showed positive correlation with NAFLD (OR=7.79, 95% confidence interval (CI): 7.08 -8.56, P <0.0001). In the minimum-adjusted model, the result did not change significantly after adjusting for age and gender (OR=6.18, 95% CI 5.59 -6.83, P <0.0001). In fully adjusted model (adjusted age, gender, ALT, AST, GGT, SP, BMI, WC, regular exercise, HDL-C, TC, HbA1C, Ethanol consumption, Smoking status), the association between TyG and NAFLD could also be detected (OR=2.45, 95% CI 2.12-2.82, P <0.0001).
For the stability of our results, TyG was treated as categorical variable (Quartile of TyG) and the P for trend in the fully adjusted model was agreed with the result when TyG was a continuous variable.

The results of the nonlinearity of TyG and NAFLD
In the present study, the non-linear correlation between TyG and NAFLD was analyzed by the Smooth curve and the Generalized additive model (Figure 1 and Table 4). After adjusting for age, gender, BMI, WC, ALT, AST, GGT, HDL-C, TC, TG, HbA1C, FPG, SBP, DBP, ethanol consumption, regular exercise, and smoking status, we detected that the correlation between TyG and NAFLD was nonlinear. Using two-piecewise binary logistic regression and recursive algorithm, the inflection point of curve was calculated to be 8.22 (log-likelihood ratio test P 0.023). When TyG was less than 8.22, the effect size and 95% CI were 3.26, 2.44 -4.35, respectively. When TyG was greater than or equal to 8.22, the effect size and 95% CI were 2.09, 1.72 -2.54, respectively.

The results of subgroup analyses
The stratification variables (age, gender, ALT, AST, GGT, HDL-C, TC, HbA1c, obesity, visceral fat obesity, hypertension, smoking status, ethanol consumption, regular exercise) were used to assess the trend of effect sizes ( Table 5). The tests of interaction were statistically significant for these variables including gender, GGT, Obesity, visceral fat obesity. For men, one unit increase of TyG was correlated with 3.20 times higher risk of NAFLD. For women, an increase of the TyG by one unit was associated with 2.24 times higher risk of NAFLD. The change in gender is significant (P for interaction = 0,0076). The stronger association was also detected in low GGT (3.73 with low GGT vs 2.25 with middle GGT, and vs 2.12 with high GGT ), non-obesity (2.72 with non-obesity vs 1.99 with obesity), non-visceral fat obesity (2.65 with non-visceral fat obesity vs 1.88 with visceral fat obesity).

ROC analyses for TyG, ALT, FPG, and TG to predict the risk of NAFLD
For predicting the risk of NAFLD, we performed ROC analyses by using TyG, ALT, FPG, and TG in the AUC was worse for TyG than ALT but better than TG and FPG (P <0.0001). For women, the AUC for the TyG was superior to ALT, FPG, and TG (P<0.0001).

Discussion
The present study indicated TyG, both as continuous and categorical variables, was positively correlated with NAFLD after adjusting relative covariates. We firstly detected that the non-linear relation between TyG and NAFLD showed a threshold effect. When the inflection point is 8. Selecting 89 patients with morbid obesity, Cazzo, E et al. 24 did not detect a significant correlation between TyG and NAFLD in a cross-sectional study. The cross-sectional study found that LAP, rather than TyG, was independently associated with hepatic steatosis which was assessed by controlled attenuation parameter (CAP) among 101 women with polycystic ovary syndrome 25 . Previous studies are controversial due to different study designs, population ethnicity, sample size, and statistical method.
In the subgroup analysis, we found that BMI, WC, GGT, and gender affect the association between TyG and NAFLD after adjusting other covariates. Zhang, S et al. 19 also detected that the stronger association between TyG and NAFLD were detected in BMI < 25 kg/m 2 (OR 4.1 with BMI < 25 kg/m 2 vs OR 2.8 with BMI > = 25 kg/m 2 ) after adjusting related covariates, however, the association was not found in different gender. A meta-analysis showed that the prevalence of NAFLD was different in gender and men were more likely than women to suffer from NAFLD 32 . The potential mechanism is the sex hormone difference between men and women. Similar results have not been reported before, we are unable to explain why the associations between TyG and NAFLD are stronger in low GGT, non-visceral fat obesity.
To the best of our knowledge, Petta, S et al. 30 first showed that TyG index, whose AUC is 0.682, predicted the severity of liver steatosis in 340 Italian patients with Genotype 1 Chronic Hepatitis C.
Then, TyG, which was correlated with IR, could be used to diagnose liver steatosis but had deficiency quantitative evaluation for the severity of fatty liver, which was proposed by Fedchuk, L et al. 23           AUROCs for the prediction of NALFD in different gender.

Figure 2
AUROCs for the prediction of NALFD in different gender.

Figure 2
AUROCs for the prediction of NALFD in different gender.