Analysis of the association between urinary glyphosate exposure and fatty liver index: a study for US adults

NHANES was established by the National Center for Health Statistics (NCHS) of the Centers for Disease Control and Prevention (CDC), and contains information on questionnaires, exams, and laboratory tests for selected study participants. The data contained in NHANES is open to the public, which allowed us to be exempt from ethical review of the study [23].

A total of 20,146 participants were included in the NHANES 2013–2016 survey. Relying on the determination of adult age in previous NHANES-based cross-sectional studies [24‐26], we first excluded participants younger than 20 years of age (n = 8658). Participants without clear uGLY information were then excluded (n = 8420), followed by those missing laboratory indicators used to calculate FLI, which were glutamyl transpeptidase (GGT) (n = 94), triglycerides (TG) (n = 2), body mass index (BMI) (n = 18), and waist circumference (WC) (n = 104). Since there is a close association between steatosis of the liver and viral hepatitis [27, 28]. Therefore, participants who were positive for hepatitis B virus (HBV) (n = 15) and hepatitis C virus (HCV) (n = 28) were also not included in this study. As urine is one of the important routes for GLY excretion out of the body [29], we excluded patients with renal weakness or renal failure (n = 78). Previous studies have confirmed that the side effects of glucocorticoids, tamoxifen and methotrexate lead to disturbances in fat metabolism of the liver [30‐32]. Therefore, we excluded participants (n = 17) who had apparently used medications that interfere with fat metabolism ( in the past month). Specific drugs included meprednisolone (n = 3), prednisolone (n = 11), tamoxifen (n = 1) and methotrexate (n = 2). World Health Organization (WHO) recommends that if a urine sample is too dilute (creatinine concentration < 30 mg/dL) or too concentrated (creatinine concentration > 300 mg/dL), urine should be recollected and analyzed for creatinine and target chemicals [33]. Therefore, samples with urine that was too dilute (n = 78) or too concentrated (n = 72) were excluded from this study. In addition, participants were excluded for excessive alcohol consumption, defined as drinking more than 30 g of alcohol/day for male (n = 179) and more than 20 g/day for female (n = 145) [34]. Ultimately, a total of 2238 participants were included in this study. More details of the participants being screened were shown in Fig. 1.

uGLY was defined as the independent variable. Urine samples from approximately one-third of participants aged 6 years and older were used to measure GLY concentrations (ng/ml) during the two survey cycles of NHANES 2013–2016 in total. Urine was collected at a mobile examination center (MEC) and subsequently aliquoted within a few hours and frozen in time for transport to the CDC’s National Center for Environmental Health (NCEH) for testing. A 200 µl urine sample was utilized, based on two-dimensional on-line ion chromatography with tandem mass spectrometry (IC-MS/MS) and isotope dilution quantification [35]. For analytes with analytical results below the lower limit of detection (LLOD), the analytic results were filled and placed using an estimated value, which was the LLOD divided by a square root of 2. The LLOD for uGLY was 0.2 ng/ml.

FLI was defined as the dependent variable, which had been validated as a valid tool for assessing NAFLD [21, 22]and was obtained by the following formula [36]:

Fatty Liver Index (FLI) = (e^{0.953×Ln (TG)} + 0.139×BMI + 0.718×Ln(GGT) + 0.053×WC − 15.745) ÷ (1 + e^{0.953×Ln (TG)} + 0.139×BMI + 0.718×Ln(GGT) + 0.053×WC − 15.745) × 100. TG, GGT, BMI and WC in the calculation formula represent triglycerides, glutamyl transpeptidase, body mass index and waist circumference, respectively. TG (mg/dl) and GGT (U/L) were derived from laboratory test information, and BMI (kg/m²) and WC (cm) were obtained from physical examination information.

In order to better demonstrate accurate independent effects between independent and dependent variables, we included covariates that could potentially affect uGLY and FLI score in the model for adjustment. Questionnaire information: age (years), gender, race/ethnicity, education level, ratio of family income to poverty (PIR), hypertension, diabetes, physical activity intensity, smoking, herbicide use, and fasting time (hours). PIR represents the ratio of family income to poverty, which is calculated using the Department of Health and Human Services (HHS) poverty guidelines as a measure. We also inducted information on dietary intake of some nutrients: energy (Kcal), protein (gm), sugar (gm), fat (gm), cholesterol (mg), alcohol (gm), and moisture (gm). All participants in NHANES 2013–2016 underwent two 24-hour dietary recalls, and we had the average of the two recall information for the final analysis [37, 38]. Laboratory test information was included for alanine aminotransferase (ALT) (U/L), aspartate aminotransferase (AST) (U/L), creatine phosphokinase (CPK) (U/L), albumin (ALB) (g/dl), globulin (GLB) (g/dl), total bilirubin (TBIL) (mg/dl), blood urea nitrogen (BUN) (mg/dl), uric acid (UA) (mg/dl), creatinine (Cr) (mg/dl), total cholesterol (TC) (mg/dl), and serum iron (ug/dl). Considering that inflammation is a major factor affecting FLI, but c-reactive protein (CRP) information was not included in NHANES 2013–2016, previous studies demonstrated a close association between systemic immune inflammatory index (SII) and liver disease [39, 40]. Therefore, we also calculated SII as a covariate with the information of lymphocyte count (LYM) (1000 cells/UL), neutrophil count (NEU) (1000 cells/UL) and platelet count (PLT) (1000 cells/UL), calculated as [41]: SII = NEU/LYM × PLT.

We analyzed all data in this study using R (http://​www.​R-project.​org) and EmpowerStats (http://​www.​empowerstats.​com), and p < 0.05 was determined to be statistically significant. NHANES employs complex multi-stage probability sampling, and the sampling weights provided should be used appropriately in the statistical analysis to make the obtained sample representative. According to the official NHANES recommendations, the uGLY sample was tested only in 1/3 of the sample and unique subweights (WTSSCH2Y) were provided to obtain the final weights included in this study from the sum of the weights from NHANES 2013–2014 and NHANES 2015–2016 cycles divided by 2 [42]. Continuous variables were presented in the form of median and quartiles, and continuous variables with no more than 10% missing values could be filled with the mean, otherwise they were converted to dichotomous variables using the median as a cut-off, and the missing values were set as a separate group. Categorical variables with more missing values were set as a separate group for inclusion in the analysis. The normality test revealed that the available uGLY data were skewed. Therefore, we transformed the log function of uGLY with the constant “e” as the base [Log_e(uGLY)] and used it for all the analyses in this study. Participants were categorized into two groups (< -1.011 ng/ml and ≥ -1.011 ng/ml) based on the median of Log_e(uGLY) level. For the comparison of all variables between the two groups, the data set was parsed and weights were obtained by using the survey design R package as recommended on the NHANES website, and p-values were calculated with weighted linear regression (continuous variables) and chi-square tests (categorical variables). We screened the selected covariates before building the final model [43]. First, a stepwise screening based on the variance inflation factor (VIF) removed covariates with too high covariance (VIF > 5) (Supplementary Table 1). Subsequently, the final included covariates (Supplementary Table 4) were identified based on the principle that the introduction of a covariate in the basic model or its exclusion from the full model would have an effect of > 10% on the regression coefficient of Log_e(uGLY) (Supplementary Table 2), or that the introduced covariate would have a statistical p-value < 0.1 on the regression coefficient of FLI (Supplementary Table 3). Multiple linear regression analysis was used to assess the association between Log_e(uGLY) and FLI, and by adjusting for different covariates we generated three models. Model 1: without adjusting for any covariates, Model 2: age, gender, race/ethnicity, education level, and PIR were adjusted, and Model 3: all covariates within Supplementary Table 4 were adjusted. To further demonstrate the association between different Log_e(uGLY) levels and FLI, we also built a model based on inverse treatment probability weighting (IPTW). Briefly, we constructed a logistic regression model with all screened covariates that could be used to determine the predicted probability of each independent participant being categorized into different Log_e(uGLY) levels (< -1.011 ng/ml and ≥ -1.011 ng/ml). The inverse of the difference between the actual probability and the predicted probability of each participant being categorized into different groups was used as weights in subsequent model. Based on the use of weights, the differences in covariates between the different Log_e(uGLY) groups could be balanced thereby more correctly assessing the association between different levels of Log_e(uGLY) and FLI. It is worth noting that the independent variables in the IPTW-based model were the groups (< -1.011 ng/ml and ≥ -1.011 ng/ml) after grouping at the median of Log_e(uGLY). Smooth curve fitting (penalized spline method) was used to evaluate whether there was a nonlinear association between Log_e(uGLY) and FLI, and a generalized additive model was further used to assess the threshold effect between the two. Threshold effect analyses aims to find a particular threshold of the independent variable and test for inconsistency in the association between the independent variable and the response variable before and after this threshold [44]. A likelihood ratio (LLR) of less than 0.05 is generally considered a criterion for the existence of a nonlinear association. Finally, we performed subgroup analysis to identify susceptible individuals for the association between Log_e(uGLY) and FLI and performed an interaction test.

A total of 2238 participants were eventually enrolled in the study. Participants with Log_e(uGLY) level ≥ -1.011 ng/ml had significantly higher FLI than those with Log_e(uGLY) level < -1.011 ng/ml, with a statistically significant difference between the two groups (P < 0.001). The baseline characteristics of the participants were shown in Table 1.

In all models (models 1–3), there was a positive association between Log_e(uGLY) level and FLI. The fully adjusted model (model 3) results suggested that each 1-unit increase in Log_e(uGLY) would be accompanied by a 2.16-unit increase in FLI (Beta coefficient = 2.16, 95% CI: 0.71, 3.61). Subsequently, we displayed Log_e(uGLY) in tertile 1 (-1.959- -1.190), Tertile 2 (-1.191- -1.030) and Tertile 3 (-1.031- 0.104). Compared to Tertile 1, the trend of positive association between independent and dependent variables was more significant as Log_e(uGLY) increased (P for trend < 0.001), and was most pronounced in Tertile 3 (Beta coefficient = 4.19, 95% CI: 1.49, 6.88). All results were presented in Table 2. In addition, we further validated the positive association between Log_e(uGLY) and FLI by employing IPTW (Beta coefficient = 2.07, 95% CI: 0.18, 3.96) (Supplementary Table 6). Information on baseline characteristics of participants after IPTW was displayed in Supplementary Table 5.

To verify the stability of the positive association between Log_e(uGLY) level and FLI in the cohort with characteristics, we performed a subgroup analysis. In the fully adjusted model, the results of the subgroup analysis showed that this positive association were more significant in participants who were female (Beta coefficient = 2.80, 95% CI: 0.74, 4.85), 40–60 years old (Beta coefficient = 3. 80, 95% CI: 0.71, 5.44), other races/ethnicities (Beta coefficient = 5.19, 95% CI: 2.29, 8.08), without hypertension (Beta coefficient = 1.84, 95% CI: -0.52, 4.20) and borderline diabetes (Beta coefficient = 2.21, 95% CI: -8.17, 12.59). In addition, we performed interaction tests for gender, age, race/ethnicity, hypertension, and diabetes to further commit the stability of the results in the subgroup analysis (P > 0.05 for interaction test). The results of the subgroup analysis were presented in Table 3.

Finally, we evaluated whether there was a nonlinear association between Log_e(uGLY) level and FLI with smooth curve fitting and threshold effect analysis. Depending on Fig. 2A; Table 4, a linear association (LLR = 0.364) was observed between independent and dependent variables. In addition, we also designed to verify whether this positive linear association was stable in each subgroup mentioned above, for which we performed smoothed curve fitting and threshold effect analysis for each subgroup respectively. The results suggested that the positive linear association between independent and dependent variables was durable for gender (Fig. 2B and Supplementary Table 7), age (Fig. 2C and Supplementary Table 8), and hypertension (Fig. 2E and Supplementary Table 9). However, when diabetes was the subgroup, a significant increase in FLI (Beta coefficient = 5.12, 95% CI: 0.69, 9.55) would occur when the Log_e(uGLY) of participants with diabetes exceeded − 1.55 (LLR = 0.007) (Table 5; Fig. 2F). In addition, the black race/ethnicity participants had an effect value of (Beta coefficient = 17.00, 95% CI: 2.74, 31.25) between Log_e(uGLY) and FLI when Ln(uGLY) exceeded − 0.12 (LLR = 0.026). Participants of other races/ethnicities showed a positive association with FLI only when Log_e(uGLY) was less than − 0.46 (Beta coefficient = 8.95, 95% CI: 4.74, 13.16) (LLR = 0.015), although when Log_e(uGLY) was more than − 0.46 it showed a negative association with FLI which was not statistically different (Beta coefficient=-5.26, 95% CI: -14.25, 3.72) (Table 6; Fig. 2D).

We excluded participants who did not have information about uGLY (n = 15,408) and FLI (n = 1286) explicitly, while retaining participants who were younger than 20 years of age, consumed large amounts of alcohol, were taking medications that interfere with fat metabolism, had viral hepatitis, had substandard urine samples, and were in renal weakness/failure to perform sensitivity analyses for exclusion criteria on the association between Log_e(uGLY) and FLI (Supplementary Fig. 1). The results demonstrated that the true association between Log_e(uGLY) and FLI would be severely affected when participants in Fig. 1 were not excluded (Beta coefficient = 0.01, 95% CI: -0.04, 0.05). The results of the sensitivity analysis to the exclusion criteria were displayed in Supplementary Table 10. In the validation results of the positive linear association between Log_e(uGLY) level and FLI, through Fig. 2A we found that there were outliers in Log_e(uGLY), for which we examined the data distribution of Log_e(uGLY) (Supplementary Table 11). Through the results we could observe that the sample size of Log_e(uGLY) over 1 is 26, which accounted for 1.162% of the total sample size, such a result belonged to small sample events (< 5%). For this reason, we removed these outliers and re-examined the association between Log_e(uGLY) level and FLI. Based on the results, we found that after removing the outliers, the supplemental results were absolutely consistent with our results above. The supplemental results were shown in Supplementary Tables 12–20, Fig. 3A-F.