The NHANES represents an ongoing, cross-sectional, and nationally representative investigation conducted in the United States. This significant research initiative is managed by the National Center for Health Statistics (NCHS) and receives authorization and funding from the Centers for Disease Control and Prevention (CDC) with the aim of evaluating the health and nutritional conditions of the civilian population in the United States who are not institutionalized. Data collection occurs biennially through a sophisticated multistage probability sampling methodology, which encompasses in-home, face-to-face interviews followed by comprehensive physical assessments at the Mobile Examination Center (MEC), during which blood and urine specimens are obtained. The NHANES protocol undergoes scrutiny and receives endorsement from the NCHS Research Ethics Review Board, ensuring that informed consent has been acquired from all participants involved. In our study, we utilized data from 10 cycles (1999–2018) of the NHANES database, which contains comprehensive test data on MASLD-HCC for participants. To account for inter-cycle variability, we applied survey weights provided in the NHANES documentation. Missing data were handled using multiple imputation methods. TPA level in plasma were measured by GC/MS as described in NHANES trans fatty acids procedure. The specific procedures included: converting fatty acids into their free forms for extraction, derivatization with pentafluorobenzyl bromide, separation via capillary gas chromatography, and detection by negative chemical ionization mass spectrometry. Fatty acids were identified by comparing retention times with standards and specific mass-to-charge ratios, while quantitative analysis was performed using a stable isotope-labeled internal standard method.

The initial stage of MASLD is considered to be characterized by abnormal hepatic fat accumulation. Due to the absence of transient elastography data for liver assessment in the 1998–2018 NHANES database cycle, this study employs the Fatty Liver Index (FLI) as a surrogate assessment indicator. The FLI is calculated using the formula:

TG denotes total triglycerides, BMI refers to body mass index, GGT represents gamma-glutamyl transferase, and WC stands for waist circumference (28–32). After excluding other liver diseases associated with the aforementioned etiological factors, participants with a FLI > 60 were diagnosed with MASLD.

From the initial cohort of 18,017 participants, exclusions were made for the following reasons: (1) removal of samples lacking FLI data (N = 3,741), (2) exclusion of samples without MASLD (N = 11,035), and (3) exclusion of samples lacking TPA (N = 2,693). The final study population comprised 548 participants, including 239 cases of MASLD-HCC and 309 MASLD samples (Figure 1).

Propensity score matching is a statistical technique utilized to address data from observational studies. In observational studies, data bias and the presence of confounding variables are common due to various reasons. The method of PSM is designed to reduce the impact of these biases and confounding factors, thereby enabling a more rational comparison between the experimental group and the control group. This method was first introduced by Rosenbaum and Rubin (33) and is frequently employed in fields such as medicine, public health, and economics to match baseline characteristics (34). In our study, we implemented a 1:1 nearest-neighbor matching algorithm to ensure balanced group comparisons.

Weighted binary logistic regression was utilized to investigate explore the possible association between levels of TPA and the occurrence of MASLD-HCC. In the regression model, TPA levels were incorporated as both continuous and categorical variables. This methodological framework enabled the estimation of odds ratios (ORs) along with their respective 95% confidence intervals (95% CIs). The TPA concentrations were first treated as a continuous variable. Subsequently, these concentrations were categorized into quartiles according to the interquartile range, organized from the lowest to the highest levels. These quartiles were designated as the first quartile (Q1), second quartile (Q2), third quartile (Q3), and fourth quartile (Q4), and were also assessed as categorical variables. In cases where TPA levels were analyzed as categorical variables, the lowest quartile (Q1) served as the reference group. The study delineated three distinct statistical analysis groups: model 1 constituted the unadjusted model, whereas model 2 incorporated adjustments for demographic factors including gender, age, race, education, and poverty-income ratio (PIR). Building on the adjustments made in model 2, model 3 additionally accounted for factors such as diabetes, lipid profiles, cholesterol levels, low-density lipoprotein, high-density lipoprotein, albumin, creatinine, blood urea nitrogen, hemoglobin, platelet count, and glycated hemoglobin. All regression analyses were conducted with the inclusion of survey weights, and non-normally distributed continuous covariates underwent transformation utilizing weighted quartiles. Furthermore, the study evaluated potential interactions among factors associated with TPA levels through the application of the variance inflation factor (VIF), where a VIF value of less than 10 indicated no interaction between the examined factor and other variables. The P values were adjusted to account for the false discovery rate (FDR).

Forest plots are primarily used for visualizing the comparison of effect sizes and CIs across multiple study results. They are commonly employed to summarize and compare outcomes from different studies, providing a more intuitive representation of the effect size (such as RR, OR, HR, or WMD) and their corresponding 95% CIs. This visualization aids in better understanding the consistency and discrepancies between different studies.

Restricted cubic spline is a concept in statistics, particularly frequently used in regression analysis and curve fitting. It is a method for fitting and modeling continuous variables by dividing the data range into several intervals and using a cubic polynomial for fitting within each interval to create a smooth curve. These polynomials are smoothly connected across adjacent intervals, often with additional smoothness constraints to avoid sharp fluctuations in the curve. In statistical modeling, RCS is commonly used to model the relationship between continuous variables and the dependent variable. In regression analysis, it allows for capturing nonlinear relationships while maintaining smoothness and avoiding overfitting. In this study, the RCS analysis was performed using the R “rms” package, with three knots automatically selected as the optimal number based on statistical algorithms.

In the realm of statistics, the multilevel logistic regression model serves as an advanced classification technique that extends traditional logistic regression to accommodate multicategory scenarios. This model is particularly designed to forecast the probabilities associated with various potential outcomes of a dependent variable that has a categorical distribution. Within the framework of a multilevel logistic regression model, the dependent variable’s predictions are made utilizing a series of independent variables, also known as features or observed variables. This model operates by computing the likelihood of achieving a specific outcome in the dependent variable through a linear combination of the independent variables along with their respective parameters, all framed within a probabilistic model. The parameters linked to the independent variables are derived from the training dataset, and they are commonly referred to as regression coefficients. Importantly, the multilevel logistic regression model empowers researchers to account for the inherent clustered structure of the data, explore the origins of variation both within and between clusters, identify which variables account for individual differences, and ascertain which factors influence variations at the cluster level.

All data processing and analytical procedures were performed utilizing R software (version 4.4.0). For the analysis of baseline characteristics, medians and interquartile ranges were utilized to characterize continuous variables that exhibited non-normal distributions. Classification variables were reported as sample counts and weighted percentages. To check for changes in variable characteristics between groups of TPA (quartiles), we used Wilcoxon rank sum tests to test continuous variables and Rao-Scott Chi-square tests to test weighted percentages of categorical variables, providing a comprehensive description of the entire population. The “mediation” package was used for mediation analysis to assess the mediating effects of lipid indicators (including cholesterol, low density lipoprotein, high density lipoprotein, and triglycerides) on the association between TPA and MASLD-HCC. The existence of mediation effects was characterized by the presence of a notable indirect effect, a significant overall effect, and a positive ratio of the mediation effect. All statistical evaluations were conducted as two-tailed tests, with a P-value of less than 0.05 deemed to indicate statistical significance. For all hypothesis tests (including primary, subgroup, and mediation analyses), Benjamini–Hochberg method was applied to adjust P values for FDR.

In this study, a final cohort of 548 patients was incorporated to evaluate the baseline characteristics, including factors such as gender, age, and race, in connection with the occurrence of HCC. The baseline characteristics of participants were presented in Table 1. Before PSM, education and smoking showed differences between the HCC and non-HCC groups (P-value = 0.038, P-value = 0.039), whereas no differences were observed after PSM (P-value = 0.1, P-value = 0.87). Some other baseline data also showed differences before and after PSM in both the HCC and non-HCC groups, such as race (P-value < 0.001 before PSM, P-value = 0.026 after PSM), BMI (P-value = 0.003 before PSM, P-value = 0.002 after PSM), diabetes (P-value = 0.013 before PSM, P-value = 0.027 after PSM), total cholesterol (P-value < 0.001 before PSM, P-value < 0.001 after PSM), triglycerides (P-value = 0.015 before PSM, P-value = 0.039 after PSM), and low-density lipoprotein (P-value = 0.001 before PSM, P-value < 0.001 after PSM). These indicators, which showed differences before and after PSM in both the HCC and non-HCC groups, could be considered for mediation effect analysis. Some baseline indicators, however, showed no differences before and after PSM in both the HCC and non-HCC groups, such as gender, age, education, and PIR. This indicates that these indicators were relatively balanced between the HCC and non-HCC groups, and their confounding factors were excluded when comparing the two groups.

We utilized box plots to illustrate the relationship between MASLD-HCC and TPA, revealing a significant difference in TPA levels between the cancer and non-cancer groups (Figure 2). To further analyze this difference, we utilized linear regression analysis and multilevel logistic regression model to demonstrate the relationship (Table 2). Model 1, without covariates, revealed a positive correlation between TPA levels and MASLD-HCC (OR = 1.524, 95% CI = 1.397–1.677, P-value < 0.001); when compared to Q1, Q2 had an OR of 3.671 (95% CI = 1.997–6.987, P-value < 0.001), Q3 had an OR of 7.519 (95% CI = 4.132–14.258, P-value < 0.001), and Q4 had an OR of 42.905 (95% CI = 20.946–94.294, P-value < 0.001), indicating that with each increase in TPA, the probability of developing MASLD-HCC increased. Model 2, with gender, age, race, education, and PIR as covariates, and model 3, further adding diabetes, blood lipids, cholesterol, low-density lipoprotein, high-density lipoprotein, albumin, creatinine, blood urea nitrogen, hemoglobin, platelet count, and glycated hemoglobin as covariates, both showed a similar positive correlation between TPA and MASLD-HCC (model 2 OR = 1.654, 95% CI = 1.487–1.862, P-value < 0.001; model 3 OR = 2.276, 95% CI = 1.921–2.767, P-value < 0.001). When grouped by quartile intervals, both model 2 and model 3 revealed that, compared to Q1, the OR values for Q2 (model 2 OR = 6.170, 95% CI = 2.983–13.500, P-value < 0.001; model 3 OR = 9.500, 95% CI = 4.069–23.801, P-value < 0.001), Q3 (model 2 OR = 15.239, 95% CI = 7.183–34.719, P-value < 0.001; model 3 OR = 26.202, 95% CI = 10.721–70.020, P-value < 0.001), and Q4 (model 2 OR = 110.166, 95% CI = 44.800–296.784, P-value < 0.001; model 3 OR = 753.702, 95% CI = 196.665–3,490.227, P-value < 0.001) increased progressively, indicating that with the addition of covariates, the probability of developing MASLD-HCC also increased with increasing TPA.

Restricted cubic spline curves were utilized to explore the association between TPA levels and the incidence of MASLD-HCC, while accounting for all pertinent covariates. A statistically significant nonlinear relationship was observed between TPA and MASLD-HCC (P non-linear < 0.001, cutoff value = 5.85 μmol/L, Figure 3A). Furthermore, a significant nonlinear relationship was found between TPA and the incidence of MASLD-HCC in the male subgroup (P non-linear < 0.001, cutoff value = 5.92 μmol/L, Figure 3B), as well as in the female subgroup (P non-linear < 0.001, cutoff value = 5.83 μmol/L, Figure 3C).

Figure 4 illustrates the association between TPA and MASLD-HCC, with the relationship analyzed across various subgroups stratified by gender, age, PIR, hypertension, and diabetes using fully adjusted multivariate logistic regression. We found that in both male and female gender subgroups, the incidence of HCC increased with increasing levels of TPA (male group: Q2 vs. Q1, OR = 4.10, 95% CI = 1.61–10.43, P-value = 0.003; Q3 vs. Q1, OR = 7.80, 95% CI = 2.95–20.64, P-value < 0.001; Q4 vs. Q1, OR = 89.27, 95% CI = 18.20–437.98, P-value < 0.001; female group: Q2 vs. Q1, OR = 10.92, 95% CI = 2.79–42.63, P-value = 0.001; Q3 vs. Q1, OR = 24.35, 95% CI = 5.99–98.90, P-value < 0.001; Q4 vs. Q1, OR = 1,148.83, 95% CI = 152.86–8,634.14, P-value < 0.001). The age young subgroup appeared to be more sensitive to TPA, with increasing OR values for Q2, Q3, and Q4 compared to Q1 (Q2 vs. Q1, OR = 6.84, 95% CI = 2.91–16.10, P-value < 0.001; Q3 vs. Q1, OR = 13.76, 95% CI = 5.64–33.54, P-value < 0.001; Q4 vs. Q1, OR = 612.11, 95% CI = 130.62–2,868.48, P-value < 0.001), while the tendency was observed in the age old subgroup only when comparing Q3 and Q4 to Q1 (Q3 vs. Q1, OR = 5.96, 95% CI = 1.16–30.71, P-value = 0.033; Q4 vs. Q1, OR = 15.17, 95% CI = 1.63–141.21, P-value = 0.017). A similar trend was observed in the PIR subgroup, with significant differences in Q2, Q3, and Q4 compared to Q1 in the high PIR group (Q2 vs. Q1, OR = 7.99, 95% CI = 3.33–19.16, P-value < 0.001; Q3 vs. Q1, OR = 18.29, 95% CI = 7.28–45.94, P-value < 0.001; Q4 vs. Q1, OR = 437.48, 95% CI = 102.84–1,860.99, P-value < 0.001), and in Q3 and Q4 compared to Q1 in the low PIR group (Q3 vs. Q1, OR = 10.00, 95% CI = 1.13–88.71, P-value = 0.039; Q4 vs. Q1, OR = 913.50, 95% CI = 13.51–61,760.65, P-value = 0.002). In the hypertension subgroup, an increase in the incidence of HCC was observed with increasing levels of TPA, regardless of whether the individual had hypertension. In the diabetes subgroup, patients without diabetes appeared to be more sensitive to TPA levels (Q2 vs. Q1, OR = 6.10, 95% CI = 2.74–13.59, P-value < 0.001; Q3 vs. Q1, OR = 13.63, 95% CI = 5.91–31.43, P-value < 0.001; Q4 vs. Q1, OR = 257.14, 95% CI = 67.62–977.92, P-value < 0.001).

Table 3 presents the associations of TPA with metabolic indicators following multivariate logistic regression. Upon controlling for all potential confounding variables, a positive correlation was observed between TPA and cholesterol levels (β = 0.644, 95% CI = 0.474–0.819, P-value < 0.001), triglycerides (β = 1.236, 95% CI = 0.994–1.493, P-value < 0.001), low-density lipoprotein (β = 0.557, 95% CI = 0.372–0.749, P-value < 0.001), albumin (β = 0.138, 95% CI = 0.08–0.198, P-value < 0.001), hemoglobin (β = 0.148, 95% CI = 0.009–0.291, P-value = 0.036), and platelet count (β = 0.004, 95% CI = 0–0.007, P-value = 0.007). It was negatively correlated with BMI (β = −0.037, 95% CI = −0.068–0.006, P-value = 0.021).

Building upon the results of the analysis of metabolic indicators, we conducted further mediation analysis to investigate the role of these indicators in the association between TPA and MASLD-HCC. The study revealed that in the mediation by triglycerides, the direct effect of TPA on HCC was 0.08, the indirect effect was 0.04, and the total effect was 0.12. Triglycerides mediated 39.18% of the association between TPA and the incidence of HCC (Figure 5). In the case of albumin, the direct effect was 0.12, the indirect effect was 0.01, and the total effect was 0.13, with a mediating proportion of 7.79%. Additionally, we evaluated the mediating roles of various metabolic parameters, including cholesterol and low-density lipoprotein, in the association (Supplementary Table 1).