The literature search followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 Statement. Six electronic databases (PubMed, Embase, Cochrane Library, SinoMed, CNKI, and Wanfang Data) were systematically searched, supplemented by manual search of gray literature including dissertations and conference proceedings. Original research papers investigating the association between HCC risk and tea intake, published before December 15, 2025, were eligible for inclusion. The full search strategies for all databases are provided in Supplementary Table 1. The full search strategy for PubMed was as follows: (“Liver Neoplasms”[Mesh] OR “Hepatocellular Carcinoma”[Mesh]) AND (“Tea”[Mesh] OR “Tea, Green”[Mesh] OR “Tea, Black”[Mesh]) OR (liver cancer OR hepatoma OR hepatic neoplasm) AND (tea OR green tea OR black tea OR tea polyphenol OR catechin). Similar strategies were adapted for other databases. This review was registered in Prospective Register of Systematic Reviews (PROSPERO) (CRD420261343011).

Studies were eligible for inclusion if they met all the following criteria: (1) published in English or Chinese; (2) publicly available population-based cohort studies or case-control studies investigating the association between tea consumption and HCC; (3) HCC cases in the study were confirmed by pathological examination or imaging diagnosis, or through validated cancer registries/International Classification of Diseases (ICD) coding systems with established diagnostic accuracy; (4) original studies provided data on the association between tea intake and HCC risk, including odds ratio (OR) or hazard ratio (HR) with corresponding 95% confidence interval (CI), or sufficient data to calculate the above effect estimates; (5) only studies with HCC as the target outcome were included. Studies were excluded based on the following criteria: (1) reviews, meta-analyses, conference abstracts and other types of studies without original primary data; (2) non-population-based studies including animal experiments and in vitro cell experiments; (3) studies with a sample size of less than 10; (4) studies with incomplete or erroneous data that cannot be corrected or supplemented; (5) for multiple publications from the same research team based on the same study population, only the study with the largest sample size and longest follow-up duration was included to avoid duplicate data bias.

Literature screening was performed independently by two trained researchers (Ling ST, Xie LZ) according to the pre-specified inclusion and exclusion criteria. Irrelevant records were first excluded based on title and abstract screening. The remaining potentially eligible records were subsequently evaluated via full-text review to determine the final inclusion. Disagreements between the two researchers were resolved by discussion; if no consensus was reached, a third senior researcher was consulted for final adjudication.

Two researchers independently assessed the methodological quality of the included cohort studies and case-control studies using the Newcastle-Ottawa Scale (NOS). Detailed evaluation criteria are provided in Supplementary Table 2. The NOS evaluates studies from three dimensions: selection of study participants, comparability between study groups, and ascertainment of exposure or outcome, with a maximum score of 9. An adapted version of the NOS was used, allowing 0.5-point increments for individual criteria to better discriminate between studies that partially met specific criteria; this adaptation was pre-specified in the study protocol and applied consistently by two independent reviewers. Only studies with a total score of ≥ 6 were included; studies were categorized as high (H: ≥ 8), medium (M: 7.5), or low (L: 6 to < 7.5) quality based on NOS scores. The average of the scores assigned by the two independent reviewers was adopted as the final score, rounded to one decimal place. If the score difference between the two reviewers was ≥ 1 point, a third researcher was invited for arbitration.

Data extraction was similarly conducted independently by two researchers. The extracted information included: first author, publication year, country/region, study design, sample size, participant baseline characteristics, tea type and exposure assessment, definition of outcome events, effect estimates (HR or OR with 95%CI), and adjustment for confounders. For studies reporting multiple effect estimates, the estimate fromthe most fully adjusted model was selected for the primary meta-analysis. Whermultiple exposure contrasts were available, the comparison of the highestversus the lowest dose category was used. If several follow-up durations were reported, tihe estimate with the longest follow-up was selected. The detailed data extraction form is shown in Supplementary Table 3. Disagreements during data extraction were resolved using the same procedure as for literature screening.

All statistical analyses were conducted using the metafor package (v4.8-0) in R (v4.5.3). Effect estimates were extracted separately for cohort studies and case-control studies. For cohort studies, we extracted HR or risk ratios (RR) as reported; for case-control studies, we used OR. No overall pooled estimate combining both study designs was calculated. Instead, subgroup analyses were performed in cohort studies.

Heterogeneity was evaluated using the Q test and I² statistic. Given expected between-study heterogeneity and the relatively small number of studies (≤ 10 per analysis), a random-effects model was applied for all pooled analyses. The Hartung-Knapp-Sidik-Jonkman (HKSJ) method was used as the primary approach, as the DerSimonian-Laird (DL) method often yields inflated type I error rates when the number of studies is small^[19].

To explore potential effect modifiers and sources of heterogeneity, we performed subgroup analyses and univariate meta-regression for cohort studies stratified by predefined covariates: study quality, tea type, geographic region, sex, and dose category (the latter based on within-study tertiles without cross-study standardization). Given the limited number of studies (k = 8), these analyses were exploratory and underpowered; no multivariable meta-regression was performed. Random-effects models with the Hartung-Knapp adjustment were used for all analyses; meta-regression additionally estimated the proportion of between-study variance explained (R²) by each covariate.

Categorical trend analysis was performed for cohort studies using a within-study categorization approach. Tea consumption levels in each study were classified into three ordered categories (low, medium, high) based on the reported dose groups, with the lowest category as the reference. For studies reporting three or more dose groups, effect estimates for the low, medium, and high groups were extracted. For studies reporting only two dose levels, only the low and high groups were extracted. Pooled effect estimates for each dose level were then calculated using random-effects models. To evaluate the trend across exposure categories, ordinal scores (1, 2, 3) were assigned to low, medium, and high levels, and meta-regression was performed. P-value for the slope was reported as the trend P-value.

To examine whether control for key confounders affected the overall estimates, subgroup analyses were stratified by whether individual studies had adjusted for each of the following predefined covariates: body mass index (BMI), coffee consumption, diabetes, education, family history of liver cancer, HBV infection, HCV infection, income, liver disease, physical activity, and sex. For each covariate, studies were classified into “Adjusted” (the factor included in multivariable adjustment) and “Not adjusted” (the factor not adjusted for). Pooled effect sizes (HR) with 95%CIs were calculated using random-effects models. When only one study was available in a subgroup, the estimate was presented descriptively. Interaction between subgroups was tested using meta-regression (P < 0.05 indicating significant modification). This analysis was restricted to cohort studies due to the limited number of case-control studies.

Sensitivity analysis was performed using the leave-one-out method. Publication bias was evaluated by funnel plot, Egger’s regression test, and Begg’s rank correlation test (P < 0.05 indicating significant publication bias).

A total of 2,868 records were retrieved. After removing 920 duplicates, 1,948 records remained for title and abstract screening. After excluding 1,880 ineligible records, 68 underwent full-text review, of which 42 were excluded. After further excluding 4 studies with overlapping populations and 7 low-quality studies, 15 studies were finally included [Figure 1].

A total of 15 studies (8 cohort studies and 7 case-control studies) were included in this meta-analysis. Among them, 14 were peer-reviewed articles published in academic journals, and one was a doctoral dissertation from Fudan University^[20]. Geographically, 13 studies were conducted in Asia (9 in China, 4 in Japan), and 2 were carried out in Europe (1 in Italy, 1 multinational study). The total sample size was 1,211,744 participants, including 5,208 HCC cases, with individual study sample sizes ranging from 200 to 482,292.

Regarding exposure factors, among the 15 included studies, 10 assessed green tea consumption, 4 assessed unspecified tea types, 4 assessed mixed tea, 3 assessed black tea, 2 assessed other tea types (including oolong or herbal tea), and 1 assessed jasmine tea. Since several studies reported on more than one tea type, the sum of these counts exceeds the total number of studies. For subgroup analyses stratified by tea type, each study was classified according to its primary tea type of interest as defined in the original study, ensuring that each study contributed only one effect estimate to a given subgroup analysis to preserve statistical independence. Exposure assessment was performed using self-administered Food Frequency Questionnaires (FFQs) in seven studies, face-to-face interviews in seven studies, and urinary biomarker measurement in one study. For liver cancer diagnosis, 6 studies adopted pathological examination or combined imaging and clinical diagnosis as the confirmation criteria, and 9 studies determined outcomes based on cancer registration systems or ICD coding. In these 9 studies, the registry- or ICD-based outcome ascertainment was derived from well-validated cancer registries or hospital record systems, where HCC diagnoses were originally confirmed by pathological examination or imaging in clinical practice. All included studies adjusted for confounding factors including smoking and alcohol consumption. Eleven studies further adjusted for sex, seven for education, and twelve for HBV or HCV infection status.

For methodological quality assessment, based on the NOS, the average score of included cohort studies was 8.0, and the average score of case-control studies was 7.2, indicating high overall quality of the included studies. The baseline characteristics of all included studies are detailed in Table 1.

Using the HKSJ method, cohort studies showed no statistically significant association between tea consumption and HCC risk, with a pooled HR of 0.93 (95%CI: 0.69-1.25). Heterogeneity among cohort studies was substantial (I² = 63.7%, τ² = 0.072, P = 0.0074); therefore, a random-effects model was used [Figure 2]. Given the substantial heterogeneity (I² = 91.9%) and very wide CI, the case-control results should be interpreted with extreme caution and are presented here as a sensitivity analysis only, not as primary evidence [Supplementary Figure 1]. Due to the substantial differences in study design and effect estimates, no overall pooled estimate combining both designs was calculated.

Given the substantial heterogeneity observed in the overall meta-analysis (cohort studies: I² = 63.7%; case-control studies: I² = 91.9%), subgroup analyses were performed for cohort studies stratified by study quality, tea type, geographic region, sex, and dose level [Table 2 and Supplementary Figures 2-6]. In most subgroups, no statistically significant association was observed (P > 0.05). However, the European subgroup, which included only a single study (Bamia et al.), showed a potential protective effect (HR = 0.41, 95%CI: 0.22-0.77, P = 0.0058)^[18]. Given that this finding is derived from only one study and has a very wide CI, it should be interpreted as exploratory rather than confirmatory. This finding requires replication in larger, multi-center European cohorts. Heterogeneity remained high in most subgroups (I² range: 51.8%-92.8%). The Baujat plot [Supplementary Figure 7] further explored individual study contributions to overall heterogeneity.

Univariate meta-regression analyses were further performed to formally test whether the five prespecified covariates explained the heterogeneity among cohort studies [Table 3]. None of the examined factors - geographic region, NOS score, tea type, sex, or dose level - was significantly associated with the effect estimates (all P > 0.05). The R² values were 0 for most covariates, indicating that these factors individually accounted for little to none of the observed between-study variance.

Sensitivity analysis using the leave-one-out method showed that after excluding each study individually, the pooled HR for cohort studies varied from 0.88 (95%CI: 0.66-1.17) to 1.00 (95%CI: 0.78-1.29) [Supplementary Table 4]. None of these recalculated pooled estimates reached statistical significance (all P > 0.05), confirming the robustness of the null association.

We further examined whether adjustment for specific confounders influenced the association between tea consumption and HCC risk. Studies were stratified by whether they adjusted for 11 pre-specified covariates: BMI, coffee consumption, diabetes, education, family history of liver cancer, HBV infection, HCV infection, income, liver disease, physical activity, and sex. Detailed adjustment factors for each included study are presented in Supplementary Table 5 and Supplementary Figures 8-16. For each factor, pooled effect estimates did not differ significantly between the “adjusted” and “not adjusted” subgroups (all interaction P > 0.05; range: 0.114-0.930). For covariates where only one study was available in a subgroup (BMI not adjusted and HCV adjusted), pooled estimates were not calculable (Supplementary Table 6, composite forest plot in Supplementary Figure 17). However, these subgroup comparisons should be interpreted with caution, as several subgroups contained only one or two studies, limiting statistical power.

Publication bias was assessed using funnel plots, Egger’s regression test, and Begg’s rank correlation test. Funnel plots [Figure 3] showed a generally symmetric distribution of effect sizes. Neither Egger’s test (P = 0.923) nor Begg’s test (P = 0.720) indicated significant publication bias.