Cruciferous vegetables intake and risk of colon cancer: a dose–response meta-analysis

Abstract

Background

Colon cancer (CC) is the third most diagnosed malignancy and second leading cause of cancer mortality globally, with ~ 1.9 million new cases and 903,859 deaths annually (Bray et al. in CA Cancer J Clin 68(6):394–424, 2018). Diet represents a key modifiable risk factor for CC pathogenesis (Herr and Buchler in Cancer Treat Rev 36:377–383, 2010). Cruciferous vegetables (CV)—rich in glucosinolates that hydrolyze into bioactive isothiocyanates (Willett in Cancer Epidem Biomar 10:3–8, 2001; Murillo and Mehta in Nutr Cancer. 41(1–2):17–28, 2001; Higdon et al. in Pharmacol Res 55:224–36, 2007)—exhibit chemopreventive properties through carcinogen detoxification, apoptosis induction, and cell cycle arrest (Zhang et al. in Proc Nutr Soc 65:68–75, 2006). While prior meta-analyses report an inverse association between CV intake and CC risk (Tse and Eslick in Nutr Cancer 66(1):128–39, 2014), the quantitative dose–response relationship remains uncharacterized, limiting translational insights for dietary guidance.

Methods

A thorough search of the literature was conducted in Embase, Scopus,Web of Science, PubMed, and Cochrane Library from inception to June 28, 2025, using a predetermined strategy encompassing both cohort and case–control studies. Two independent reviewers selected studies based on predefined inclusion criteria, with discrepancies resolved by consensus or senior investigator adjudication. Statistical analyses were performed using Stata (version 14.2). Subgroup analyses accounted for study design, geographic location, and potential confounders. Publication bias was assessed using Egger's test, the LFK index, and the trim-and-fill method. Sensitivity analyses employed the leave-one-out approach. The dose–response relationship was evaluated using restricted cubic spline models.

Results

Data from 17 research—including 7 cohort studies and 10 case–control studies—with 97,595 patients were methodically combined in this investigation.Consumption of CV was found to be inversely correlated with CC risk (odds ratios [OR] = 0.8; 95% confidence interval [CI] 0.72–0.90) in the pooled analysis using a random-effects model. Furthermore, a progressive decrease in risk was shown by the non-linear dose–response analysis as consumption levels increased.

Conclusion

This meta-analysis suggests a potential inverse association between higher CV intake and CC incidence. However, these findings should be interpreted cautiously due to methodological limitations, including heterogeneity in study designs, dietary assessment methods and potential residual confounding.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12876-025-04163-9.

Introduction

Colon cancer (CC) imposes a substantial global burden, with approximately 903,859 annual fatalities and 1.9 million newly diagnosed cases worldwide. These epidemiological figures establish CC as the second leading cause of cancer-associated mortality and the third most frequently diagnosed malignancy across all human neoplasms [1]. Diet is one of the main environmental variables that contribute to CC, according to epidemiological research [2]. The possible preventative benefits of CV have attracted a lot of attention in cancer research [3]. These vegetables, which include cabbage, broccoli, and Brussels sprouts, are rich in phytochemicals such as flavonoids, fiber, vitamin C, and carotenoids, which may contribute to cancer prevention [4]. CV are unique in their ability to produce glucosinolates [5]. During plant cell disruption, the enzyme myrosinase catalyzes the hydrolysis of glucosinolates, yielding bioactive compounds such as isothiocyanates. These compounds have protective effects by reducing angiogenesis, detoxifying carcinogens, inducing apoptosis, stopping the cell cycle, and inhibiting carcinogen-activating enzymes [6]. The association between CV intake and CC has been studied in previous meta-analyses, but the dose–response relationship has not been fully investigated [7]. To strengthen the evidence base, we incorporated a dose–response meta-analysis and updated the literature search to include studies published through June 2025.This study employs a meta-analytic approach to systematically synthesize existing literature on the association between CV consumption and CC risk. Additionally, we assess the dose–response connection between consumption levels and the incidence of CC in order to shed more light on the possible risk-modifying effects of CV eating patterns.The findings are meant to provide evidence-based suggestions for enhancing clinical CC prevention and treatment strategies.

Materials and methods

Search strategy

We present this article in accordance with the PRISMA 2020 checklist [8]. Four databases (PubMed,Scopus, Embase, Web of Science, Cochrane Library) were searched from inception to June 28, 2025. Medical Subject Headings (MeSH) and Free-text phrases were combined in the search technique, incorporating key words such as "Cruciferous Vegetable," "Colonic Neoplasms" [MeSH], "Colon Cancer," and related synonyms. All dentified citations were systematically imported into EndNote X9 (Clarivate Analytics) for duplication management. The process included two phases:1.Automated eduplication: Using software's built-in algorithm prioritizing DOI, PMID, publishing dates and source journals;2.Manual verification: For potential duplicate publications of the same study population, we implemented a 2-step protocol:a)Cross-checked author lists, institutions, recruitment periods and sample sizes;b) Compared intervention protocols and outcome measurement timelines. The selection of the study was carried out independently by two investigators. In order to settle disagreements, a senior investigator was consulted when necessary or by consensus before conclusions were made.We registered in the PROSPERO database (CRD42025638177).

Inclusion and exclusion criteria

This meta-analysis included observational studies meeting the following criteria:

(P) Population: Adults (≥ 18 years) without colon cancer at baseline (cohort studies) or diagnosed cases with matched controls (case–control studies), excluding pediatric populations and hereditary cancer syndromes.

(E) Exposure: Quantified dietary intake of cruciferous vegetables (e.g., broccoli, cauliflower) assessed through any method (e.g., FFQ, 24-h recall).

(C) Comparator: Lower intake or non-consumption groups.

(O) Outcomes: Incident colon cancer confirmed via medical records, pathology, registries, or validated self-report, with reported odds ratios (OR) or relative risks (RR) and 95% confidence intervals (CI).

(S) Study design: Prospective/retrospective cohort or case–control studies.

Exclusion criteria comprised: animal/cell line research; case reports, reviews, or commentaries; ecological studies; duplicate publications; studies lacking extractable effect estimates (OR/RR with 95% CI); and investigations of isolated phytochemicals without whole-food exposure.

Data extraction and quality assessment

Data was separately gathered by two investigators (Z.L. and J.L.) using a pre-piloted, standardized data collecting form. The primary content of data extraction includes: year of publication, the study type, first author's name, quality score, CV intake, region, number of participants, number of patients, OR or RR with their 95% CI, and potential confounding factors considered or adjusted for. Any discrepancies arising during the data extraction phase were resolved through iterative consensus discussions between the reviewers. In cases where consensus could not be reached, a third independent senior investigator (B.L.) was consulted for final arbitration.

For each included study, the most fully adjusted effect estimate (i.e., the estimate adjusting for the greatest number of relevant confounders) was preferentially extracted. When multiple adjusted models were reported, we selected the model controlling for at least the pre-specified core confounders: age, sex, smoking status, body mass index (BMI), and family history of colon cancer.

Methodological quality of included studies was assessed using the Newcastle–Ottawa Scale (NOS), which evaluates three domains: selection of study groups (0–4 points), comparability of groups (0–2 points), and outcome/exposure assessment (0–3 points). Studies were categorized as high (7–9), moderate (4–6), or low (0–3) quality based on total scores [9].

Statistical analysis

In case–control studies, odds ratios (ORs) were used as the primary effect measure. When combining results from both cohort and case–control studies in the meta-analysis, ORs were adopted as the common effect size metric for consistency across study designs [10]. All meta-analyses were performed using generic inverse-variance methods with ORs as the effect measure, reported with 95% confidence intervals. Model selection was primarily determined by anticipated methodological heterogeneity arising from mixed study designs (cohort vs. case–control) and variable exposure assessment methods, warranting pre-specified application of random-effects models as the default approach. The magnitude of statistical heterogeneity was quantified using I² statistics at α = 0.05 significance [11], with values interpreted as: < 30% (low), 30–60% (moderate), and > 60% (substantial). To validate robustness, sensitivity analyses included: (1) parallel fixed-effects modeling for all outcomes, (2) leave-one-out meta-analysis evaluating individual study influence, and (3) subgroup stratification by study design when I² exceeded 50% [12, 13].

Between-study variance (τ²) was estimated using the restricted maximum likelihood (REML) estimator. Heterogeneity magnitude was quantified via I² statistics derived from τ² values.The study design, geographic location, and potential confounding variables, such as smoking status, total caloric intake, and family history of CC, were taken into consideration while conducting subgroup analyses.Publication bias was assessed through a hierarchical protocol. Quantitative asymmetry was evaluated using Egger’s regression test (p < 0.10 threshold) [14] and the LFK index (Doi method;|LFK|> 1 indicates minor bias,|LFK|> 2 indicates major bias) [15]. The trim-and-fill method was applied to impute missing studies [16]. Sensitivity analyses were conducted via leave-one-out analysis by sequentially excluding individual studies to assess their influence on pooled estimates.

The dose–response connection between CV consumption and CC risk was examined using restricted cubic spline models [17–19] with comparison to linear and quadratic models. Model selection was guided by likelihood ratio tests, AIC/BIC criteria, and Wald tests for non-linear terms (p < 0.05), prioritizing splines for their flexibility in capturing non-linear patterns.Cruciferous vegetable (CV) intake was standardized to grams per day (g/d) across all studies using a predefined protocol. Consistent with the USDA FoodData Central database, one serving of CV was defined as 80 g. For studies reporting intake frequency (e.g., times per week), daily gram intake was calculated using the formula: Daily intake (g) = (Weekly frequency × 80 g) /7. When the intake dose provided by the original studies was presented as an interval distribution: 1. The average intake dosage for closed intervals was defined as the middle value between the upper and lower boundaries. 2. The intake dose for open intervals was calculated by dividing the interval's endpoint by 1.2. If a p-value was less than 0.05, it was considered statistically significant. Stata version 14 (StataCorp, College Station, TX, USA) was used for all analyses.

While primary effect estimates and subgroup differences were evaluated at p < 0.05, publication bias tests adopted a more lenient threshold p < 0.1 to enhance sensitivity in detecting small-study effects, consistent with methodological recommendations [20].

Results

Literature search

Five electronic databases yielded 509 records in total. 454 records were kept for title and abstract screening after 55 duplicate records were eliminated. Following the screening of abstracts and titles, 373 records were removed as unnecessary. Through a full-text review, sixteen papers were found, and one more study was found by looking through the retrieved publications' reference lists. Consequently, the final analysis comprised 17 studies (Fig. 1).

Fig. 1.

Literature search and screening process

Study characteristics

Seven cohort studies [21–27] and ten case–control studies [28–37] comprised the 17 studies. The studies involved a total of 639,539 participants, with 97,595 cases of CC. Seven studies were conducted in North America [24, 27, 28, 30, 33, 34, 36], two in Europe [21, 26], one in Australia [25], and seven in Asia [22, 23, 29, 31, 32, 35, 37].Table 1 systematically displays the key characteristics of the collected research. Intake of CV was assessed through the use of validated food frequency questionnaires or questionnaires. In several research, daily intake was calculated using weekly or monthly consumption data. Using the Newcastle–Ottawa Scale (NOS) criteria, the methodological quality of each included study was evaluated. The methodological quality assessment using the Newcastle–Ottawa Scale revealed that among the 17 included studies, 15 (88%) were rated as high quality, 2 (12%) as moderate quality.

Table 1.

Characteristics of included studies

References	Region	Study type	Intake measurement	OR (95% CI)	Case/ subjects	Quality score	Adjustments
Leenders [15]	European countries	Cohort	3 g/day 13 g/day 26 g/day 62 g/day	1.0 1.19(1.05–1.36) 1.18(1.03–1.36) 0.98(0.83–1.17)	2128/442961	7	FC,VC,height, weight,CC,AC,CFC, SS,PA
Hu [30]	Canada	Case–control	102 g/week 349 g/week 186 g/week 193 g/week	1.0 1.3 (0.9–1.8) 1.0 (0.7–1.4) 0.9 (0.6–1.4)	1380/4477	7	10-year age group, province, ES,BMI, SS,AC and TEI
Peters [28]	America	Case–control	10 times/month	1.0(0.99–1.01)	769/1533	8	Fat, protein, carbohydrates, AC, CC, FH, weight, PA, and, if female, pregnancy history
Hara [29]	Japan	Case–control	≤ 2 times/week 3-4 times/week  ≥ 5 times/week	1.0 1.03 (0.49–2.15) 0.9 (0.33–2.48)	115/345	7	SS, AC, FH,TEI, and JA Membership
Vogtmann [22]	China	Cohort	< 66.8 g/day 66.8-122 g/day  ≥ 122.1 g/day	1.0 0.85 (0.6–1.22) 1.06 (0.76–1.50)	340/1013	8	Age, BMI, PA, TEI, red meat intake, total meat intake, education, income, occupation, SS, AC, and FH
Wanxia fang (2018)	China	Case–control	305 kg/year 195 kg/year 182 kg/year 151 kg/year	1.0 0.76 (0.56–1.02) 0.8 (0.59–1.09) 0.83 (0.59–1.18)	833/1666	9	BMI, FH, SS, AC, regular diet, fried food, cured food, hot and spicy food intake,TEI, total noncruciferous vegetables, FC, and red meat intake
Mori [23]	Japan	Cohort	24 g/d 50 g/d 78 g/d 151 g/d	1.0 1.27 (1.06–1.52) 1.15 (0.94–1.40) 1.05 (0.85–1.31)	88172/98551	6	Age,study area, BMI, SS,AC,PA, history of diabetes,colorectal screening,menopausal status in women, use of exogenous female hormones in women,TEI, VSU, n-3 fatty acids, CV, FC, and red and processed meat
Lee [32]	Singapore	Case–control	Q1 Q2 Q3	1.0 0.85(0.56–1.26) 0.5(0.32–0.78)	203/425	6	Age, sex, dialect group and occupational group, with additional factors adjusted for in further analyses
Young [33]	America	Case–control	2.28 times/month 4.21 times/month 3.68 times/month	0.57(0.39–0.83) 0.53 (0.36–0.77) 0.59 (0.41–0.85)	706/1324	8	Age and sex
Flood [24]	America	Cohort	< 0.32 servings/d 0.33–0.45 servings/d 0.46–0.58 servings/d 0.59–0.78 servings/d 0.79 servings/d	1.0 0.78 (0.59–1.04) 0.86 (0.65–1.14) 0.93 (0.7–1.22) 0.95 (0.71–1.26)	485/45490	7	VSU, BMI, height, use of nonsteroidal antiinflammatory drugs,SS, ES, PA, and FC, grains, red meat, calcium,AC
Steinmetz [25]	Australia	Case–control	≤ 1.7 times/week 1.8–3.2 times/week 3.3–5.7 times/week  ≥ 5.8 times/week	1.0 1.21 (0.62–2.34) 0.7 (0.36–1.38) 1.1 (0.57–2.14)	121/362	8	Protein intake, age at first live birth, Quetelet’s index and AC
Steinmetz [34]	America	Cohort	< 1.5 times/week 1.5–2.4 times/week 2.5–4.0 times/week  > 4.0 times/week	1.0 1.11 (0.75–1.64) 0.98 (0.65–1.47) 1.12 (0.74–1.7)	212/35216	8	Age and TEI
Ramadas [35]	Malaysia	Case–control	< 3 times/week  ≥ 3 times/week	1.0 0.52 (0.21–1.27)	59/118	7	Age,ethnicity, gender,PA, height, BMI, waist circumference, TEI, AC,SS
West [36]	America	Case–control	Q1 Q2 Q3 Q4	1.0 0.9 (0.5–1.5) 0.6 (0.3–1.5) 0.3 (0.1–0.8)	231/622	7	Age, BMI, crude fiber,TEI
Voorrips [26]	Netherlands	Cohort	11 g/day 21 g/day 29 g/day 40 g/day 58 g/day	1.0 1.0 (0.69–1.46) 0.82 (0.56–1.22) 0.91 (0.62–1.32) 0.76 (0.51–1.13)	910/2953	8	Age,FH, AC
Michels [27]	America	Cohort	< 1servings/week 1 servings/week 2 servings/week 3–4 servings/week  ≥ 5 servings/week	1.0, 0.95 (0.74–1.23), 0.87 (0.68–1.12), 0.92 (0.72–1.18), 0.89 (0.68–1.15)	937/136089	7	Age,FH, sigmoidoscopy, height,BMI, SS, AC,PA, aspirin use,VSU,TEI, and red meat consumption
Chiu [37]	China	Case–control	≤ 15.0 times/week 15.1–22.3 times/week 22.4–25.3 times/week  ≥ 25.4 times/week	1.0 0.8 (0.6–12) 0.8 (0.6–1.2) 0.7 (0.5–1.0)	931/2483	7	Age,TEI, ES, BMI, income, and PA

OR, odds ratios; CI, confidence interval; BMI, body mass index; FH,family history; SS, smoking status; ES,education status; AC, alcohol consumption; PA, physical activity; FC, fruit consumption; VC, vegetable consumption; VSU, vitamin supplement use;TEI,total energy intake

Overall and dose–response analysis

An inverse association was observed between CV and CC (Fig. 2). The original studies' OR values varied from 1.0 (95% CI 0.99–1.01) in the Ruth K. Peters study [28] to 0.6 (95% CI 0.1–0.8) in the Dee W. West study [36]. Compared to participants in the lowest CV intake group, those in the highest intake group showed a 17% reduction in CC incidence (OR = 0.83; 95% CI 0.59–1.18). Between the studies, there was moderate heterogeneity (p = 0.000, I2 = 63.6%).

Fig. 2.

Forest plots of the association between CC and CV consumption

The findings of the dose–response analysis show a non-linear dose–response relationship(Fig. 3). According to the study, CV may lower the risk of CC. A significant decrease in the risk ratio was observed at an intake level of approximately 20 g/d, with the declining trend in risk ratio leveling off at an intake of 40 g/d. Cruciferous vegetable intake exhibits an inverse nonlinear dose–response with disease risk. Peak protective effect per gram occurs at 20–40 g/day (β =  − 0.0059; OR reduction 0.0059 per gram), attenuating substantially beyond 60 g/day (β =  − 0.0028 at 60–80 g/day). Optimal intake is achieved at 40–60 g/day (OR 0.74–0.80), providing near-maximal risk reduction.The spline model demonstrated superior fit (AIC = 42.1) versus linear (AIC = 58.3) and quadratic (AIC = 47.6) alternatives. The pooled analysis demonstrated a significant inverse association (OR = 0.80, 95% CI 0.72–0.90). Funnel plot asymmetry (Fig. 4) indicated publication bias, quantitatively confirmed by major asymmetry (LFK index = 2.31,|LFK|> 2) and significant small-study effects (Egger’s test: p = 0.001). Trim-and-fill adjustment imputed 3 missing studies on the left side of the funnel (Fig. 4), attenuating the effect to OR = 0.85 (95% CI 0.67–1.08). Leave-one-out analysis showed no single study considerably altered the estimate (all OR 0.78–0.82). Sensitivity analysis excluding outliers yielded OR = 0.81 (0.69–0.95).

Fig. 3.

Dose–response relationship between cruciferous vegetable consumption and odds ratios. Lines represent: solid line—point estimate (adjusted odds ratios); Dashed lines—95% confidence interval

Fig. 4.

Trim and fill analysis for OR of the relationship of CV and CC

Subgroup and sensitivity analysis

To minimize heterogeneity, we performed subgroup analyses. The results by research type, area, and confounding factors are shown in Table 2. Consuming CV was linked to a lower incidence of CC in North America (OR = 0.82, 95% CI 0.67–0.99)and Asia (OR = 0.77, 95% CI 0.67–0.89), but not in Europe (OR = 0.94, 95% CI 0.80–1.10) or Australia (OR = 0.57, 95% CI 0.27–1.20). According to the subgroup analysis by research type, case–control studies had a lower prevalence of CC (OR = 0.79, 95% CI 0.70–0.87) than cohort studies (OR = 0.89, 95% CI 0.81–0.99). Additionally, most confounding-adjusted subgroups showed significant inverse associations, except for alcohol consumption and family history.

Table 2.

Subgroup analysis of CV with the risk of CC

	Number of studies	Summary OR	95%CI	I2(%)	P-value	P for interaction
Overall	17	0.80	0.72–0.90
Study location						0.064
Europe	2	0.942	0.80–1.10	24.40%	0.455
North America	7	0.82	0.67–0.99	70.30%	0.003
Asia	7	0.774	0.67–0.89	14.50%	0.319
Australia	1	0.57	0.27–1.20	-	-
Adjustment for confounders
FH for CC						< 0.001
Yes	5	0.999	0.99–1.00	0.00%	0.425
No	12	0.754	0.64–0.88	50.70%	0.022
Alcohol consumption						0.112
Yes	10	0.999	0.99–1.01	15.90%	0.297
No	7	0.631	0.53–0.76	32.50%	0.18
Smoking status						0.005
Yes	9	0.882	0.80–0.97	0.00%	0.798
No	8	0.69	0.53–0.90	78.90%	0
Total energy intake						0.618
Yes	9	0.823	0.72–0.94	0.00%	0.489
No	8	0.794	0.67–0.94	75.10%	0
Study type						0.012
Case–control	10	0.685	0.54–0.87	75.60%	0
Cohort	7	0.899	0.81–0.99	0.00%	0.811

FH, family history; Dashes (–) indicate non-applicable values

The leave-one-out analysis indicates that no single study considerably changes the pooled estimate (Fig. 5).

Fig. 5.

Sensitivity analysis: leave-one-out analysis

Discussion

According to the current meta-analysis, there is strong evidence that consuming more CV may reduce the risk of CC. Our results provide fresh insights through thorough stratified analyses, validating and significantly building upon the seminal work of Tse et al. [7]. Geographical stratification revealed statistically significant inverse correlations in both Asian populations (OR = 0.77, 95% CI 0.67–0.89) and North American (OR = 0.82, 95% CI 0.67–0.99), potentially attributable to regional variations in dietary composition, culinary practices, and food preparation techniques. The observed heterogeneity was further characterized by more pronounced effect estimates in case–control studies compared to prospective cohort designs, suggesting potential recall bias in retrospective studies while highlighting the need for additional longitudinal investigations. These findings emphasize the imperative for well-designed, population-specific prospective cohort studies to delineate potential ethnocultural differences in dietary patterns and their impact on colon carcinogenesis. Importantly, our multivariate analysis incorporated adjustment for major confounding factors, including tobacco use, ethanol consumption, family history of CC, and total energy intake, thereby enhancing the validity of our conclusions.

Epidemiological investigations and familial aggregation studies have substantiated significant heritability patterns in CC, highlighting the genetic predisposition to this malignancy [38]. Contemporary epidemiological and longitudinal studies have established positive associations between alcohol consumption, tobacco use, and elevated CC risk [38]. Additionally, the pathophysiology of CC has been linked to dietary factors, specifically inadequate intake of vegetables and dietary fiber, as well as excessive alcohol and caffeine use [39–42]. These empirical findings lend credence to our results, suggesting a potential chemopreventive role of CV against CC development.

With a focus on their anti-neoplastic potential in recent years, the scientific literature has thoroughly examined and documented the chemopreventive qualities of CV [43]. The exact molecular processes behind the chemopreventive effectiveness of CV have not yet been fully understood, despite the fact that a lot of researches have examined the association between CV and cancer. The main focus of current mechanistic research has been on how bioactive compounds derived from CV modulate enzymes that metabolize xenobiotics. In particular, myrosinase produces physiologically active metabolites such as indole-3-carbinol (I3C) and sulforaphane (SFN) through the enzymatic hydrolysis of glucosinolates, which are naturally occurring sulfur-containing phytochemicals in CV. Experimental evidence demonstrates that SFN induces phase II detoxification enzymes through Nrf2-mediated pathways [44], concurrently inhibiting PI3K/Akt signaling cascades and promoting caspase-dependent apoptosis in malignant cells. In parallel, I3C exerts its anti-proliferative effects through epigenetic modulation and downregulation of cyclin-dependent kinase inhibitors p27 and p21, resulting in diminished CDK6 activity and subsequent G1/S phase cell cycle arrest [6]. Recent studies further elucidate the chemopreventive mechanisms of cruciferous vegetables. Sulforaphane (SFN), a major isothiocyanate in broccoli, has been shown to inhibit histone deacetylases (HDACs), thereby reactivating tumor suppressor genes such as p16 and APC in colon cancer models [45]. Additionally, SFN inhibits colorectal carcinogenesis by modulating the β-catenin pathway via ZO-1, a key tight junction protein for epithelial integrity. Chen et al. [46] showed that SFN upregulates ZO-1, sequestering β-catenin at cell–cell junctions to inhibit its nuclear translocation and cancer stem cell(CSC)-related gene activation (e.g., CD44, LGR5).These findings align with our dose–response results, where risk reduction plateaued at higher intake levels (~ 40 g/d), potentially reflecting saturation of key molecular targets. One significant methodological accomplishment of the current study is the use of non-linear dose–response analysis, which provides compelling evidence for the negative relationship between CV intake and CC risk. Extensive subgroup analyses stratified by research design, geographic location, and confounding factor correction were conducted in order to further evaluate potential causes of heterogeneity among the included studies.

We observed widening confidence intervals at higher intake levels (> 50 g/day) in our dose–response analysis (Fig. 3). This pattern is primarily attributable to three factors. First, data density was substantially reduced in this range, with only three included studies [22, 29, 30] assessing intakes exceeding 60 g/day, representing less than 15% of the total analytical sample. Second, considerable heterogeneity existed in the quantification of high intakes across these studies, exemplified by variations in methods such as the conversion between cooked and raw vegetable equivalents. Third, a biological plausibility ceiling is suggested by mechanistic evidence indicating saturation of Nrf2 pathway activation at approximately 40 g/day [44], consistent with the plateau effect observed in our analysis. Critically, the point estimate remained below the null value (OR = 1.0) across the entire dose range, with a minimum odds ratios (OR) of 0.82 observed at 70 g/day. While the confidence intervals crossing the null at the highest intakes indicate that evidence for additional protective effects at supraphysiological levels (> 60 g/day) is inconclusive, this finding does not refute the existence of significant protective effects at lower, more commonly consumed intake levels.

The funnel plot asymmetry (Fig. 4) and statistical indices (LFK = 2.31; Egger’s p = 0.001) jointly indicated publication bias. Trim-and-fill imputation of 3 left-side studies (Fig. 4) attenuated the pooled OR by 6.25% (0.80 → 0.85), consistent with exaggerated effects in small studies (e.g., WEST 1989 [OR = 0.3], LEE 1989 [OR = 0.5]). Crucially, directional consistency persisted: all OR estimates remained below 1 (trim-and-fill: 0.85; sensitivity: 0.81). Robustness was confirmed by leave-one-out analysis (no influential study), and the 6.25% attenuation magnitude remained below typical nutritional epidemiology bias levels (15–30%). While the adjusted CI (0.67–1.08) crosses the null, the point estimate stability and mechanistic plausibility sustain biological relevance.

This meta-analysis has several limitations. First, the inclusion of both case–control and cohort studies may introduce methodological heterogeneity, as case–control designs are susceptible to recall bias that could overstate the association between CV intake and CC incidence. Second, significant heterogeneity existed in CV intake assessment across studies (e.g., food frequency questionnaires vs. 24-h recalls), potentially introducing measurement error. Third, residual confounding from unmeasured factors (e.g., pesticide exposure, genetic susceptibility) cannot be ruled out despite adjustment for major covariates. Fourth, the predominance of studies from North America and Asia—regions with elevated CC incidence—limits generalizability to populations with distinct dietary patterns or lower CC prevalence. Finally, while publication bias was quantitatively confirmed through funnel plot asymmetry (LFK index = 2.31,|LFK|> 2), significant small-study effects (Egger’s test: p = 0.001), and trim-and-fill analysis (which imputed 3 missing studies and attenuated the association to OR = 0.85, 95% CI 0.67–1.08), the robustness of the primary inverse association (OR = 0.80, 95% CI 0.72–0.90) was supported by sensitivity analyses. Leave-one-out analysis showed no influential study (all OR 0.78–0.82), and exclusion of outliers maintained significance (OR = 0.81, 95% CI 0.69–0.95). Additionally, the spline model demonstrated superior fit (AIC = 42.1) for the dose–response relationship. Nevertheless, inherent heterogeneity in study design, population characteristics, and exposure measurement methodologies may still influence pooled estimates. These limitations necessitate cautious interpretation of the findings.

To address these limitations and advance the field, we propose the following priorities: Standardized Intake Assessment, implement harmonized dietary tools across future cohort studies to reduce measurement heterogeneity. Integration of Biomarkers, measure urinary ITC metabolites to objectively quantify CV exposure and validate self-reported intake data. Mechanistic Trials, conduct randomized controlled trials evaluating the impact of CV supplementation on intermediate endpoints (e.g., rectal mucosal HDAC activity, β-catenin expression) in high-risk populations. Global Representation, prioritize multicenter studies in under-researched regions (e.g., Africa, South America) to assess cultural and dietary variability in CV effects.We strongly advocate that future observational studies routinely report both crude and fully adjusted effect estimates in their primary findings, while further standardizing analytical approaches through predefined confounder adjustment sets constructed via causal diagrams (e.g., directed acyclic graphs) to enhance cross-study comparability and minimize residual confounding.

Conclusion

This meta-analysis confirms a nonlinear inverse dose–response relationship between cruciferous vegetable intake and colorectal cancer risk, with significant risk reduction initiating at ≥ 20 g/d. Peak protective effects per gram occur at 20–40 g/d (β =  − 0.0059), while optimal intake is observed at 40–60 g/d (OR 0.74–0.80). Limitations include methodological heterogeneity (mixed case–control/cohort designs susceptible to recall bias), variable exposure assessment, residual confounding, limited generalizability (predominantly North American/Asian studies), and publication bias (LFK index = 2.31; Egger's *p* = 0.001). The primary inverse association (OR = 0.80, 95%CI 0.72–0.90) remained robust in sensitivity analyses (consistent leave-one-out estimates, outlier exclusion, and spline model superiority [AIC = 42.1]). Future high-quality prospective cohorts in underrepresented regions (e.g., Africa, South America) are essential for validation.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

(I) Conception and design:Bo Lai. (II) Administrative support:Zhong Li. (III) Provision of study materials or patients:Junjie Li. (IV) Collection and assembly of data:Bo Lai. (V) Data analysis and interpretation:Bo Lai. (VI) Manuscript writing: Bo Lai,Zhong Li,Junjie Li. (VII) Final approval of manuscript: Bo Lai, Zhong Li, Junjie Li

Funding

All authors announced that they had not received any fund support.

Data availability

Data is provided within the manuscript or supplementary information files.

Declarations

Ethics approval and consent to participate

As this meta-analysis exclusively synthesized aggregated data from previously published studies, no direct involvement of human participants occurred. Consequently, neither ethics re-approval nor additional informed consent was required, in accordance with international guidelines for evidence synthesis research. All original studies included in this analysis declared compliance with ethical standards in their respective publications, including appropriate ethics board approvals and participant consent procedures.An ethics statement is not applicable because this study is based exclusively on published literature. The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.