Risk assessment of digestive cancers with the modified EAT-Lancet diet pattern: a prospective cohort study based on UK Biobank Authors: Daixin Wu, Junwei Huang, Juncheng Chen, Zhengyu Wu, Jiale Jiang, Yuan Yuan, Rufu Chen, Chonghui Hu, Yong Jiang, Shangyou Zheng Published: 2025 ABSTRACT AbstractBackgroundWhile diet is a modifiable risk factor for digestive cancers, existing dietary scores are not optimized for site-specific risk assessment. We aimed to develop and evaluate cancer-specific dietary scores for improved risk stratification.ObjectivesTo modify the EAT-Lancet diet (ELD) score into cancer-specific versions and evaluate their association with the risk of seven major digestive cancers.MethodsBased on 111,785 UK Biobank participants, we first expanded the ELD score and then derived cancer-specific scores (Score3) for seven digestive cancers using LASSO-Cox regression. Associations with cancer incidence were assessed using Cox models, Kaplan-Meier curves, and dose-response analyses. Model discrimination was evaluated using the C-index, and improvements over the original ELD score were quantified by net reclassification improvement (NRI). Subgroup analyses were conducted based on sex, smoking status, and polygenic risk score (PRS), with additive interactions between the dietary score and PRS assessed using attributable proportion (AP) and relative excess risk due to interaction (RERI), and multiplicative interactions evaluated using interaction terms in the Cox models.ResultsCancer-specific Score3 showed improved discriminatory ability compared to the original ELD score, with C-index of 0.623, 0.630, 0.611, and 0.604 for esophageal, liver, rectal, and pancreatic cancers, respectively, and demonstrated positive NRI for these cancers. In multivariable-adjusted models, the lowest Score3 category was associated with a significantly higher risk for these four cancers: esophageal cancer (HR = 1.702, 95% CI: 1.202–2.410), liver cancer (HR = 1.667, 95% CI: 1.084–2.561), rectal cancer (HR = 1.662, 95% CI: 1.273–2.171), and pancreatic cancer (HR = 1.576, 95% CI: 1.201–2.069). Correspondingly, the low-score group exhibited a higher absolute risk per 10,000 person-years compared to the high-score group. A strong association was observed between low Score3 and high rectal cancer risk among individuals with high genetic predisposition (PRS); significant additive interaction was observed (AP = 0.286, RERI = 0.687). Score3 performance varied by sex and was substantially amplified among smokers.ConclusionCancer-specific modifications of the ELD score may improve risk assessment for digestive cancers, particularly esophageal, liver, rectal, and pancreatic cancers. Across digestive cancers, the associations of these modified scores also vary by sex and smoking, highlighting their potential relevance for personalized dietary guidance. The utility of the modified score is most evident in individuals with high polygenic risk for rectal cancer.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12876-025-04558-8.Keywords: EAT-Lancet diet, Digestive cancer, Risk, UK Biobank Introduction Digestive cancers, including cancers of the esophagus, liver, stomach, rectum, biliary tract, colon, and pancreas are highly prevalent and among the leading causes of cancer-related mortality worldwide [1–3]. According to 2018 data, 4.8 million new cases and 3.4 million deaths were reported globally, accounting for 26% of all cancers and 35% of cancer-related fatalities, respectively [2, 4, 5]. This substantial burden underscores the critical need for effective prevention strategies, especially given that these cancers are often diagnosed at advanced stages despite therapeutic advances [6]. In this context, modifiable lifestyle factors, particularly diet, are of paramount importance. A growing body of evidence indicates that healthy dietary patterns are inversely associated with the risk and mortality of digestive cancers [7, 8]. In 2019, the EAT-Lancet Commission introduced the “EAT-Lancet diet” (ELD), promoting increased plant-based food intake and reduced consumption of red meat and added sugars to balance human health and environmental sustainability [9] Unlike other established dietary indices—such as the Mediterranean diet, Dietary Approaches to Stop Hypertension (DASH), and the Alternative Healthy Eating Index (AHEI), which were primarily developed for cardiovascular and metabolic outcomes [10–12] —the ELD score may be more directly related to lower digestive cancer risk [13–15]. Indeed, the link between red and processed meat consumption and digestive cancer risk is well-established. The International Agency for Research on Cancer (IARC) classifies processed meat as a Group 1 human carcinogen and red meat as a Group 2 A probable human carcinogen [16]. Recent systematic reviews and meta-analyses consistently associate processed meat intake with increased risks of digestive cancers, supported by mechanistic pathways such as heme-iron–induced oxidative stress, endogenous N-nitroso compound formation, and nitrite-related exposures [17–20]. Evidence also highlights the relevance of total red-meat intake. Several large cohort and pooled studies demonstrate that higher overall consumption of red and processed meat increases the risk of colorectal and other gastrointestinal cancers [21–23]. These findings indicate that carcinogenic risk reflects cumulative meat exposure rather than any single subtype, supporting the use of total red-meat intake as a stable, biologically meaningful metric and providing a rationale for constructing a digestive-cancer–oriented dietary score. To date, the ELD score has not been specifically modified or validated for assessing digestive cancer risk. Therefore, utilizing data from the UK Biobank, this study aimed to construct and evaluate cancer-specific modifications of the ELD score. Specifically, our objective was to develop a novel dietary risk assessment framework by constructing separate, cancer-type-specific scores for seven digestive cancers and to examine their associations with cancer incidence. Methods Study population This study was conducted using data from the UK Biobank (UKB; https://www.ukbiobank.ac.uk/). The UKB is a large-scale prospective cohort study that recruited over 500,000 participants aged 37–73 years from 22 assessment centers between 2006 and 2010. All participants completed touchscreen questionnaires, brief interviews, physical measurements, and provided biological samples. The study was approved by the North West Multi-Center Research Ethics Committee, and all participants provided written informed consent [24, 25]. We performed a landmark analysis. The date of the first dietary assessment was defined as baseline, with a landmark set at three years thereafter. Participants who were diagnosed with cancer or died between baseline and this landmark were excluded. Person-time for the remaining eligible participants began accruing from the landmark point. Incident cancer cases were ascertained through linkage with the National Cancer Registry and hospital patient records, using International Classification of Diseases (ICD-9 and ICD-10) codes (Table S1). Follow-up continued until the earliest occurrence of cancer diagnosis, death, or the end of the study period (October 13, 2022) (Table S2). The final analytical cohort was established by further applying the following exclusion criteria: participants who completed fewer than two 24-hour dietary recall questionnaires, those with missing covariate information, and individuals with implausible energy intake (< 800 or > 4000 kcal/day for males and < 500 or > 3500 kcal/day for females), as detailed in Fig. 1. Flowchart for selecting the UK Biobank study population Dietary assessment and EAT-Lancet diet score In the UKB, dietary intake was assessed at five time points using the “Oxford WebQ,” a validated web-based 24-hour dietary recall tool designed for large-scale prospective studies. This tool has been validated against interviewer-administered 24-hour recalls and biomarkers [26]. The 24-hour recall assessments were open-ended, with participants selecting from a list of 206 foods and 32 beverages that they reported consuming in the past 24 h. The assessments were conducted at five intervals: assessment center (April 2009–September 2010), online cycle 1 (February–April 2011), online cycle 2 (June–September 2011), online cycle 3 (October–December 2011), and online cycle 4 (April–June 2012). The consumption of each food or beverage within 24 h was calculated by multiplying each item’s specified portion size by the quantity consumed (Table S3), and this data is available in the UK Biobank [27, 28]. To represent habitual intake, data from all available recalls were aggregated at the food-group level, and the mean intake of each component was calculated for each participant. These mean values were used to construct all dietary scores. The analysis used the EAT-Lancet Commission’s sustainable dietary recommendations. Adherence to these recommendations was assessed using the Stubbendorff ELD score [25], which applied an ordinal scale to each dietary component (Table S8), and was designated as Score1 in this study. To improve scientific accuracy, we replaced the original "added sugar" component with "free sugars" in accordance with WHO and UK SACN definitions, where free sugars include all monosaccharides and disaccharides added to foods plus sugars naturally present in honey, syrups, and fruit juices, thereby offering a broader and more physiologically relevant measure than added sugars alone [29, 30]. For the fat component, fat intake was defined using the ratio of total unsaturated fat to saturated fat. Unsaturated fat included n-3 fatty acids, n-6 fatty acids, and monounsaturated fatty acids, whereas saturated fat referred exclusively to saturated fatty acids. Higher ratios reflected more favorable fat quality and received higher ordinal scores. Modification of dietary score To improve the relevance of the ELD scoring system for evaluating digestive cancer risk, we first expanded the original Stubbendorff ELD score (Score1) by incorporating beef, pork, and lamb into the red meat category and adding processed meat intake [16, 31]. This modification resulted in Score2, which included 14 dietary components (Table S4). Using Score2 as the foundation, we constructed cancer-specific Score3 for each digestive cancer subtype (esophagus, liver, stomach, rectum, biliary tract, colon, and pancreas). All 14 dietary components were standardized before being entered into a LASSO-Cox regression model. Cross-validation was applied to identify the penalty parameter that minimized the cross-validated error (minimum λ). Dietary components with non-zero coefficients at the minimum λ were retained. The standardized coefficients were then used as weights and multiplied by the corresponding dietary component scores. The weighted components were summed to generate cancer-specific Score3. All scores (Score1, Score2, and the seven cancer-specific Score3) were categorized into approximate tertiles (low, middle, high) and applied in the evaluation of digestive cancer incidence. Sensitivity analysis Seven digestive cancer types that demonstrated significant associations with the modified dietary score were further examined through sensitivity analyses to assess the robustness of these findings. Two adjusted Cox proportional hazards models were applied. Model 1 included the covariates sex, age, ethnicity, BMI, smoking status, alcohol consumption, total energy intake, and the Townsend deprivation index (TDI), and excluded participants who developed cancer within 2 years after the 3-year landmark (5 years for pancreatic cancer) to minimize potential reverse causation. TDI is a measure of material deprivation, considering factors like unemployment, non-homeownership, and overcrowded living conditions [32]. Model 2 included the same covariates except BMI and did not apply any exclusion based on early cancer incidence, thereby allowing evaluation of the associations without time-related restrictions. PRS interaction analysis Polygenic risk score (PRS) was constructed to examine the relationship between genetic susceptibility and the dietary score in relation to risks of four digestive cancers. For rectal cancer, we derived the PRS based on summary statistics from a previous genome-wide association study (GWAS) by Rashkin et al. [33]. To ensure consistency with the ancestry of the UK Biobank cohort, SNP selection was restricted to variants with available effect estimates from populations of predominantly European ancestry. Linkage disequilibrium (LD) pruning was performed using a clumping window of 10,000 kb and an LD threshold of r2 = 0.3, yielding a set of independent SNPs for PRS construction. The PRS was calculated as the weighted sum of risk alleles using the published GWAS beta coefficients (Table S4). For three additional digestive cancers—pancreatic (Table S5), esophageal (Table S6), and liver cancers (Table S7)—the PRSs were constructed using SNPs and effect sizes directly reported in previously published studies: Sharma et al. for pancreatic cancer [34], Lin et al. for esophageal cancer [35], and Yu et al. for liver cancer [36]. These studies provide validated SNP panels and effect estimates derived from large-scale GWAS, which we adopted without modification to maximize reproducibility and consistency with prior literature. For all cancers, PRS values were computed as: where βn represents the effect size (or weight) of each SNP, and SNPn is the genotype at the corresponding SNP. The PRS was categorized into low, median, and high groups based on tertiles to evaluate the stability of the dietary score’s predictive ability for the risk of the four digestive cancers across different PRS subgroups. On this basis, we further explored the interaction between PRS and the modified dietary score in relation to the risk of the four digestive cancers. Statistical analyses Continuous variables with a normal distribution were compared using t-tests, while non-normally distributed continuous variables were analyzed using the Mann–Whitney U test. Categorical variables were compared using the Chi-square test, and two-sided p-values < 0.05 were considered statistically significant. Associations between dietary scores (Score1, Score2, and cancer-specific Score3) and digestive cancer risk were assessed using Cox proportional hazards regression and Kaplan–Meier (KM) analyses. The KM method and log-rank tests were applied to estimate cumulative incidence for seven digestive cancers—pancreatic, colon, stomach, liver, rectal, esophageal, and biliary tract cancers—across dietary score groups, with incidence differences assessed separately for each cancer type. Dose–response relationships between cancer-specific Score3 and the risk of the seven digestive cancers were further examined using restricted cubic spline Cox regression, immediately following the survival curve analyses. For univariable Cox models, hazard ratios (HRs), C-index (with 95% confidence intervals), ΔC-index, net reclassification improvement (NRI), and 10-year flexible calibration curves were reported. Absolute risk was additionally reported per 10,000 person-years, standardized by sex, across dietary score categories. For cancer-specific Score3, internal validation was performed, including bootstrap-based optimism-corrected C-index, calibration slopes, overfitting estimates, and both 70/30 training/test set splits and nested five-fold cross-validation to assess discrimination and calibration. Multivariable Cox models were fitted with covariate adjustment, and the Fine-Gray subdistribution hazards model was applied to account for death as a competing event, yielding subdistribution hazard ratios (SHR). Schoenfeld residuals were examined, and E-values were calculated in all primary models to verify the proportional hazards assumption and evaluate the potential impact of unmeasured confounding. For both the univariable Cox models of the three dietary scores (Score1, Score2, and cancer-specific Score3) and subgroup analyses, multiple comparisons across related cancer types were controlled using the Benjamini–Hochberg false discovery rate (FDR) to account for family-wise testing. Subgroup analyses stratified by PRS were conducted to evaluate associations between cancer-specific Score3 and digestive cancer risk. Within these PRS-stratified models, additive interactions were quantified using relative excess risk due to interaction (RERI) and attributable proportion (AP), while multiplicative interactions were assessed using the corresponding interaction coefficients with 95% confidence intervals. Results Baseline demographic characteristics In the UKB study, a total of 111,785 participants were included, comprising 62,313 females (55.7%) and 49,472 males (44.3%) (Table S9). The mean age at baseline was 55.72 years (SD = 7.83), and the median follow-up duration was 4,882.45 days (SD = 436.80). During the follow-up period, 224 esophageal cancer cases, 146 liver cancer cases, 163 stomach cancer cases, 339 rectal cancer cases, 75 biliary tract cancer cases, 914 colon cancer cases, and 325 pancreatic cancer cases were identified. Individuals who were male, had higher BMI, or reported previous or current smoking generally exhibited an increased risk of developing these cancers. Modification of the dietary score Score2 (Table S8) was derived by adding the processed meat component to Score1. Subsequently, the 14 dietary components of Score2 were entered into LASSO regression models for each of the seven digestive cancer outcomes. Using 10-fold cross-validation to select the optimal λ, cancer-specific dietary components and their corresponding weights were identified (Table S10) and summed to construct the cancer-specific Score3. All three scores (Score1–Score3) were then categorized into low, middle, and high groups for comparison. As shown in Table S11, the C-index of Score3 for esophageal, liver, stomach, rectal, biliary tract, colon, and pancreatic cancer were 0.623, 0.630, 0.591, 0.611, 0.647, 0.541, and 0.604, representing improvements of 0.044, 0.079, 0.078, 0.074, 0.081, 0.019, and 0.033 over Score1, respectively. The corresponding NRI estimates (Table S13) were 0.041 (95% CI: − 0.060 to 0.139), 0.113 (95% CI: 0.019 to 0.203), 0.089 (95% CI: 0.001 to 0.176), 0.085 (95% CI: 0.036 to 0.133), 0.134 (95% CI: 0.024 to 0.258), 0.025 (95% CI: − 0.010 to 0.057), and 0.058 (95% CI: 0.004 to 0.118). Score3 demonstrated good internal validity across all cancer models (Table S12). The bootstrap-corrected calibration slopes were close to 1 across outcomes, indicating minimal overfitting. The C-index obtained from the 70/30 train-test split and 5-fold cross-validation were comparable to the optimism-corrected bootstrap C-index, supporting robust discrimination. As shown in Figure S1, the proportional hazards assumption was satisfied for all Score3 models based on Schoenfeld residual tests. Across all cancer types, lower Score3 categories were consistently associated with higher hazard ratios and higher absolute risks per 10,000 person-years. The increased risks observed in low-score groups were statistically significant for all cancers except biliary tract cancer (P_FDR < 0.05). Association between cancer-specific Score3 and the seven digestive cancers After adjustment for age, sex, ethnicity, BMI, smoking status, alcohol consumption frequency, TDI and energy, lower Score3 categories were associated with substantially higher risks of esophageal, liver, rectal, and pancreatic cancers compared with the high-score group (Fig. 2). Specifically, the low-score group had 70.2%, 66.7%, 66.2%, and 57.6% higher risks for these four cancers, respectively, with all confidence intervals excluding 1. The SHRs were consistent, and the corresponding E-values were 2.796, 2.720, 2.712, and 2.529. Kaplan–Meier curves for these cancers showed clear separation across Score3 categories (Fig. 3). In addition, their 10-year predicted survival probabilities aligned well with the 45° reference line, indicating good calibration (Figure S2). Dose–response analyses demonstrated significant linear associations for all four cancers (P for overall < 0.05), and pancreatic cancer also showed a significant nonlinear component (P for non-linear = 0.014) (Fig. 4). Figure S5 Associations between Score3 (cancer-specific) and risks of seven digestive cancers. HRs were obtained from multivariable Cox proportional hazards models adjusted for sex, age, ethnicity, smoking status, alcohol frequency, Townsend deprivation index and energy. E-values were calculated to assess the robustness of associations to unmeasured confounding. For subtypes with competing events, SHRs were estimated using Fine-Gray models Cumulative risk of four digestive cancers during follow-up according to Score3 (cancer-specific). a Esophagus cancer b Liver cancer c Rectal cancer d Pancreatic cancer Dose-response association of the risks of four digestive cancers with Score3 (cancer-specific). Models were adjusted for sex, age, ethnicity, BMI, smoking status, alcohol frequency, Townsend deprivation index and energy. a Esophagus cancer b Liver cancer c Rectal cancer d Pancreatic cancer In contrast, for stomach, biliary tract, and colon cancers, Score3 categories displayed graded separation in the KM analyses (Figure S3), but the multivariable-adjusted associations were not statistically significant (Fig. 2). The corresponding 10-year predicted survival probabilities showed relatively narrow prediction intervals and poorer calibration performance (Figure S2). Dose–response analyses for these cancers did not identify significant linear or nonlinear trends (Figure S4). Sensitivity analysis The significant associations for esophageal, liver, rectal, and pancreatic cancers proved robust under alternative model specifications (Figure S5). In Model 1, which extended the landmark exclusion to further address reverse causation, the associations persisted with similar effect sizes. In Model 2, which excluded body mass index (BMI) to evaluate its role as a potential confounder or mediator, the hazard ratios were virtually unchanged. The consistency across these models supports the robustness of the primary findings for these four cancers. In contrast, stomach, biliary tract, and colon cancers again showed no clear risk differentiation across Score3 categories. Subgroup and interaction analyses of cancer-specific Score3 and genetic risk As shown in Fig. 5, clear effect modification was observed for rectal cancer. Across low, median and high PRS categories, participants with low dietary scores had substantially higher risks compared with those with high scores, with risk increases of 101.8%, 112.9%, and 78.6%, respectively (all P_FDR < 0.05). Correspondingly, additive interaction was statistically significant (Table S14), with a RERI of 0.687 (95% CI: 0.045–1.260) and an AP of 0.286 (95% CI: 0.028–0.491). No evidence of multiplicative interaction was found (HR for multiplicative interaction: 1.323; 95% CI: 0.871–2.010). Associations of risks of rectal cancer with Score3 (cancer-specific), stratified by PRS subgroup For the other three cancers, no significant interactions were detected. Nevertheless, compared with the high dietary score group (Figure S6), esophageal cancer risk was elevated in the low dietary score category among both high-PRS and low-PRS groups (203.0% and 135.6% increases, respectively; P_FDR < 0.05). For pancreatic cancer, participants with median PRS and low dietary scores had a 213.3% higher risk (P_FDR < 0.05). Although the associations for liver cancer did not remain statistically significant after FDR correction, a consistent pattern of higher risk with lower dietary scores was still observed. Subgroup analyses by sex and smoking status The association between Score3 and cancer risk was modified by both sex and smoking status. In the sex-stratified analyses (Figure S7), the associations between Score3 and cancer risk differed across subgroups. Among females, the low-score group showed substantially higher risks of esophageal and pancreatic cancers compared with the high-score group, with increases of 102.8% and 82.1%, respectively. Among males, the low-score group exhibited greater increases in the risks of liver and rectal cancers, rising by 132.7% and 101.8%, respectively. In the smoking-stratified analyses (Figure S8), stronger associations were observed among previous or current smokers. Within this subgroup, individuals in the low-score group demonstrated markedly higher risks of esophageal, liver, and pancreatic cancers, increasing by 161.5%, 194.6%, and 125.2%, respectively, with all corresponding P_FDR values < 0.05. For rectal cancer, however, the differences between smoking categories were minimal, with comparable risk patterns observed in previous or current smokers and never smokers. Discussion In this study, we developed cancer-specific dietary scores (cancer-specific Score3) for seven digestive cancers. Score3 demonstrated consistent predictive value for esophageal, liver, rectal, and pancreatic cancers, with C-index of 0.623, 0.630, 0.611, and 0.604, respectively—representing clear improvements over the original ELD score (ΔC-index: 0.044–0.081). NRI analyses further indicated better risk classification for liver, rectal, and pancreatic cancers (0.113, 0.085, and 0.058, respectively). KM curves showed clear separation across low-, middle-, and high-score groups. In multivariable Cox models adjusting for covariates, individuals in the low-score group had substantially higher risks of esophageal, liver, rectal, and pancreatic cancers—by 70.2%, 66.7%, 66.2%, and 57.6%, respectively—with confidence intervals excluding unity and corresponding increases observed in absolute risk per 10,000 person-years. Dose–response analyses supported largely linear associations, and sensitivity analyses produced generally consistent results, supporting the robustness of these findings. For stomach, biliary tract, and colon cancers, risk gradients were apparent descriptively in KM curves, although associations were not statistically significant after multivariable adjustment. These results indicate that Score3 provides meaningful discrimination for esophageal, liver, rectal, and pancreatic cancer risk, while performance is more limited for others. Currently, mainstream dietary indices—including the Mediterranean Diet Score (MDS) [37], the Dietary Approaches to Stop Hypertension (DASH) [10], and the Alternative Healthy Eating Index (AHEI) [11] —primarily target health outcomes such as cardiovascular disease, metabolic syndrome, and all-cause mortality. These indices emphasize higher consumption of vegetables, fruits, whole grains, nuts, and fish, and lower intake of red and processed meats and sugar-sweetened beverages. The MDS highlights monounsaturated-fat–rich foods, the DASH diet prioritizes sodium reduction to improve blood pressure, and the AHEI focuses on dietary components linked to metabolic and inflammatory pathways. However, their relevance to digestive cancers is limited, as they do not specifically capture dietary exposures most strongly linked to digestive carcinogenesis. The ELD score, by contrast, is structured around a predominantly plant-based dietary pattern, prioritizing whole grains, legumes, nuts, fruits, and vegetables while restricting red meat intake. Its design integrates nutritional principles with environmental sustainability, providing a framework that aligns with dietary components relevant to digestive tract health. Building on this foundation, the modified ELD scores (cancer-specific Score3) were tailored to each digestive cancer type using data-driven derivation of cancer-specific weights, reflecting distinct dietary risk patterns across the seven cancers. This progression from a general dietary framework to cancer-specific scoring enables each modified score to more directly represent the dietary risk profile of its corresponding cancer type. We further examined the dietary components contributing to the observed associations. Evidence from recent cohort studies consistently shows that higher consumption of red and processed meat is associated with increased risks of rectal cancer, pancreatic cancer, and liver cancer [38–40]. These relationships are biologically plausible, as heme-iron–related nitrosation, lipid peroxidation, and meat-derived mutagens (HCAs and PAHs) can promote mucosal proliferation, oxidative stress, and chronic inflammation, while nitrite-rich processed meat may further aggravate hepatic metabolic stress and bile-acid–related carcinogenic pathways [16, 22, 41]. Potatoes, a food item with a high glycemic index within the modified score, can induce rapid postprandial glucose and insulin responses that contribute to insulin resistance and IGF-1–mediated proliferative signaling, mechanisms that may be involved in colorectal and pancreatic carcinogenesis [42–44]. Conversely, vegetables provide fiber and bioactive compounds (e.g., isoflavones) with potential protective effects [45–47], and omega-3 fatty acids have been associated with reduced digestive cancer risk, likely via anti-inflammatory pathways and modulation of bile-acid metabolism [48]. These protective effects likely operate through anti-inflammatory pathways and modulation of bile-acid metabolism. In addition, our findings indicate that higher dairy intake may be linked to lower risks of rectal and certain other digestive cancers, which is mechanistically supported by the capacity of calcium and vitamin D to bind carcinogenic bile acids, attenuate colonic epithelial proliferation, and regulate intestinal inflammation [38, 49]. Collectively, these patterns provide biological coherence to the observed performance of the modified ELD score. Nevertheless, Score3 appeared less effective for colon, stomach, and biliary tract cancers after covariate adjustment. Regarding colon cancer, the unadjusted association had a relatively low E-value, suggesting that even modest unmeasured confounding could explain the observed effect. After adjusting for lifestyle and metabolic factors, the association was markedly attenuated. Combined with minimal absolute risk differences and low discriminative performance (C-index), this indicates that the initial crude association was likely confounded by correlated risk factors rather than reflecting a direct effect of the diet score. For stomach cancer, the lack of association may reflect that the score, although including processed meat, may not fully capture other dietary exposures relevant to gastric cancer, particularly traditional high-salt preserved foods or nitrosamine- or PAH-containing items, which may interact with Helicobacter pylori to increase cancer risk [50, 51]. Additionally, the score does not account for cooking methods, potentially introducing exposure misclassification and attenuating true associations. For biliary tract cancer, differences in risk across score groups were minimal, likely due to its relatively low incidence. Among the numerous risk factors for digestive cancers, genetic predisposition plays an important role [52]. In rectal cancer, individuals with higher dietary scores experienced lower cancer risk across different genetic risk strata, with a significant additive interaction between PRS and the dietary score. This finding suggests that individuals with elevated genetic susceptibility may particularly benefit from adherence to the modified dietary pattern. For esophageal cancer, individuals with high genetic risk in the low-score group had the greatest risk elevation compared with high-score counterparts. However, for esophageal, liver, and pancreatic cancers, neither multiplicative nor additive interactions between the dietary score and PRS were statistically significant, likely reflecting that genetic susceptibility and dietary exposures operate through largely independent pathways. Emerging evidence indicates that sex and smoking also influence digestive cancer risk [53, 54]. Our analyses showed that the predictive performance of Score3 varied across subgroups defined by these factors. Sex-stratified analyses revealed that Score3 had stronger discriminatory ability for esophageal and pancreatic cancers among females, and for liver and rectal cancers among males, suggesting that metabolic and inflammatory pathways linking diet to carcinogenesis may differ by sex. Smoking status also modified associations: among current or former smokers, the difference in cancer risk between low and high Score3 categories was larger for esophageal, liver, and pancreatic cancers. In contrast, rectal cancer risk showed minimal variation by smoking status, indicating limited effect modification. These results underscore that dietary risk scores may not perform uniformly across subgroups and highlight the need to consider sex and smoking when interpreting diet–cancer associations. There are several important limitations to this study. Firstly, our analyses were restricted to the UKB cohort, which may limit the external validity of the findings in more heterogeneous populations. Importantly, the UKB participants are generally healthier and less ethnically diverse than the general UK population, and differ substantially from global populations, which further constrains the generalizability and real-world applicability of the proposed dietary score. Secondly, although multiple 24-hour dietary recalls were used, these short-term assessments are inherently prone to recall error and may not adequately reflect long-term habitual diet; the lack of repeated long-term measurements also hampers our ability to capture within-person dietary changes and to address regression dilution bias. Moreover, excluding individuals with fewer than two dietary assessments or with missing covariate data may disproportionately remove participants with poorer health or more irregular dietary behaviors, potentially introducing selection bias and leading to conservative estimates of associations. Thirdly, covariates were measured only at baseline, and subsequent lifestyle, behavioral, or metabolic changes during follow-up could introduce time-dependent misclassification. Fourthly, although extensive adjustment was undertaken, key unmeasured factors—such as detailed physical activity assessments, gut microbiome composition, medication use, and other lifestyle or metabolic traits—could contribute to residual confounding, even though the E-value analysis suggests that strong unmeasured confounding would be required to explain most associations. Nevertheless, the interplay of several moderate unmeasured confounders could collectively bias the results in unknown directions. Fifthly, the possibility of reverse causation remains, as early subclinical disease may have influenced dietary behaviors despite the exclusion of early incident cancer cases. Sixthly, outcome misclassification related to cancer registry data, particularly incomplete capture of cancer-specific deaths, may have led to under-ascertainment of events. If such misclassification is non-differential with respect to dietary exposure, it would likely attenuate the hazard ratios and reduce the statistical power to detect true associations. Finally, the modified ELD score has not yet undergone external calibration or independent validation, and its potential use as a dietary risk-stratification tool requires confirmation in other cohorts. Without such validation, its generalizability and performance in different populations with varying dietary patterns and cancer incidence rates remain uncertain. These limitations highlight the need for future studies incorporating repeated long-term dietary assessments, more comprehensive measurement of confounders, time-updated covariate information, and validation across diverse populations to further refine and evaluate the predictive utility of modified ELD-based dietary scores. Conclusions In conclusion, higher adherence to the modified ELD dietary pattern was associated with a lower incidence of digestive cancers, particularly esophageal, liver, rectal, and pancreatic cancers. This inverse association was most pronounced among individuals with high genetic predisposition to rectal cancer, suggesting that dietary modification may be relevant even in genetically susceptible populations. Overall, these findings support the potential utility of the modified ELD score in evaluating digestive cancer risk and may inform dietary strategies for primary prevention. Supplementary Information Supplementary Material 1: Table S1. International Classification of Disease Codes used to identify cancer in the UK Biobank cohort. Table S2. Data fields and date of cancer diagnosis in the UK Biobank cohort. Table S3. Definition of portion size and food items used in this study. Table S4. Summary results of 8 independent SNPs used for PRS construction of rectal cancer. Table S5. Summary results of 22 independent SNPs used for PRS construction of pancreatic cancer. Table S6. Summary results of 14 independent SNPs used for PRS construction of esophagus cancer. Table S7. Summary results of 5 independent SNPs used for PRS construction of liver cancer. Table S8. Cut-offs for each component of the EAT-Lancet diet score(Score2). Table S9. Baseline characteristics of participants in the UK Biobank (N=111785). Table S10. components of cancer-specific Score3 and coefficients in LASSO-Cox regression. Table S11. C-index (95% CI) and ΔC-index (Score3 vs Score1) for predictive discrimination of dietary scores (Score1–Score3) across digestive cancer types. Table S12. Internal validation of the cancer-specific Score3: bootstrap correction, train-test split, and cross-validation performance. Table S13. Net reclassification improvement (NRI) for dietary scores (Score2 and Score3) versus Score1 across seven digestive cancer types. Table S14. Multiplicative and additive interactions between dietary scores and genetic risk on four digestive cancers incidence, including RERI (Relative Excess Risk due to Interaction) and AP (Attributable Proportion due to Interaction). Figure S1 Associations between Score1–Score3 (cancer-specific) and risks of digestive cancers. Figure S2 Flexible 10-year calibration curves of Score3 (cancer-specific) for predicting survival probability of seven digestive cancers. Figure S3 Cumulative risk of other three digestive cancers during follow-up according to Score3 (cancer-specific). Figure S4 Dose-response association of the risks of other three digestive cancers with Score3 (cancer-specific). Figure S5 Associations between Score3 (cancer-specific) and risks of seven digestive cancers. Figure S6 Associations of risks of other three digestive cancers with Score3 (cancer-specific), stratified by PRS subgroups. Figure S7 Associations of Score3 (cancer-specific) with risks of four digestive cancers, stratified by sex subgroups. Figure S8 Associations of Score3 (cancer-specific) with risks of four digestive cancers, stratified by smoking subgroups Acknowledgements We are grateful to the participants and the team of the UK Biobank study. Declaration of generative AI in scientific writing None. Abbreviations UK Biobank EAT-Lancet diet Polygenic risk score Hazard ratio Townsend deprivation index Body mass index Relative excess risk due to interaction Attributable proportion Authors’ contributions DW, JH, JC, and ZW conducted data curation, formal analysis, software development, and wrote the original draft. JJ performed data curation and visualization. YY carried out data curation and validation. RC, CH, and SZ were responsible for funding acquisition and manuscript review and editing. YJ contributed to conceptualization, methodology, project administration, and supervision. All authors reviewed the final manuscript. Funding This work was supported by the National Natural Science Foundation of China [grant numbers 82372859 (to S.Z.), 82072639 and 82372858 (to R.C.), 82203691 (to C.H.)]; the Special Fund of “Dengfeng Plan” of Guangdong Provincial People’s Hospital, China [grant numbers KJ012019509 and DFJH2020027 (to R.C.)]; the National Key Clinical Specialty Construction Project [2021–2024, grant number 2022YW030009 (to R.C.)]; the Guangzhou Science and Technology Bureau Basic and Applied Research Project [grant numbers 2025A04J4765 (to S.Z.), 2024A1515013216 (to C.H.)]. Data availability This study was carried out utilizing resources from the UK Biobank, with application number 85224. The UK Biobank is an open-access resource and bona fide researchers can apply to use the UK Biobank dataset by registering and applying at http://ukbiobank.ac.uk/register-apply/. The other data, models, and the code that support the findings of this study are available from the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate UK Biobank has ethics approval from the North West Multi-Centre Research Ethics Committee (REC reference: 21/NW/0157). Informed consent was appropriately obtained from participants, and ethical approval was granted by the UK Biobank. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Daixin Wu, Junwei Huang, Juncheng Chen and Zhengyu Wu contributed equally to this work. Contributor Information Yong Jiang, Email: yongjiang29@qq.com. Shangyou Zheng, Email: zhengshangyou@gdph.org.cn. References Associated Data Supplementary Materials Supplementary Material 1: Table S1. International Classification of Disease Codes used to identify cancer in the UK Biobank cohort. Table S2. Data fields and date of cancer diagnosis in the UK Biobank cohort. Table S3. Definition of portion size and food items used in this study. Table S4. Summary results of 8 independent SNPs used for PRS construction of rectal cancer. Table S5. Summary results of 22 independent SNPs used for PRS construction of pancreatic cancer. Table S6. Summary results of 14 independent SNPs used for PRS construction of esophagus cancer. Table S7. Summary results of 5 independent SNPs used for PRS construction of liver cancer. Table S8. Cut-offs for each component of the EAT-Lancet diet score(Score2). Table S9. Baseline characteristics of participants in the UK Biobank (N=111785). Table S10. components of cancer-specific Score3 and coefficients in LASSO-Cox regression. Table S11. C-index (95% CI) and ΔC-index (Score3 vs Score1) for predictive discrimination of dietary scores (Score1–Score3) across digestive cancer types. Table S12. Internal validation of the cancer-specific Score3: bootstrap correction, train-test split, and cross-validation performance. Table S13. Net reclassification improvement (NRI) for dietary scores (Score2 and Score3) versus Score1 across seven digestive cancer types. Table S14. Multiplicative and additive interactions between dietary scores and genetic risk on four digestive cancers incidence, including RERI (Relative Excess Risk due to Interaction) and AP (Attributable Proportion due to Interaction). Figure S1 Associations between Score1–Score3 (cancer-specific) and risks of digestive cancers. Figure S2 Flexible 10-year calibration curves of Score3 (cancer-specific) for predicting survival probability of seven digestive cancers. Figure S3 Cumulative risk of other three digestive cancers during follow-up according to Score3 (cancer-specific). Figure S4 Dose-response association of the risks of other three digestive cancers with Score3 (cancer-specific). Figure S5 Associations between Score3 (cancer-specific) and risks of seven digestive cancers. Figure S6 Associations of risks of other three digestive cancers with Score3 (cancer-specific), stratified by PRS subgroups. Figure S7 Associations of Score3 (cancer-specific) with risks of four digestive cancers, stratified by sex subgroups. Figure S8 Associations of Score3 (cancer-specific) with risks of four digestive cancers, stratified by smoking subgroups Data Availability Statement This study was carried out utilizing resources from the UK Biobank, with application number 85224. The UK Biobank is an open-access resource and bona fide researchers can apply to use the UK Biobank dataset by registering and applying at http://ukbiobank.ac.uk/register-apply/. The other data, models, and the code that support the findings of this study are available from the corresponding author upon reasonable request.