Our research was based on the Preferred Reported Items for Systematic Reviews and Meta-analyses (PRISMA) checklist (Table S1 in Supplementary material S1) (13). The relevant literatures were searched from Web of Science, Pubmed, and Scopus. We retrieved articles using the following Boolean search terms (“vimentin”) AND (“pancreatic cancer” OR “pancreatic ductal adenocarcinoma”) AND (“immunohistochemistry”; as shown in Table S2 in Supplementary material S2). The literature searching was available until July 8th, 2025. No language restrictions were applied at the search level to avoid “language bias.”

Inclusion criteria: (1) histologically confirmed PDACPDAC; (2) case-control or cohort studies; (3) assessment of vimentin in primary tumor cells; (4) reported prognostic or clinicopathological data.

Exclusion criteria: (1) animal studies (murine/porcine); (2) in vitro studies; (3) systematic-reviews or meta-analysis; (4) studies that data was insufficient to calculate OR or HR value; (5) case reports, editorials, commentaries, and conference abstracts (unless they provided sufficient quantitative data for HR/OR extraction).

Two authors (I-O.C. and B.S.U.) screened titles, abstracts, and full texts for eligibility, using Ryann (14), a web-based tool designed for systematic reviews and meta-analyses, and extracted data from the included studies independently. The process involved: blinded screening (each reviewer independently categorized records as “Include,” “Exclude,” or “Maybe,” with conflicts automatically flagged by Rayyan), and full-text review (screened abstracts were advanced to full-text assessment if marked “Include” or “Maybe” by either reviewer). Extracted data included study characteristics, population demographics, methods of vimentin assessment, and clinical features (tumor stage, lymph node metastasis, histological differentiation, stage, lymphatic invasion, venous invasion, neural invasion). Data on vimentin immunoexpression were extracted as originally reported, scoring was limited to tumor cells when specified, and studies without clear differentiation between tumor and stromal staining were included as such, this methodological variability being recognized as a limitation. Disagreements were resolved by consensus or a third reviewer (A.T-S.). The quality of included studies was assessed using the Newcastle-Ottawa Scale (NOS) for cohort studies (15).

Publication bias was assessed via visual inspection of funnel plots and, where applicable, by Kendall's Tau correlation and Egger's regression test to check for funnel plot asymmetry.

R (version 4.5.1–1.2204.0, R Foundation, Vienna, Austria) with the metafor package (16) and RevMan (version 5.4, Cochrane Collaboration, 2020) were used. Hazard ratios (HRs) with 95% CIs were used to assess the association between OS of PC and vimentin expression. Pooled effect sizes were calculated using Odds Ratios (ORs) with 95% Confidence Intervals (CIs) for categorical outcomes. In instances where HRs were not directly reported, survival data were extracted from Kaplan-Meier curves using WebPlot Digitizer (Version 4.7, Austin, Texas, USA) (17). Meanwhile, logOR or logHR with 95% CIs was applied to observe the correlations clinical features of PC with vimentin expression. Statistical pooling was performed using either fixed-effects or random-effects models. The choice between models was guided by the presence of heterogeneity (I² and Cochran's Q test). Where heterogeneity was low (I² = 0% or p > 0.10), a fixed-effects model using the inverse-variance method was applied. In cases of significant heterogeneity (I² > 0%), a random-effects model was utilized with the Restricted Maximum Likelihood (REML) estimator. REML was chosen over the traditional DerSimonian–Laird approach to provide a more robust estimate of between-study variance (τ²) given the number of studies included. Given the methodological heterogeneity between studies, including differences in immunohistochemical protocols, scoring systems, and cut-off for defining vimentin expression, the classification of “low/high” was performed according to the original definitions of each study. Pooling was performed using these specific classifications to preserve internal validity and avoid reclassification of cases by applying uniform cut-off in a heterogeneous methodological context. The amount of heterogeneity was reported with tau², Q-test for heterogeneity results and the I² statistic. In case of any amount of heterogeneity was detected (tau² > 0, regardless of the results of the Q-test), a prediction interval for the true outcomes was also provided. Studentized residuals and Cook's distances were used to examine whether studies may be outliers and/or influential in the context of the model. Studies with a Cook's distance larger than the median plus 6 times the interquartile range of the Cook's distances are considered to be influential.

A PRISMA flow diagram (Figure 1) was used to document the study selection process for this systematic review and meta-analysis of vimentin in pancreatic cancer. The initial search across three databases: PubMed (N = 140), Web of Sciences (N = 155), and Scopus (N = 148), yielded a total of 443 articles. In the Identification phase, 110 articles were excluded due to duplication, leaving 333 articles for title and abstract screening. This initial screening was performed independently by two reviewers (O-I.C. and B.S.U.) utilizing the Rayyan systematic review software. During the Screening phase, 132 articles were excluded for various reasons, including: 42 non-human in vivo studies (comprising various animal models such as murine, porcine, and aquatic) and 61 in vitro cell-line studies; five literature reviews; and 20 case reports. This resulted in 201 articles undergoing full-text assessment, which was also conducted independently by the two reviewers. Any discrepancies regarding inclusion were resolved through consensus or by arbitration with a third reviewer (A.T-S.). Of these, 192 articles were excluded, primarily because they lacked vimentin expression data (N = 98), did not use immunohistochemistry (N = 67), or did not focus on pancreatic ductal adenocarcinoma (N = 27). Ultimately, a consensus was reached, and nine articles were selected and included in the final quantitative synthesis (meta-analysis), containing 721 PC patients, 558 low expression vimentin and 163 with high expression vimentin. Table 1 contains the characteristics of the included studies (18–26). Stage classification was evaluated using UICC 7th TNM classification of the Union for International Cancer Control in one study (22, 24) or the 6th edition AJCC in two studies (19, 25).

The methodological quality of the studies was assessed using the Newcastle-Ottawa Scale (NOS), with scores ranging from 4 to 9, as shown in Table S3 of Supplementary material S2. The highest quality was observed in the studies conducted by Chouat et al. (19), Maehira et al. (24) and Sánchez-Ramírez et al. (25), each achieving 9/9 points through rigorous cohort selection, appropriate multivariable adjustment and complete follow-up reporting. The studies by Kokumai et al. (22), Lee et al. (23) and Wang et al. (18) scored 8/9. These studies included well-defined cohorts, applied appropriate exposure assessment methods and performed adjustments for relevant confounders. However, they did not clearly report the proportion of patients lost to follow-up or provide an explicit statement regarding the percentage of patients followed until the endpoint. Consequently, although the internal validity of the association appears strong, the uncertainty regarding follow-up losses requires caution when interpreting the risk of bias related to follow-up. Two studies reached only 7/9 points. The study by Xu et al. (26) lost points in the comparability domain due to insufficient details on multivariable adjustment for key confounders, as well as in the adequacy of follow-up, since the proportion of patients lost was not reported, while Javle et al. (21) applied a limited multivariable model and provided insufficient follow-up data, raising concerns about potential bias. Guo et al. (20) obtained the lowest score (4/9) reflecting modest methodological quality. The outcome of interest was assessed at the time of resection rather than as a subsequent event, no multivariable adjustments were performed, and analyses were restricted to simple statistical tests. The authors explicitly acknowledged these limitations and did not provide survival data.

A meta-analysis was performed on three studies (n = 263 patients) to investigate the association between vimentin expression and tumor stage in PC, including 192 patients with low expression vimentin and 71 patients with high expression vimentin, as shown in Figure 2A. The observed logOR ranged from −0.658 to 0.858, with the majority of estimated being positive (67%). The estimated average logOR based on fixed-effects model was −0.10 (95% CI: −0.87 to 0.66). Therefore, the average outcome did not differ significantly from zero (z = −0.27, p = 0.791). Based on the absence of significant heterogeneity as determined by the Q-test (Q(2) = 1.79, p = 0.41, I² = 0%), a fixed-effects model was employed. This confidence interval crosses the line of no effect (0), indicating that there is no statistically significant association between vimentin expression and T-stage in this meta-analysis. Clinically, these data suggest that while vimentin is a marker of cellular migration, its expression does not correlate with primary tumor size or local invasiveness (T-stage) in PDAC, indicating it may be a more specific indicator of metastatic spread rather than local tumor growth.

Neither the rank correlation nor the regression test indicated any funnel plot asymmetry (p = 1.0 and p = 0.41, respectively). The funnel plot in Figure 2B appears symmetrical, with studies distributed relatively evenly around the pooled effect. The studies, despite their varying precision, fall within the expected funnel shape. This symmetry suggests that there is no apparent publication bias in the included literature regarding the association between vimentin expression and T-stage. The meta-analysis results are likely robust against bias, though the small number of studies means caution is warranted in overinterpreting symmetry.

A meta-analysis of five studies was conducted to assess the association between tumor-cell vimentin expression and lymph node metastasis, including in the analysis (340 patients with low expression vimentin and 103 patients with high expression vimentin), as shown in Figure 3A. The observed logOR ranged from 0.182 to 2.498, with all of the estimated being positive. The estimated average logOR based on the fixed-effects model was 0.58 (95% CI: 0.083–1.078), with the average outcome significantly different from zero (z = 2.29, p = 0.022), indicating a statistically significant positive association between vimentin expression and the presence of lymph node metastasis. The significant association with nodal involvement suggests vimentin could help identify patients at high risk for occult lymphatic metastasis who might benefit from more extensive lymphadenectomy.

The forest plot visually confirms this result. While the individual studies—Guo et al. (20), Kokumai et al. (22), Maehira et al. (24), and Wang et al. (11)—show wide confidence intervals that cross the line of no effect, the study by Xu et al. (26) demonstrates a significant positive association [2.50, 95% CI: (0.33, 4.67)]. The overall pooled result is driven by the consistent, albeit non-significant, trend toward a positive association across most studies, culminating in a statistically significant combined effect. The minimal heterogeneity (I² = 0%) suggests that the true effect is consistent across the included studies.

According to the Q-test, there was no significant amount of heterogeneity in the true outcomes (Q(4) = 3.91, p = 0.42, I² = 0%). The funnel plot does not indicate significant asymmetry, as shown in Figure 3B and Table S4 of Supplementary material S2, supporting the robustness of the meta-analysis results, assuming no other biases are present.

A meta-analysis was conducted on three studies to examine the association between tumor-cell vimentin expression and metastasis in PC, including 182 patients with low expression vimentin and 81 patients with high expression vimentin, as shown in Figure 4A. The observed logOR ranged from −1.082 to 0.67, with the majority of estimated being negative (67%). The estimated average logOR based on fixed-effects model was 0.08 (95%CI: −0.95 to 1.10), with average outcome different significantly from zero (z = 0.15, p = 0.88). The confidence interval for the pooled effect size crosses the line of no effect (0), indicating that the association between vimentin expression and metastasis is not statistically significant in this analysis. The confidence intervals for all three studies are wide and cross the line of no effect, meaning that none of the individual studies demonstrated a statistically significant association. The lack of association with distant metastasis suggests that vimentin-driven EMT may be a more localized driver of lymphatic spread than of systemic hematogenous dissemination.

According to the Q-test, there was no significant amount of heterogeneity in the true outcomes (Q(2) = 1.39, p = 0.498, I² = 0%). The visual inspection of the forest plot suggests also low heterogeneity, as the confidence intervals of the individual studies largely overlap. The funnel plot (Figure 4B) and Table S3 in Supplementary material S2 show no asymmetry, indicating minimal risk of publication bias. While results are robust, the limited study count advises cautious interpretation.

A total of four studies were included in the analysis (298 patients with low expression vimentin and 105 patients with high expression vimentin), as shown in Figure 5A. The observed logOR ranged from −1.082 to 2.84, with half of estimated being negative (50%). The estimated average logOR based on random-effects model was 0.57 (95% CI: −1.02 to 2.16), with average outcome did not differ significantly from zero (z = 0.71, p = 0.48). According to the Q-test, there was significant amount of heterogeneity in the true outcomes (Q(3) = 13.39, p = 0.0039, τ² = 1.95, I² = 80.56%), likely due to the use of different staging systems (UICC 7th vs. AJCC 6th). A 95% prediction interval for the true outcomes is given by −2.59 to 3.74. Hence, although the average outcome is estimated to be positive, in some studies the true outcome may be negative. An examination of the studentized residuals revealed that one study (25) had a value larger than ± 2.394 and may be a potential outlier in the context of this model. According to the rank correlation, the regression test (Table S3 in Supplementary material S2) and the funnel plot asymmetry (Figure 5B), minimal risk of publication bias was observed. While results are robust, the limited study count advises cautious interpretation.

Clinically, these results suggest that vimentin-driven mesenchymal transition is not a reliable predictor of the overall TNM stage grouping, indicating that vimentin expression may occur independently of the cumulative anatomic extent of the disease in PDAC.

A total of three studies were included in the analysis, as shown in Figure 6A. The observed logOR ranged from −1.66 to 1.67, with the majority of estimated being negative (67%). The estimated average logOR based on random-effects model was 0.46 (95%CI: −1.37 to 2.28), with average outcome did not differ significantly from zero (z = 0.49, p = 0.624). According to the Q-test, there was no significant amount of heterogeneity in the true outcomes (Q(2) = 6.96, p = 0.031, τ² = 1.912, I² = 74.57%). From a clinical perspective, this implies that the EMT marked by vimentin is a distinct biological phenomenon that can occur across the spectrum of histological differentiation, rather than being confined to poorly differentiated tumors. A 95% prediction interval for the true outcomes is given by −2.811 to 3.72. Hence, although the average outcome is estimated to be positive, in some studies the true outcome may be negative. The Egger's regression test (Table S3 in Supplementary material S2) indicated funnel plot asymmetry (p = 0.0087) but not the Begg's rank correlation test, which is more conservative and less powered with very few studies, like in our case (p = 0.333). The visual asymmetry (Figure 6B) and Egger's significant result suggest a potential small-study effect or publication bias, but this is a tentative conclusion, given the small number of studies.

A meta-analysis of six studies, including 360 patients with low vimentin expression and 84 patients with high vimentin expression, evaluated the impact of high vimentin expression on overall survival in pancreatic cancer, as shown in Figure 7A. The pooled analysis showed that high vimentin expression was significantly associated with worse OS, with a combined HR of 1.39 (95% CI: 0.11–2.68), based on a random-effects model, as shown in Figure 7A. However, there was substantial heterogeneity among studies (Q(5) = 78.68, τ² = 2.36, I² = 93.64%, p < 0.001), suggesting considerable variation in effect sizes. The overall effect estimate remained statistically significant despite this heterogeneity, as shown by the Z statistic (Z = 2.12, p = 0.034), suggesting that high vimentin expression may be a negative prognostic factor for survival in pancreatic cancer. The increased hazard ratio for overall survival indicates that tumor-cell vimentin status serves as a potent indicator of poor biological behavior and shorter life expectancy in PDAC.

The high heterogeneity is visually apparent in the forest plot, as the individual study confidence intervals do not overlap significantly. For example, the study by Chouat et al. (19) shows a large positive effect (6.44), while the studies by Kokumai et al. (22) and Lee et al. (23) show a negative effect (−0.24). The remaining studies show varying degrees of positive effect. This wide range of outcomes highlights the importance of using a random-effects model, which acknowledges and incorporates this variability into the final pooled estimate. The weighting of each study, as shown in the forest plot, indicates their proportional contribution to the overall effect.

Visual inspection of the funnel plot (Figure 7B) suggested mild asymmetry, with most of the studies on the left side of the funnel (a potential for publication bias), possibly indicating small-study effects. However, Egger's regression test was statistically significant (p = 0.026), suggesting that there is strong evidence of publication bias, as in Table S3 in Supplementary material S2.