The present systematic review was undertaken in full alignment with the PRISMA guidelines for reporting meta-analyses and systematic syntheses (10). The protocol of the present meta-analysis was not prospectively registered in PROSPERO or any other registry. To capture all relevant evidence, a comprehensive search of PubMed (https://pubmed.ncbi.nlm.nih.gov/), EMBASE (https://www.embase.com/) and the Cochrane Library (https://www.cochrane.org/) was performed, encompassing studies published up to November 18, 2025. The retrieval strategy employed predetermined terms related to the LCR ratio. The PubMed search strategy was as follows: [Lymphocyte-to-C-reactive protein ratio(Title/Abstract)] OR [Lymphocyte-C-reactive protein ratio(Title/Abstract)] OR [Lymphocyte/C-reactive protein ratio(Title/Abstract)] OR [Lymphocyte-to-CRP ratio(Title/Abstract)] OR [Lymphocyte-CRP ratio(Title/Abstract)] OR [Lymphocyte/CRP ratio(Title/Abstract)]. This strategy was adapted appropriately for EMBASE and the Cochrane Library according to the syntax and indexing rules of each database. Furthermore, the reference lists of included articles were manually screened to identify any additional work not detected through database queries. To reduce selection bias, two reviewers independently evaluated the literature, and disagreements were adjudicated by consultation with a senior investigator.

Studies were included if they met the following criteria: i) Studies specifically investigating patients with GC, as defined according to the diagnostic criteria used in the original reports; ii) studies evaluating the prognostic significance of pretreatment LCR, with LCR analyzed as a dichotomized variable; iii) studies reporting survival outcomes, including overall survival (OS) and/or disease-free survival (DFS); and iv) studies providing hazard ratios (HRs) with corresponding 95% confidence intervals (CIs), or sufficient data from which these estimates could be obtained. No additional restriction was imposed regarding explicit histological confirmation beyond the diagnostic criteria reported in the original studies. In addition, no minimum follow-up duration was required, provided that extractable survival data were available.

For each study that met the inclusion criteria, all pertinent information was systematically gathered, including the leading author, publication year, duration of the study period, geographical setting, therapeutic modalities applied, overall cohort size, and fundamental demographic variables such as age distribution and sex composition. The thresholds used to categorize LCR values in individual reports were also recorded. When available, HRs with corresponding 95% CIs were preferentially extracted from multivariate Cox regression models to minimize the influence of potential confounding. If multivariate estimates were not available, univariate HRs were extracted. When HRs and 95% CIs were not reported directly but Kaplan-Meier survival curves were available, the log HR and its standard error were estimated using the established methods described by Tierney et al (11). Survival probabilities, numbers at risk, total events and log-rank information were extracted when available to reconstruct the corresponding HR estimates. Multivariate estimates were prioritized over reconstructed or univariate estimates in the quantitative synthesis. When a single publication reported more than one independent and non-overlapping cohort with separately available survival estimates, each cohort was extracted and analyzed as an independent dataset. Specifically, Xiong et al (12) reported both a discovery cohort and a validation cohort, which were treated as two separate cohorts in the quantitative synthesis because they included distinct patient populations and provided separate prognostic data (12).

The methodological robustness of the included observational studies was appraised using the Newcastle-Ottawa Scale (NOS), a widely accepted instrument for evaluating potential bias in non-randomized cohort and case-control research (13,14). Investigations earning six or more points on this scale were regarded as having satisfactory methodological rigor. Data extraction and quality assessment were independently performed and cross-verified by two authors, with any discrepancies resolved by a senior investigator.

All quantitative analyses were performed with Stata (version 18.0; StataCorp LP), and the pooled estimates were illustrated using forest diagrams. To evaluate variability across studies, Cochran's Q statistic was applied together with the I² metric; heterogeneity was regarded as meaningful when the I² value was greater than 50% or when the Q test yielded a P-value below 0.1 (15). When substantial divergence among studies was detected, the DerSimonian-Laird random-effects approach was adopted. By contrast, datasets exhibiting minimal heterogeneity were synthesized using a fixed-effects framework based on inverse-variance weighting. To evaluate whether variations in LCR thresholds influenced the pooled estimates, an exploratory subgroup analysis was performed according to the median cut-off value reported by the included studies. Because no universally accepted LCR threshold has been established for GC, this analysis was considered exploratory.

Potential publication bias was assessed using Begg's rank-based test and Egger's regression methodology (15). To determine the stability of the synthesized effect estimates, sensitivity analyses were performed, in which each study was removed one at a time before recalculating the pooled HRs. Statistical significance was defined by a two-sided P-value of 0.05.

A comprehensive literature search in multiple electronic databases, complemented by manual screening of reference lists, yielded 491 articles. After the elimination of 126 duplicates, the remaining articles were evaluated at the title and abstract levels, leading to the exclusion of 341 articles that did not fulfill the established selection criteria. Full-text review was subsequently performed for 24 articles, of which 15 were removed because they did not meet the eligibility requirements related to study design, participant characteristics, or reporting of survival endpoints. In total, eight publications comprising nine independent cohorts were included in the quantitative synthesis (12,16–22) (Fig. 1). Of note, Xiong et al (12) reported two independent cohorts, a discovery cohort and a validation cohort, which were analyzed separately because the patient populations were non-overlapping and survival estimates were reported independently.

A comprehensive summary of the principal features of the included studies is presented in Table I. Altogether, the nine included studies contributed data from 4,429 patients with GC, with individual cohort sizes ranging between 123 and 774 participants. In terms of geographic distribution, four studies were performed in China, four in Japan, and one in Türkiye. This distribution highlights the regional pattern of currently available evidence.

Regarding treatment context, seven studies focused specifically on patients undergoing surgical treatment, whereas the remaining two studies included individuals receiving multimodal therapy. All studies adopted a retrospective observational framework, a design that, despite certain inherent constraints, enabled the extraction of detailed clinical information and laboratory parameters.

Quality assessment based on the NOS yielded scores ranging from 6 to 8, suggesting a generally low risk of bias and reinforcing confidence in the robustness of the synthesized findings. Overall, these study characteristics reflect high methodological quality and provide essential context for interpreting subsequent meta-analysis results.

A total of nine studies fulfilled the eligibility criteria, contributing data from 4,429 individuals diagnosed with GC. In each dataset, patients were categorized into high- and low-LCR groups based on the cut-off values specified in the respective studies. When these data were synthesized, a pronounced survival disadvantage emerged for individuals with reduced LCR values. The pooled hazard ratio for OS was 2.24 (95% CI, 1.97–2.55; P<0.001), indicating that the mortality risk in low-LCR patients was more than twice that in high-LCR patients (Fig. 2).

To assess the consistency of these findings across studies, heterogeneity was examined using both Cochran's Q statistic and the I² index. Neither metric suggested meaningful variability: The Q test showed no statistical significance, and the I² value was 0% (P=0.453). This absence of detectable heterogeneity suggested that the effect estimates were highly concordant among included analyses. Consequently, a fixed-effects modeling approach was selected as the most appropriate framework for deriving the final pooled results.

To evaluate the robustness of the pooled effect, a leave-one-out sensitivity analysis was conducted, in which each study was removed sequentially and the overall HR recalculated. This procedure tested whether any single dataset disproportionately affected the combined estimate. The results showed that the exclusion of individual studies produced negligible changes in both effect size and direction, indicating strong internal consistency across the included analyses (Fig. 3A). Although minor fluctuations were observed, the overall pattern remained stable. Removing the study by Miyatani et al (20) produced the lowest risk estimate (HR=2.16; 95% CI: 1.89–2.50; P<0.001), whereas omission of the dataset from Wang et al (22) resulted in the highest estimate (HR=2.33; 95% CI: 2.01–2.69; P<0.001). Importantly, all recalculated values were directionally aligned and statistically significant, demonstrating that no single investigation materially shaped the final outcome. Collectively, these findings confirm that the observed association between low LCR levels and unfavorable OS in GC is highly robust and not driven by outlier studies. This reinforces the methodological reliability of the quantitative synthesis.

To further evaluate potential distortions arising from selective publication, Begg's rank correlation and Egger's regression tests were applied. Neither test identified significant evidence of bias, with P-values of 0.466 and 0.350, respectively. This conclusion was supported by the symmetrical appearance of the funnel plot, which likewise indicated the absence of publication-related skewness (Fig. 3B).

To determine whether the prognostic value of LCR was influenced by analytical approach, region, or treatment strategy, a series of predefined subgroup analyses were performed. When stratified by the source of effect estimates, the association between reduced LCR and adverse OS remained highly consistent. Effect sizes derived from univariate analyses indicated a substantial survival disadvantage for patients with low LCR levels (HR=2.66; 95% CI: 2.04–3.46; P<0.001), and this relationship persisted even after adjustment for confounding factors in multivariate models (HR=2.12; 95% CI: 1.83–2.46; P<0.001) (Fig. 4A). These findings suggest that the prognostic relevance of LCR is robust and not dependent on statistical modeling choices.

Geographically stratified analyses further reinforced this conclusion. In studies conducted in Japan, low LCR values were strongly associated with inferior OS, with a pooled HR of 2.31 (95% CI: 1.85–2.87; P<0.001). A similar magnitude of risk was observed among cohorts from China, where low LCR remained a significant predictor of poor survival (HR=2.19; 95% CI: 1.84–2.60; P<0.001) (Fig. 4B). The consistent effect across regions suggests that the prognostic value of LCR transcends geographic and demographic variability.

It was further examined whether treatment modality and LCR cut-off value influenced the prognostic association between LCR and OS. In the treatment-based subgroup analysis, low LCR remained significantly associated with unfavorable OS both in patients undergoing gastrectomy (HR=2.31; 95% CI: 1.98–2.70; P<0.001) and in those receiving other treatment modalities (HR=2.07; 95% CI: 1.63–2.64; P<0.001; Fig. 5A). Similarly, when studies were stratified by the median LCR cut-off value, the association remained significant in both the high cut-off group (>6451.6: HR=2.31; 95% CI: 1.85–2.87; P<0.001) and the low cut-off group (≤6451.6: HR=2.20; 95% CI: 1.88–2.59; P<0.001; Fig. 5B).

Collectively, the subgroup analyses provide compelling evidence that the prognostic effect of low LCR is consistent across diverse analytical frameworks, geographic populations, cut-off and clinical management strategies. These findings highlight the potential of LCR as a broadly applicable biomarker for risk stratification in GC.

A total of five eligible studies, encompassing 2,987 individuals diagnosed with GC, were included to investigate the prognostic significance of pretreatment LCR in relation to DFS. In each cohort, patients were stratified into high- and low-LCR subgroups based on the cut-off values defined by the original studies. When these datasets were quantitatively synthesized, low pretreatment LCR was consistently associated with a markedly increased risk of earlier recurrence. The pooled HR for DFS was 2.11 (95% CI: 1.78–2.51; P<0.001), indicating that patients with reduced LCR levels experienced substantially shorter DFS compared with those with higher levels (Fig. 6A).

Between-study heterogeneity was minimal. Cochran's Q test did not reach statistical significance, and the I² statistic was 0% (P=0.301), demonstrating that the effect estimates were highly concordant across all included cohorts. Owing to this strong consistency, a fixed-effect model was adopted to generate the final summary estimate. These findings collectively suggest that LCR may serve as a simple and clinically accessible biomarker capable of identifying patients at elevated risk for early relapse.

Sensitivity analyses further supported the robustness of the pooled results. Sequential omission of individual studies revealed no material changes in the magnitude or direction of the hazard ratio, confirming that the association between reduced LCR and inferior DFS was not driven by any single dataset (Fig. 6B).

Evaluation of publication bias using Begg's rank correlation and Egger's regression also demonstrated no statistically significant evidence of small-study effects (Begg's P=0.462; Egger's P=0.391). The symmetrical appearance of the funnel plot provided additional reassurance that the synthesized results were unlikely to be influenced by selective reporting practices (Fig. 6C). Geographically stratified analyses further reinforced this conclusion (Fig. 7).

Overall, these findings reinforce the prognostic value of pretreatment LCR for predicting disease recurrence in GC and highlight its potential role as a readily applicable tool for risk stratification.