The analysis performed in this study was conducted following the guidelines of the PRISMA statement (16). On April 23, 2023, a comprehensive literature search was carried out using PubMed, EMBASE, and the Cochrane Library. Various search terms, including MeSH terms and keywords, were used to retrieve relevant studies, such as “Immune Checkpoint Inhibitors [MeSH]”, “PD-1 Inhibitors”, “PD-L1 Inhibitors”, “CTLA-4 Inhibitors”, “Pembrolizumab”, “Nivolumab”, “Atezolizumab”, “Ipilimumab”, “Avelumab”, “Tremelimumab”, “Durvalumab”, “Cemiplimab”, “Prognostic Nutritional Index”, “PNI”. Search restricted to English literature. A detailed description of the search strategies is provided in Table S1 . Additionally, gray literature was searched using Google Scholar, and the reference lists of eligible studies were screened manually.

In our study, we strictly included research articles that met the following criteria: patients who were diagnosed with GIC underwent treatment with ICIs, and the prognostic value of PNI was evaluated. Furthermore, these articles reported on at least one of the following outcomes: overall survival (OS), progression-free survival (PFS), objective response rate (ORR), and disease control rate (DCR). We excluded conference abstracts, comments, and case reports from our analysis. In situations where studies had overlapping patients, we prioritized those with the most comprehensive data and robust methodology (3).

In this study, we gathered diverse information from the chosen articles, including the names of the authors, year of publication, duration and location of the study, drugs used for treatment, cancer type, sample size, patient age and gender, and relevant cut-off values and outcomes. We placed greater emphasis on obtaining data from multivariate analyses of hazard ratios (HR) compared to univariate analyses. We also employed the Newcastle-Ottawa Scale (NOS) to appraise the quality of observational studies and classified those with a NOS score of 6 or above as high-quality literature (17). All of the above steps were completed and cross-checked independently by two authors, with decisions sought from the corresponding author on points of dispute.

Statistical analyses were performed using Stata 15.0 software. To assess heterogeneity, we utilized the chi-squared test. We employed a random effects model when the p-value < 0.1 or the I² statistic was > 50%, and a fixed effects model otherwise. We estimated publication bias using both Egger’s and Begg’s tests, and if bias was detected, we utilized the “trim and fill” method to evaluate the influence of the bias on the pooled results. Furthermore, we conducted a sensitivity analysis by excluding each study independently to assess the robustness of the results. p < 0.05 was considered statistically significant.

After excluding duplicates and screening titles and abstracts, we identified 25 articles for full-text evaluation, among which 17 met the eligibility criteria, resulting in a total of 2883 patients (18–34). The PRISMA flowchart in Figure 1 depicts the study selection process. Table 1 provides a comprehensive overview of the characteristics of the eligible studies. We assessed the risk of bias using the Newcastle-Ottawa Scale (NOS), with scores ranging from 6 to 8, indicating a low risk of bias in all included studies. Of the 17 studies, 16 were retrospective, and one was prospective. Four studies were in esophageal squamous cell carcinoma (ESCC) patients; five studies were in gastric cancer (GC) patients; and four studies were in hepatocellular carcinoma (HCC) patients. In addition, there is one study on esophageal cancer (EC) patients, one study on intrahepatic cholangiocarcinoma (ICC) patients, one study on upper gastrointestinal cancer (UGIC) patients and one study on biliary tract cancers (BTC) patients.

Through the analysis of data from 14 studies with 2293 patients, we aimed to explore the correlation between PNI levels and OS in ICI-treated GIC patients. Considering no significant heterogeneity across the studies (I² = 33.2%, p = 0.109), we employed a fixed-effects model to estimate the pooled HR. The results revealed that high PNI levels were significantly related to longer OS (HR: 0.530, 95% CI: 0.456-0.616, p < 0.001), as depicted in Figure 2 .

We performed a subgroup analysis based on cancer type, cut-off values, and the Cox model. We found that a high PNI was associated with a better prognosis in patients with EC, GC, HCC, or BTC ( Figure 3 ). Differences in PNI cutoff values do not affect the correlation between PNI and OS in ICI-treated GIC patients ( Figure 4 ). Both univariate and multivariate analyses confirmed the above findings ( Figure S1 ).

We also investigate the correlation between PNI and PFS in GIC patients treated with ICIs using data from 12 studies with 1733 patients. We demonstrated that patients with high PNI levels had a lower risk of progression (HR: 0.740, 95% CI: 0.649-0.844, p < 0.001, Figure 5 ). No significant heterogeneity was observed (I² = 22.1%, p = 0.226); we used a fixed-effects model for our analysis.

The results of the subgroup analysis showed that the relationship between high PNI levels and longer PFS was consistent in GIC patients with EC, GC, and HCC ( Figure 6 ). Notably, PNI predicted PFS in patients with GIC only when the cut-off value was between 40 and 45 ( Figure 7 ).

The findings of the multivariate analysis also support these findings; although the univariate analysis revealed that PNI was not associated with patient PFS, we consider the conclusions drawn from the former to be more reliable due to its more methodologically rigorous nature and larger number of inclusions ( Figure S2 ).

Subsequently, we conducted an analysis to investigate the correlation between PNI levels and response to ICI therapy in GIC patients. The presence of notable heterogeneity was not observed in the results presented in Figures 8A, B , and hence a fixed-effect model was implemented. Our findings revealed that GIC patients with high PNI levels had a higher ORR (5 studies with 1430 patients, OR: 1.622, 95% CI: 1.251-2.103, p < 0.004, Figure 8A ) and DCR (7 studies with 1551 patients, OR: 1.846, 95% CI: 1.428-2.388, p < 0.001, Figure 8B ).

To assess the robustness of the findings, a sensitivity analysis was conducted by iteratively excluding each study and examining the impact on the overall results. Our analysis indicated that the exclusion of any individual study did not significantly affect the pooled HR for OS. Specifically, the HR estimates for OS ranged from 0.474 (95% CI: 0.398-0.566) when excluding Muhammed et al., 2021 (25) to 0.547 (95% CI: 0.469-0.637) when excluding Yang et al., 2022 (34), as depicted in Figure 9A . Furthermore, the sensitivity analysis demonstrated that the removal of any individual study did not significantly impact the overall results for PFS. The range of HR values varied from 0.662 (95% CI: 0.547-0.800) after excluding Persano et al., 2023 (27) to 0.754 (95% CI: 0.659-0.862) after excluding Yang et al., 2022 (33) ( Figure 9B ). Similarly, sensitivity analysis showed that removing any of the studies did not affect the ORR and DCR results ( Figures 10A, B ).

We performed Begg’s and Egger’s tests to evaluate the potential publication bias in our meta-analysis. The findings indicated no considerable publication bias for ORR (Egger’s test: p = 0.241, Begg’s test: p = 0.230) and DCR (Egger’s test: p = 0.086, Begg’s test: p = 0.081). Nevertheless, we detected publication bias in OS (Egger’s test: p = 0.003, Begg’s test: p = 0.021) and PFS (Egger’s test: p = 0.027, Begg’s test: p = 0.064) based on Egger’s test. To address this issue, we utilized the trim and fill method to estimate the number of potential missing studies in OS. The results showed no change in pooled HR without the missing study being incorporated ( Figure 11 ).