This systematic review and meta-analysis was conducted in compliance with the Preferred Reporting Items for Systematic review and Meta-analysis (PRISMA) guideline (21). Candidate studies investigating the prognostic value of blood PD-L1 markers in NSCLC patients undergoing ICIs therapy were comprehensively searched in PubMed, EMBAS, Web of Science, and Cochrane Library from inception to November 30, 2023. The following search terms were applied: (exosome OR exosomal OR extracellular vesicles OR soluble OR circulating tumor cells OR neoplastic circulating cells OR CTCs) AND (programmed cell death ligand 1 OR PD-L1 OR CD274 OR sPD-L1 OR exoPD-L1) AND (NSCLC OR non-small cell lung cancer OR lung adenocarcinoma OR lung cancer OR lung carcinoma). Language was not restricted. The reference lists of included studies and relevant reviews were manually reviewed for additional eligible articles.

The eligibility of retrieved studies was judged according to the following criteria (1): NSCLC patients were treated with ICIs alone or in combination with other therapies; (2) at least one of blood PD-L1, including sPD-L1, CTC PD-L1 and exoPD-L1, were detected; (3) pre-treatment level, post-treatment level or dynamic change of blood PD-L1 was determined as prognostic marker; (4) the association of PD-L1 marker with progression-free survival (PFS) and/or overall survival (OS) was investigated. Case reports, reviews, meta-analyses, conference abstracts, and functional studies were excluded. If recruiting various cancers including NSCLC, the study was included only when survival outcomes of NSCLC subgroup were reported. Otherwise, it was discarded. Since we aimed for the prognostic value of functional PD-L1, studies detecting mRNA expression on exosomes were excluded.

Newcastle-Ottawa Scale (NOS) for cohort study was used to assess the quality of included studies (22). The risk of bias with regard to selection, comparability, and outcome was judged. Selection and outcome domains contained 4 and 3 items, respectively, and 0 or 1 score was awarded to each item. In comparability domain, 0, 1 or 2 scores were awarded. Thus, the total score ranged from 0 to 9. A study with ≥7, 4–6, and ≤5 scores were deemed to have a high, moderate, and low quality, respectively. Quality assessment was performed by two independent authors (JC, PL), and conflicts were resolved by a third author (XT).

Two independent authors (JC, PL) extracted the following information of each study: first author, publication year, study design, region, cancer stage, sample size, age, percentage of male, percentage of smoker, type and treatment line of ICIs, CTC enrichment method, PD-L1 detection method, enzyme-linked immunosorbent assay (ELISA) kit, cut-off value, cut-off determination method, time point of marker detection, hazard ratio (HR) with corresponding 95% confidence interval (CI) of survival outcomes (PFS, OS), and follow-up duration. We extracted HR and 95%CI from multivariate analysis if available. Otherwise, HR results from univariate analysis were extracted. If HR was not directly reported, we estimated it after extracting the survival data from Kaplan-Meier curve if available using Engauge Digitizer v12.1. Disagreement was resolved by a third author (XT).

The statistical analysis was conducted by using STATA 16.0 (Stata Corporation, TX, USA). The model for quantitative synthesis was determined by between-study heterogeneity. A fixed-effect model was applied when there was low heterogeneity (I²<50% and Q test P>0.10). Otherwise, a random-effect model was used. The association strength between PD-L1 blood markers was calculated by pooled HR with 95%CI. The prognostic values of pre-treatment, post-treatment, and dynamic change of PD-L1 blood markers in predicting PFS and OS was assessed separately. Subgroup analysis and meta-regression were performed to explore the potential source of heterogeneity. Subgroup analyses were stratified by study design, region, PD-L1 detection method, ELISA kit, CTC enrichment method, cut-off value, cut-off determination method, HR analysis, HR extraction, and sample size. Meta-regression analysis was done to explore the modulation effect of baseline variables, including median age, percentage of males, percentage of smokers, sample size, percentage of first-line immunotherapy, and cut-off value, on prognostic value of PD-L1 blood markers. Funnel plot was viewed for symmetry and Egger’s test was performed to assess potential publication bias. In case of potential publication bias, we implemented a trim-and-fill analysis to explore the impact of publication bias on pooled results. P<0.05 indicated statistical significance.

As illustrated in Figure 1 , we obtained a total of 487 articles by comprehensive literature search and discarded 440 after reviewing titles and abstracts as they did not fulfill the inclusion criteria. We further reviewed the full texts of the remaining 47 studies for eligibility and excluded 21 with the following reasons: survival outcomes not investigated (n=5), non-ICI treatments (n=6), various treatments including ICIs (n=2), ICIs treatment unknown (n=2), no NSCLC patients (n=1), survival results not provided (n=3), survival analysis at specimen level (n=1), PD-L1 mRNA expression in exosomes (n=1). These articles with detailed exclusion reasons were listed in Supplementary Table 1 . Finally, 26 studies comprising 1606 NSCLC patients receiving ICIs treatment were included in the meta-analysis (16, 17, 19, 20, 23–44). The prognostic significance of sPD-L1, CTC PD-L1 and exoPD-L1 were investigated in 16, 9 and 3 studies, respectively. The correlations of survival outcomes with pre-treatment blood PD-L1 markers, post-treatment markers and dynamic changes from baseline were reported by 21, 5 and 7 studies, respectively. There were 4 retrospective studies and 22 prospective studies. These studies were conducted in East Asian countries (n=12), European countries (n=12) or America (n=2). The sample size ranged greatly from 14 to 233. PFS outcome was analyzed in 26 studies while OS outcome was analyzed in 19 studies. According to NOS, all included studies had high methodological quality. The baseline characteristics of all studies included in meta-analysis were summarized in Table 1 .

For pre- or post-treatment sPD-L1 and exoPD-L1, cut-off value of PD-L1 concentrations was determined to categorize high and low groups. For dynamic change, cut-off value of change fold was set to segregate patients with up-regulation and without up-regulation after immunotherapy. For pre- or post-treatment CTC PD-L1, cut-off value was used to define PD-L1⁺ CTCs. Seven studies determined the cut-off value by using the median value of PD-L1 concentrations. Nine studies established the cut-off value as the optimal point showing the best performance by various analyses, such as receiver operating characteristic (ROC) curve, Classification and regression tree (CART) analysis, and or log-rank test. Details regarding CTC enrichment, PD-L1 detection, cut-off determination, HR analysis and extraction, and follow-up duration were shown in Supplementary Table 2 .

The association between pre-treatment sPD-L1 and PFS was investigated in 11 studies involving 1011 NSCLC patients treated with ICIs. There was substantial level of between-study heterogeneity (I^{2 =} 60.3%, P=0.005), and a random-effect model was used. Pooled analysis demonstrated patients with high pre-treatment sPD-L1 concentrations had significantly worse PFS than those with low concentrations (HR=1.49, 95%CI 1.13–1.95, P=0.004, Figure 2 ). The correlation of pre-treatment sPD-L1 with OS was analyzed in 10 studies comprising 995 patients. Using a random-effect model, meta-analysis revealed significantly shorter OS in patients with high sPD-L1 levels compared with those with low sPD-L1 levels (HR=1.83, 95%CI 1.25–2.67, P=0.002, Figure 3 ). Further analysis stratified by baseline characteristics were performed ( Table 2 ). There were no significant between-subgroup differences (all P values >0.05), indicating that baseline features were not the main sources of heterogeneity.

Post-treatment sPD-L1 levels were determined for association with PFS and OS in 3 (153 patients) and 4 (177 patients) studies, respectively. High post-treatment sPD-L1 was significantly correlated with worse PFS (HR=2.09, 95%CI 1.40–3.12, P<0.001, Supplementary Figure 1 ) and OS (HR=2.60, 95%CI 1.09–6.20, P=0.032, Supplementary Figure 2 ) compared with low post-treatment sPD-L1.

The prognostic value of dynamic changes of sPD-L1 concentrations was evaluated in 5 studies including 139 NSCLC patients. Up-regulation of sPD-L1 levels from baseline to several cycles of immunotherapy was not significantly associated with survival outcomes (PFS: HR=0.76, 95%CI 0.47–1.23, P=0.259; OS: HR=0.68, 95%CI 0.39–1.19, P=0.117; Supplementary Figure 3 ).

Meta-analysis encompassing 7 studies (238 patients) using a random-effect model revealed a tendency of pre-treatment PD-L1⁺ CTCs to be correlated with favorable PFS at marginal level of significance (HR=0.63, 95%CI 0.39–1.02, P=0.062, Figure 4 ). By pooling together 4 studies (158 patients), we found significantly improved OS favoring patients with pre-treatment PD-L1⁺ CTCs (HR=0.58, 95%CI 0.36–0.93, P=0.024, Figure 5 ). Subgroup analysis revealed that CTC enrichment method may be the main source of heterogeneity ( Supplementary Table 3 ). Pre-treatment PD-L1⁺ CTCs were associated with both improved PFS and OS in studies applying EpCAM-based CTC enrichment ( Supplementary Figure 4 , 5 ), whereas not significant association was found in studies implementing size-based CTC enrichment. The between-subgroup differences were statistically significant (P=0.003 and 0.036, respectively).

The prognostic value of post-treatment PD-L1⁺ CTCs and dynamic change of PD-L1⁺ CTCs was both evaluated in only 1 study. Ikeda M et al. found patients with ≥7.7% PD-L1 positivity rate in CTCs had a longer PFS than those with <7.7% positivity rate but did not have OS benefit (30). Spiliotaki M et al. observed longer PFS in patients with decreased PD-L1low CTCs after pembrolizumab treatment (33).

Two studies comprising 173 patients detected pre-treatment exoPD-L1 levels (16, 34). Wang Y et al. reported results of prognostic value of exoPD-L1 in patients receiving mono-immunotherapy and those with combination immunotherapy separately [34]. They were included in quantitative analysis as two datasets. Meta-analysis using a fixed-effect model revealed patients with high pre-treatment exoPD-L1 concentrations had significantly worse PFS compared to those with low concentrations (HR=4.24, 95%CI 2.82–6.38, P<0.001, Figure 6 ). Two studies involving 170 patients detected dynamic changes of exoPD-L1 levels immunotherapy using near 2-fold up-regulation as cut-offs (34, 35). Patients with up-regulation of exoPD-L1 levels after immunotherapy had significantly favorable PFS than those without obvious up-regulation (HR=0.36, 95%CI 0.23–0.55, P<0.001, Figure 6 ). The correlation between dynamic exoPD-L1 concentrations and OS was reported by only 1 study (35), in which patients with ≥1.86-fold un-regulation had significantly prolonged OS than those with <1.86-fold change (HR=0.24, 95%CI 0.08–0.68, P=0.008).

We conducted meta-regression to analyze whether baseline features modulated the association between pre-treatment sPD-L1 and survival outcomes in NSCLC patients treated with ICIs ( Supplementary Table 4 ). Meta-regression showed that median age, percentage of males, percentage of smokers, sample size, percentage of first-line ICIs, and cut-off value did not modulate the association of sPD-L1 with PFS. Moreover, these factors did not modulate the correlation between sPD-L1 and OS except for cut-off value. After excluding an outlier (3357 ng/ml) that was extremely larger than the other cut-offs (27), meta-regression demonstrated a positive linear relationship between cut-off value and effect size (P=0.019, Supplementary Figure 6 ). Cut-off value may partially explain the source of heterogeneity in analysis of pre-treatment sPD-L1 associated with OS.

Publication bias was assessed by viewing the symmetry of funnel plots and Egger’s test ( Supplementary Table 5 ). There was potential publication bias in meta-analysis of dynamic change of sPD-L1 concentrations associated with PFS and OS (P<0.05). Trim-and-fill analysis after imputing 1 and 2 studies showed that the association between dynamic change of sPD-L1 and survivals remained insignificant (PFS: HR=0.86, 95%CI 0.54–1.37; OS: HR=1.00, 95%CI 0.61–1.63; Supplementary Figure 7 ). Thus, publication bias had no obvious impact on results of pooled analysis. We did not observe obvious publication bias in other meta-analyses ( Supplementary Figure 8 ).