A Bayesian Network Meta-Analysis of First-Line Treatments for Non-Small Cell Lung Cancer with High Programmed Death Ligand-1 Expression

We performed a Bayesian network meta-analysis (NMA) to suggest frontline treatments for advanced non-small cell lung cancer (NSCLC) showing high programmed death ligand-1 (PD-L1) expression. A total of 5237 patients from 22 studies were included. In terms of progression-free survival, immune checkpoint inhibitors (ICIs) plus bevacizumab plus chemotherapy had the highest surface under the cumulative ranking curve (SUCRA) value (98.1%), followed by ICI plus chemotherapy (82.9%). In terms of overall survival (OS), dual immunotherapy plus chemotherapy had the highest SUCRA value (79.1%), followed by ICI plus bevacizumab plus chemotherapy (73.4%). However, there was no significant difference in survival outcomes among treatment regimens combined with immunotherapy. Moreover, ICI plus chemotherapy failed to reveal a significant OS superiority to ICI monotherapy (hazard ratio = 0.978, 95% credible internal: 0.771–1.259). In conclusion, this NMA indicates that ICI plus chemotherapy with/without bevacizumab might to be the best options in terms of OS for advanced NSCLC with high PD-L1 expression. However, considering that there was no significant difference in survival outcomes among treatment regimens incorporating immunotherapy and that ICI plus chemotherapy failed to show significant survival benefits over ICI monotherapy, ICI monotherapy may be reasonable as first-line treatment for advanced NSCLC with high PD-L1 expression.


Introduction
Lung cancer is the leading cause of cancer-related death all over the world [1]. Nearly half of all patients with non-small cell lung cancer (NSCLC) are presented in the advanced or metastatic stages, which limits the treatment options. For a long time, platinum-based doublet chemotherapy was the first-line standard treatment for patients with advanced NSCLC without driver somatic mutations [2].

Searching Strategy
We searched the Cochrane Central Register of Controlled Trials, PubMed, and EM-BASE for articles that included the following search terms in their titles, abstracts, or keyword lists: 'metastatic or advanced', 'non-small cell', 'lung', 'malignant or neoplasm or cancer or carcinoma', 'treatment', 'chemotherapy', 'immune checkpoint inhibitor or immunotherapy' and 'randomized or randomised'. All eligible studies were retrieved, and their bibliographies were checked for other relevant publications. We also scanned the reference lists of relevant articles and reviews. In addition, we used the 'related articles' features in PubMed to identify other potentially eligible articles. In the case of duplicate publications, the more recent paper was selected. Two independent reviewers examined the titles, abstracts, and full articles to determine the eligibility of the identified trials. Any disagreements were resolved through consensus or consultation with a third reviewer.

Selection Criteria
All potentially eligible studies identified using the search strategy were screened. Clinical trials that met the following criteria were reviewed for the NMA: (i) prospective randomized phase II or III trials for advanced NSCLC; (ii) trials comparing treatment regimens in the first-line setting; (iii) trials reporting the efficacy according to the level of PD-L1 expression or studies conducted for advanced NSCLC with greater than or equal to 50% PD-L1 expression.

Data Extraction
Two independent reviewers extracted the complete data from each included trial using a standardized data extraction form. Extracted data included the details of the trials (year of publication, treatments, number of patients, and histology) and outcome measures (PFS and OS). The risk of bias for each trial was assessed by the Cochrane risk of bias method. Discrepancies in data extraction were resolved through discussion.

Data Analysis
The primary outcomes intended to analyze were OS and PFS, which were reported as a hazard ratio (HR) and its 95% confidence intervals (CIs). A Bayesian NMA was conducted to evaluate the treatment effects by direct pairwise and indirect comparisons and to provide a hierarchical ranking for the treatments without direct comparisons between them. Considering the heterogeneity between included trials, a random-effects model was incorporated and an informative prior of a log-normal (−3.95, 1.342) distribution was set in the Bayesian framework [38].
The posterior distributions were obtained using Markov-chain Monte Carlo process with 5000 burn-ins and 50,000 iterations of four chains, which were thinned after every 10th simulation to reduce autocorrelation [39]. The convergence of the model was assessed by evaluating the trace plots and Gelman-Rubin diagnostics with a cut-off value of 1.05 [40]. The effect sizes of the Bayesian NMA were presented as the HR with 95% credible intervals (CrIs). To provide the rankings of each treatment, the surface under the cumulative ranking curve (SUCRA) values were calculated. A higher SUCRA value indicates a higher the likelihood that the treatment option would be in the top rank [40].
The statistical heterogeneity was evaluated using the statistic inconsistency index (I 2 ). I 2 values of <25%, 25-50%, and >50% indicate low, moderate, and high heterogeneity across randomized controlled trials, respectively. To discover the consistency, the node splitting analysis was performed to check the differences between direct and indirect comparisons among closed loops of each network. Sensitivity analysis was carried out to explore the robustness or consistency of the results and to determine whether a certain study has a high risk of bias. Egger's test and Begg's test were applied to determine publication bias across included trials where p values of <0.05 indicated publication bias. The statistical software R (R version 4.0.5, https://www.r-project.org/, accessed on 6 March 2022) and the R package GeMTC (version 1.0-1) were used to perform the NMA.

Characteristics of the Included Studies
The detailed characteristics of the included trials are summarized in Table 1

Network Analysis Diagrams
In this NMA, the treatment regimens were assigned into one of the following nodes: ICI monotherapy, ICI plus doublet, double ICIs, double ICIs plus doublet, ICI plus bevacizumab plus doublet, or doublet chemotherapy. The network analysis diagrams for OS and PFS are shown in Figure 2.

Risk of Bias Assessment
Since all studies were well-designed randomized controlled trials, the risk of bias was low in general across the studies ( Figure S1). Although there was no information about the methods of randomization and allocation concealment in several trials, selection and attrition bias seemed to be minimal. However, the studies with open-labeled design (63.6%) were scored as having a high risk of bias in terms of blinding of participants and personnel. Because almost all studies were analyzed based on the intention-to-treat population and reported sufficient endpoints, a low risk of bias was observed with respect to the incomplete outcome data and selective reporting.

Network Analysis Diagrams
In this NMA, the treatment regimens were assigned into one of the following nodes: ICI monotherapy, ICI plus doublet, double ICIs, double ICIs plus doublet, ICI plus bevacizumab plus doublet, or doublet chemotherapy. The network analysis diagrams for OS and PFS are shown in Figure 2.

Risk of Bias Assessment
Since all studies were well-designed randomized controlled trials, the risk of bias was low in general across the studies ( Figure S1). Although there was no information about the methods of randomization and allocation concealment in several trials, selection and attrition bias seemed to be minimal. However, the studies with open-labeled design (63.6%) were scored as having a high risk of bias in terms of blinding of participants and personnel. Because almost all studies were analyzed based on the intention-to-treat population and reported sufficient endpoints, a low risk of bias was observed with respect to the incomplete outcome data and selective reporting.

Progression-Free Survival
Seven network nodes covering 17 treatment regimens were included in the Bayesian NMA for PFS. The Gelman-Rubin diagnostic statistic value of 1.006 supported the model convergence, and the statistical heterogeneity was low across the trials (I 2 = 14%) by fitting the random-effects model with appropriate informative prior distributions. Egger's and Begg's tests with a funnel plot indicated that there was no significant publication bias (Egger's p = 0.300, Begg's p = 0.082). The node-splitting model indicated that there were no significant differences between direct and indirect comparisons, suggesting no inconsistency in the network ( Figure S2A). Sensitivity analysis showed that the results were relatively stable except for some small changes of SUCRA values.
The forest plot revealed that four network nodes had significant superiority to doublet chemotherapy (Figure 3). ICI plus doublet chemotherapy had a significantly better

Progression-Free Survival
Seven network nodes covering 17 treatment regimens were included in the Bayesian NMA for PFS. The Gelman-Rubin diagnostic statistic value of 1.006 supported the model convergence, and the statistical heterogeneity was low across the trials (I 2 = 14%) by fitting the random-effects model with appropriate informative prior distributions. Egger's and Begg's tests with a funnel plot indicated that there was no significant publication bias (Egger's p = 0.300, Begg's p = 0.082). The node-splitting model indicated that there were no significant differences between direct and indirect comparisons, suggesting no inconsistency in the network ( Figure S2A). Sensitivity analysis showed that the results were relatively stable except for some small changes of SUCRA values.
The forest plot revealed that four network nodes had significant superiority to doublet chemotherapy ( Figure 3). ICI plus doublet chemotherapy had a significantly better PFS over ICI monotherapy (HR = 0.571, 95% CrI: 0.454-0.709). The relative effects of all network node pairs on PFS are summarized in Table 2

Overall Survival
In this Bayesian NMA, 14 treatment regimens were available for OS analysis and assigned into seven network nodes. Model convergence was confirmed based on the Gelman-Rubin diagnostic statistic value of 1.009 and diagnostic plots. Statistical heterogeneity was found to be low across the included trials (I 2 = 0%) after applying the random-effects model with appropriate informative prior distributions. Significant publica-

Overall Survival
In this Bayesian NMA, 14 treatment regimens were available for OS analysis and assigned into seven network nodes. Model convergence was confirmed based on the Gelman-Rubin diagnostic statistic value of 1.009 and diagnostic plots. Statistical heterogeneity was found to be low across the included trials (I 2 = 0%) after applying the random-effects model with appropriate informative prior distributions. Significant publication bias was not observed when Egger's and Begg's tests with a funnel plot were performed (Egger's p = 0.868, Begg's p = 0.371). The node-splitting analysis revealed that there were no significant differences between the direct and indirect estimates, indicating no inconsistency in the network ( Figure S2B). The consistency of results was verified by sensitivity analysis.
Except for bevacizumab plus doublet chemotherapy, all treatments demonstrated a significantly reduced risk of death compared with doublet chemotherapy (Figure 4). However, none of the treatment regimens incorporating ICI showed significantly better OS than others in patients with NSCLC showing high PD-L1 expression. Especially, ICI plus doublet chemotherapy failed to show a significant superiority over ICI monotherapy (HR = 0.978, 95% CrI: 0.771-1.259). random-effects model with appropriate informative prior distributions. Significant publication bias was not observed when Egger's and Begg's tests with a funnel plot were performed (Egger's p = 0.868, Begg's p = 0.371). The node-splitting analysis revealed that there were no significant differences between the direct and indirect estimates, indicating no inconsistency in the network ( Figure S2B). The consistency of results was verified by sensitivity analysis.
Except for bevacizumab plus doublet chemotherapy, all treatments demonstrated a significantly reduced risk of death compared with doublet chemotherapy (Figure 4) However, none of the treatment regimens incorporating ICI showed significantly better OS than others in patients with NSCLC showing high PD-L1 expression. Especially, ICI plus doublet chemotherapy failed to show a significant superiority over ICI monotherapy (HR = 0.978, 95% CrI: 0.771-1.259).
The relative effects of all network node pairs for OS are presented in Table 2. The ranking of each treatment strategy was estimated according to the SUCRA values Double ICIs plus doublet chemotherapy had the highest SUCRA value (79.1%), followed by ICI plus bevacizumab plus doublet chemotherapy (73.4%). ICI plus doublet chemotherapy (64.9%) and ICI monotherapy (61.8%) had a similar SUCRA value indicating that they are equally effective against NSCLC with high PD-L1 expression in terms of OS.  The relative effects of all network node pairs for OS are presented in Table 2. The ranking of each treatment strategy was estimated according to the SUCRA values. Double ICIs plus doublet chemotherapy had the highest SUCRA value (79.1%), followed by ICI plus bevacizumab plus doublet chemotherapy (73.4%). ICI plus doublet chemotherapy (64.9%) and ICI monotherapy (61.8%) had a similar SUCRA value, indicating that they are equally effective against NSCLC with high PD-L1 expression in terms of OS.

Safety Analysis
Safety was analyzed according to all-grade adverse events (AEs) and Grade 3-5 AEs. Bayesian NMA included all network nodes. The incidence of toxicities was lowest for ICI monotherapy followed by double ICIs and doublet chemotherapy in both analyses (Table 3). Especially, ICI monotherapy and double ICIs showed significantly lower odds ratios (ORs) compared to the rest of the regimens combined with ICIs. The addition of chemotherapy and/or anti-angiogenic drug to ICIs elevated the toxicity.
The Bayesian NMA results for the efficacy and safety were summarized in the scatter plot based on the SUCRA values of OS, PFS, and Grade 3-5 AEs in Figure 5.
An increasing number of studies have suggested that there may be the synergistic anti-tumor effects between ICIs and chemotherapy. Cytotoxic agents may exhibit positive immuno-modulatory effects by releasing a high level of tumor antigens and changing the tumor micro-environment [44,45]. Accordingly, the combination of ICI and chemotherapy may reveal greater efficacy than chemotherapy alone, particularly in patients with lower PD-L1 expression levels. In fact, many randomized clinical trials have suggested that combining an anti-PD-1 mAb (pembrolizumab, camrelizumab, tislelizumab, sintilimab) or anti-PD-L1 inhibitor (atezolizumab) with platinum-doublet chemotherapy could significantly improve PFS or OS compared with chemotherapy alone in both squamous and nonsquamous advanced NSCLC, irrespective of the level of PD-L1 expression [16,18,19,23,32,34,41].
Besides ICIs targeting PD-l/PD-L1, cytotoxic T-lymphocyte-associated protein-4 (CTLA-4) checkpoint inhibitors also enhance T-cell activity against tumors with different complementary mechanisms. The first phase III study of dual immunotherapy, Check-Mate 227, investigated the efficacy of nivolumab plus ipilimumab compared with plati-
An increasing number of studies have suggested that there may be the synergistic anti-tumor effects between ICIs and chemotherapy. Cytotoxic agents may exhibit positive immuno-modulatory effects by releasing a high level of tumor antigens and changing the tumor micro-environment [44,45]. Accordingly, the combination of ICI and chemotherapy may reveal greater efficacy than chemotherapy alone, particularly in patients with lower PD-L1 expression levels. In fact, many randomized clinical trials have suggested that combining an anti-PD-1 mAb (pembrolizumab, camrelizumab, tislelizumab, sintilimab) or anti-PD-L1 inhibitor (atezolizumab) with platinum-doublet chemotherapy could significantly improve PFS or OS compared with chemotherapy alone in both squamous and nonsquamous advanced NSCLC, irrespective of the level of PD-L1 expression [16,18,19,23,32,34,41].
Besides ICIs targeting PD-l/PD-L1, cytotoxic T-lymphocyte-associated protein-4 (CTLA-4) checkpoint inhibitors also enhance T-cell activity against tumors with different complementary mechanisms. The first phase III study of dual immunotherapy, CheckMate 227, investigated the efficacy of nivolumab plus ipilimumab compared with platinumbased chemotherapy as frontline treatment of advanced NSCLC without EGFR or ALK mutations [25]. The updated results of the CheckMate 227 part 1 were recently reported [41]. With a median follow-up of 54.8 months, OS remained longer with nivolumab plus ipilimumab versus chemotherapy not only in patients with PD-L1 greater than or equal to 1% (HR = 0.76, 95% CI: 0.65-0.90) but also in patients with PD-L1 less than 1% (HR = 0.64, 95% CI: 0.51-0.81). In the CheckMate 9LA trial, first-line nivolumab plus ipilimumab combined with two cycles of chemotherapy improved OS versus chemotherapy alone (median OS 15.6 vs. 10.9 mo, HR = 0.66, 95% CI: 0.55-0.80) in patients with advanced NSCLC [26]. Interestingly, Ando et al. indirectly compared nivolumab plus ipilimumab versus other immunotherapies using NMA in PD-L1 positive (≥1%) advanced NSCLC [46]. The SUCRA ranking showed that pembrolizumab plus doublet chemotherapy had the highest efficacy for PFS, followed by nivolumab plus ipilimumab, nivolumab, doublet chemotherapy, and pembrolizumab. The safety outcome analysis revealed that nivolumab plus ipilimumab was well tolerated compared to existing immunotherapy regimens. These results indicate the possibility of dual immunotherapy with nivolumab and ipilimumab as a new therapeutic option in PD-L1-positive advanced NSCLC.
As we reviewed above, available data indicate that the addition of ICIs to chemotherapy with or without an anti-angiogenic agent increases survival benefits in advanced NSCLC, regardless of PD-L1 expression proportions [16][17][18][19]23,32,34]. Except for the KEYNOTE-598 study [43], however, no randomized clinical trials comparing the efficacy of ICIs as monotherapy versus combination with other treatment options are now available for patients with advanced NSCLC highly expressing PD-L1. Because this subgroup may achieve greater survival benefits from anti-PD-1/PD-L1 mAbs than chemotherapy, sparing those patients the risk of increased toxicities with the combination of other agents should be an important consideration.
In this Bayesian NMA of 22 randomized phase II or III trials with a total of 5237 patients, we indirectly compared survival outcomes and AEs of the seven treatment nodes (ICI monotherapy, ICI plus doublet chemotherapy, double ICIs with or without doublet chemotherapy, bevacizumab plus doublet chemotherapy with or without ICI, and doublet chemotherapy) as first-line treatment for advanced NSCLC with high PD-L1 expression. The toxicities were most tolerable for ICI monotherapy followed by double ICIs and doublet chemotherapy. Especially, ICI monotherapy and double ICIs showed significantly lower OR compared to the rest of the treatment regimens. The addition of chemotherapy and/or an anti-angiogenic drug to ICIs elevated the toxicities. Other network metaanalyses have also demonstrated that ICI monotherapy had significantly lower odds of any AEs than chemotherapy or a combination of ICI and chemotherapy [4,36]. In terms of PFS, four network nodes (ICI monotherapy, ICI plus chemotherapy, and bevacizumab plus chemotherapy with or without ICI) showed significant superiority, compared with chemotherapy alone. Interestingly, ICI plus chemotherapy had a significantly better PFS over ICI monotherapy (HR = 0.571, 95% CrI: 0.454-0.709). When the ranking of each treatment was estimated according to SUCRA values, ICI plus bevacizumab plus chemotherapy had the highest probability of being the most effective regimen (98.1%), followed by ICI plus chemotherapy (82.9%). In terms of OS, all treatment regimens except for bevacizumab plus chemotherapy demonstrated longer survival compared with chemotherapy alone. However, none of treatment regimens incorporating ICI showed significantly better OS than others. Especially, ICI plus chemotherapy failed to show a significant superiority over ICI monotherapy (HR = 0.978, 95% CrI: 0.771-1.259), indicating that ICI plus chemotherapy has no survival advantage compared with ICI monotherapy for patients with PD-L1 expression of at least 50%. Based on the SUCRA values, dual immunotherapy plus chemotherapy had the highest value (79.1%), followed by ICI plus bevacizumab plus chemotherapy (73.4%). However, it should be considered that only two studies were included in these treatment nodes. Moreover, dual immunotherapy (64.9%) and ICI monotherapy (61.8%) have similar SUCRA values, suggesting that they are equally effective in terms of OS against advanced NSCLC with high PD-L1 expression. Our findings were in concordance with the result from the recent meta-analysis of randomized controlled trials [36]. Li et al. compared the efficacy and safety of PD-1/PD-L1 inhibitors plus chemotherapy versus PD-1/PD-L1 inhibitors in advanced NSCLC using a network analysis. When stratifying patients according to PD-L1 expression level, patients with high PD-L1 expression receiving PD-L1 inhibitors plus chemotherapy had improved PFS, but not OS, compared to PD-L1 inhibitors as monotherapy.
Several limitations of this study need to be mentioned. First, this NMA was performed using aggregated data of results from the eligible trials, not individual patient data. Second, this study also included a randomized phase II study, and data from this kind of study may not be as reliable as data from phase III trials. However, only a single phase II study was included in the NMA, and thus it is less likely to have debatable impacts on the results [18]. Third, only one or two studies were included in two treatment nodes (double ICI plus doublet and ICI plus bevacizumab plus doublet), which could result in estimates with lower statistical power. Fourth, we did not stratify patients according to the histology (squamous or non-squamous) because of the limited number of available studies for each histology. Finally, the IHC methods measuring PD-L1 expression level were different among studies, which might cause patients to be misclassified.
In conclusion, we combined both direct and indirect evidence in this NMA of randomized trials to suggest frontline treatments for advanced NSCLC with high PD-L1 expression (≥50%). The results indicated that treatment regimens combined with immunotherapy reveal better survival outcomes compared with chemotherapy alone. Based on the SUCRA values, ICI plus bevacizumab plus chemotherapy or ICI plus chemotherapy might to be the best option in terms of OS. However, there was no significant difference in survival outcomes among treatment regimens combined with immunotherapy. Moreover, ICI plus chemotherapy failed to reveal significant survival benefits over ICI monotherapy. In addition, ICI monotherapy was most tolerable in terms of AEs, followed by double ICIs and doublet chemotherapy. In terms of both OS and safety, therefore, ICI monotherapy may also be reasonable as first-line treatment for advanced NSCLC with high PD-L1 expression and no targetable aberrations. Considering no prospective direct comparison is now available, however, the choice of treatment should be determined based on patient-specific factors after open discussion with the patient on the benefits, cost, and risks of each option. Ran-domized clinical trials are still warranted in order to identify the best therapeutic strategy for patients with advanced NSCLC highly expressing PD-L1.
Supplementary Materials: The following supporting information can be downloaded at: https:// www.mdpi.com/article/10.3390/jcm11061492/s1. Table S1: Detailed characteristics of the 22 studies included in the Bayesian network meta-analysis. Figure S1: Graphic (A) and tabular representation (B) of risk of bias in all studies included. Figure