Therapeutic options for advanced epidermal growth factor receptor (EGFR)-mutant non-small cell lung cancer: a Bayesian network secondary analysis

The most favorable treatments for advanced EGFR-mutant NSCLC are less indicated. Forty-one studies were eligible for this Bayesian network secondary analysis. For PFS, erlotinib (Erlo)+bevacizumab (Bev) (HR 0.26, 95% CrI: 0.08-0.75 vs placebo), osimertinib (Osi) (HR 0.29, 0.11-0.70 vs placebo), and afatinib (Afa) were top-ranking individual treatments, while immunotherapy (IT)+anti-VEGFR (aVEGFR)+platinum-based therapy (Plat) (HR 0.42, 0.06-2.63 vs placebo), EGFR-TKI (ET)+aVEGFR (HR 0.35, 0.14-0.85 vs placebo), and ET+aVEGFR+Plat were top-ranking medication classes. For OS, Osi (HR 0.52, 0.10-2.00 vs placebo), cetuximab (Cet)+Bev+Plat (HR 0.51, 0.06-3.38 vs placebo), and cilengitide (Cil)+Cet+Plat were top-ranking individual treatments, while ET+aVEGFR+Plat, ET+Plat, and third-generation EGFR-TKI (3rd ET) were top-ranking medication classes. For PFS regarding the EGFR genomic aberration status, Erlo+Bev, Osi, and Afa were superior for exon 19 deletion status, whereas ET+Bev, Osi, and gefitinib (Gef)+pemetrexed (Peme) were excellent for exon 21 L858Arg mutation status. The results were consistent in terms of the ORR and DoR and remained robust across sensitivity analyses. However, Erlo + Bev had the most grade 3 or higher adverse events. Osi, Erlo+Bev, and Erlo+Bev+Plat are reasonably recommended to balance PFS and OS, but adverse events should be considered. IT+aVEGFR+Plat shows potential superiority, but more clinical evidence is needed.


INTRODUCTION
Non-small cell lung cancer (NSCLC) represents approximately 85% to 90% of lung cancer cases and is the leading cause of cancer-related death worldwide, with a lower than 15% 5-year survival [1,2]. Since treat-ment selections have become increasingly related to the biological subtypes of NSCLC, attention has been drawn to tumors harboring epidermal growth factor receptor (EGFR) mutations, which are estimated to exist in 10%-15% of patients with nonsquamous NSCLC [3]. The identification of EGFR mutations has led to the AGING development of targeted therapies, including small molecule tyrosine kinase inhibitors (TKIs) directed at the signal transduction pathway as well as immunotherapies incorporating checkpoint monoclonal antibodies that bind to and inactivate the receptors on cell membranes [4].
As a monotherapy, gefitinib, erlotinib and, more recently, afatinib have been licensed and recommended as first-line treatment regimens for EGFR-mutant NSCLC patients by the European Society for Medical Oncology (ESMO) guidelines. In August 2015, the American Society of Clinical Oncology (ASCO) clinical guidelines recommended two cytotoxic drugs, docetaxel and pemetrexed, and two EGFR-TKIs, erlotinib and gefitinib, to patients who experienced treatment failure with conventional first-line chemotherapy [4]. Nevertheless, several new regimens have been approved by the US FDA, such as the combination of docetaxel and ramucirumab, nivolumab, pembrolizumab, and atezolizumab. At the same time, more than 40 therapeutic options are being assessed in randomized controlled trials (RCTs) [5]. With more clinical trials emerging , the FLAURA trial [42] has shown that osimertinib has superior efficacy compared with standard EGFR-TKIs in treating advanced EGFR-mutant NSCLC with less serious adverse effects (18.9 months vs 10.2 months for progression-free survival (PFS), P<0.001). The newest National Comprehensive Cancer Network (NCCN) guidelines also regarded osimertinib as category 1 for advanced EGFR-mutant NSCLC.
There is an urgent need to identify complete information on the most effective and latest treatment for advanced EGFR-mutant NSCLC. Conventional meta-analyses have only partially captured the available evidence for treating the intended populations; their outcomes are not comprehensive. This work is a generalized version of a pairwise meta-analysis integrating direct and indirect evidence  to aid in clinical decision making. Thus, the aim of this article is to comprehensively evaluate the effectiveness and safety of various therapeutics for advanced EGFR-mutant NSCLC.

Study selection and characteristics of the included studies
We identified a total of 1749 records from a database search and 34 records from other available literature; of these, 1721 were excluded based on the selection criteria. Subsequently, 62 potential articles went through full-text review, and 41 studies were ultimately eligible for inclusion (Appendix Figure 1 in the Supplementary Data).

Risk of bias and quality assessment
In the quality assessment, we found that a vast majority of the included studies had a low risk of bias (Appendix Table 3 in the Supplementary Data). Two [24,46] studies had "other bias" according to the Cochrane risk of bias tool, and 16 [10, 13, 15, 19, 20, 22, 26-29, 31, 35, 38, 39, 41, 44] had an unclear "other bias". Most of the risk stemmed from the blinding of participants and personnel and blinding of outcome assessment categories due to the open-label method.

Bayesian NMA at the individual-treatment level
A total of 37 RCTs with 37 arms and 21 unique treatment levels were included in the treatment level analysis for PFS; a study by Reck et al. [45] was omitted for not contributing to a complete loop ( Figure  1A). Twenty-one unique nodes were presented in this analysis, with every node representing a unique A total of 21 RCTs with 21 arms and 18 unique treatment levels were considered in the OS analysis; a study by Reck et al. [45] was excluded from the network ( Figure 2A). Eighteen nodes were included in the treatment-level analysis for OS. The most direct interactions were in nodes between Erlo and Plat ( Figure 2B). Osi (HR 0.52, 95% CrI: 0.10-2.00), Cet+Bev+Plat (HR 0.51, 95% CrI: 0.06-3.38), Cil+Cet+Plat (HR 0.59, 95% CrI: 0.06-4.30), and Cet+Plat (HR 0.63, 95% CrI: 0.10-3.03) appeared to benefit OS over placebo even though the favorable efficacy was nonsignificant. Osi (ranking probability: 0.183) ranked first, followed by Cet+Bev+Plat (0.157), Cil+Cet+Plat (0.117), and Cet+Plat (0.147) ( Figure 2C, 2D). The results of the comparisons among those treatments and all treatments are shown in Figure 2C and Appendix Table 4 in the Supplementary Data. The statistical model was good, and no evidence of heterogeneity or loop inconsistency was observed ( Table 2). A total of 13 trials with 10 unique treatment levels were analyzed for ORR after excluding Reck et al. [45] and Yang et al.   Table 6 in the Supplementary Data). No evidence of heterogeneity was found. A total of 17 trials with 13 unique treatment levels focused on grade 3 or higher AEs (Appendix Figure 3 in the Supplementary Data). The primary data are shown in Appendix Table 7 in the Supplementary Data. Erlo + Bev had the most grade 3 or higher AEs (OR 24.22, 95% CrI: 0.64-NA vs Ico), followed by Gef +Peme (OR 15.28, 95% CrI: 0.43-NA), Plat (OR 12.61, 95% CrI: 0.86-NA) and Dac (OR 11.09, 95% CrI: 0.33-NA) (Appendix Table 8 in the Supplementary Data). No significant heterogeneity was noted ( Table 2). The outcome of DoR was not analyzed for insufficient comparisons.

Bayesian NMA at the medication-class level
There were 35 RCTs with 36 arms and 15 unique class levels included in the class-level analysis for PFS; the studies of Yang et al. [39], Urata et al. [32], and Shi et al. [17] were excluded for having two of the same class interventions. A total of 15 nodes were presented, in which the most direct interventions were between 1stgen ET and Plat, 1st-gen ET and 2nd-gen ET, and 1stgen ET and Placebo ( Figure 3A and Figure 3B). ET+aVEGFR (HR 0.35, 95% CrI: 0.14-0.85) and 3rdgen ET (HR 0.39, 95% CrI: 0.16-0.91) were better than placebo, while ET+aVEGFR+Plat (HR 0.40, 95% CrI: 0.11-1.37) and IT+aVEGFR+Plat (HR 0.42, 95% CrI: 0.06-2.63) seemed to be superior than placebo, but the data were not statistically significant ( Figure 3C). Our work demonstrated that statistically, ET+aVEGFR ranked first, followed by 3rd-gen ET, ET+aVEGFR+Plat, and IT+aVEGFR+Plat ( Figure 3D). The results of the comparisons among those treatments and all treatments are shown in Figure 3C and Appendix Table 9 in the Supplementary Data. The model was good, and no heterogeneity or inconsistency was observed.
In total, 21 RCTs with 11 arms and 11 unique class levels were selected for the class-level analysis of OS; studies by Yang et al. [39] and Reck et al. [45] were excluded for comparing the same class interventions ( Figure 4A). The 11 nodes with the most direct interactions were between 1st-gen ET and Plat ( Figure 4B). When compared with placebo, the HR for   AGING ET+aVEGFR+Plat was 0.65 (95% CrI: 0.13-2.71), the HR for ET+Plat was 0.76 (95% CrI: 0.17-2.65), and the HR for 3rd-gen ET was 0.66 (95% CrI: 0.16-2.04) ( Figure 4C). Bayesian analysis revealed that ET+aVEGFR+Plat was significantly ranked first, followed by ET+Plat, and then 3rd-gen ET ( Figure 4D). Additional data on the interactions can be found in Figure 4C and Appendix Table 9 in the Supplementary Data. The fitness for this model was good, and no significant heterogeneity or inconsistencies were observed.
In medication-class level analysis for ORR (Appendix Figure 4 in the Supplementary Data, Appendix  Table 10 in the Supplementary Data). Analysis for DoR (Appendix Figure 5 in the Supplementary Data, Appendix Table 11 in the Supplementary Data) revealed that 2nd-gen ET had the longest DoR (MD 5.08, 95% CrI: -11.44 to 21.63 vs ET + aVEGFR), followed by 3rd-gen ET (MD 3.04, 95% CrI: -11.29 to 17.31) and ET + CT (MD 3.87, 95% CrI: -12.47 to 20.38) (Appendix Table 12 in the Supplementary Data). Analysis for grade 3 or higher AEs (Appendix Figure 6 in the Supplementary Data, Appendix  Table 13 in the Supplementary Data). There was low to moderate heterogeneity among medication-level comparisons on ORR, DoR and grade 3 or higher AEs.

EGFR genomic status
Multiple treatment-level comparisons (10 treatments) were specifically performed for 19 del and 21 L858R mutations regarding PFS. We found that Erlo+Bev, Osi, Afa, and Erlo were top-ranking alternatives for patients harboring 19 del mutations to prolong PFS; additionally, Erlo+Bev, Osi, Gef+Pem, and Dac were optimal treatments among the available treatments for patients with 21 L858R mutations (Table 3). Overall, Erlo+Bev, Osi, and 2nd-gen ET (including Afa and Dac) were the most viable treatment options, and no significant differences were found in terms of the best strategies for 19 del and 21 L858R mutations. The evidence was insufficient to assess these mutations in terms of OS as well as uncommon mutations.

Sensitivity analysis
Sensitivity analyses helped to confirm the robustness of these findings, which put more weight on main outcomes. The results restricted to phase III trials (trial number = 34 and patient number = 7448) did not show significant deviations compared with the original network Bayesian analyses; however, they showed a more robust status of Osi that had better OS and PFS. Superiority of Erlo + Bev on PFS was also stressed. The results of the primary meta-analysis remained stable across sensitivity analyses by removing Reck et al. [45], (IMpower150) and Soria et al. [42], (FLAURA), respectively. The superiority of Osi was confirmed to be robust. When restricted to Asia (trial number = 12 and patient number = 1977), the superiority of Osi on PFS slightly declined, but Osi was still among the topranking treatments. Erlo + Bev and EGFR-TKI + aVEGFR + Plat showed favorable PFS across Asian and non-Asian populations (trial number = 29 and patient number = 6408).

DISCUSSION
With increasing molecular research focused on biomarkers for NSCLC, attention has been drawn to targeted therapies and immunotherapy. The available guidelines have helped improve the level of clinical treatment, but they still need to be updated as more evidence and clinical trials emerge. In the present study, we gathered evidence from 41 RCTs assessing the role of 22 treatment-level options and 22 class-level options for the intended populations. Herein, we found that Erlo+Bev, Osi, Afa, and Erlo are optimal treatment-level options in terms of PFS and that Osi, Cet+Bev+Plat, Cil+Cet+Plat, and Cet+Plat are successively optimal options in terms of OS. For medications at the class level, ET+aVEGFR, 3rd-gen ET, ET+aVEGFR+Plat,  ET monotherapy has been established as the standard treatment for patients with EGFR-positive NSCLC, and a meta-analysis involving patients who had not previously received treatment showed a median PFS of 11 months with ET (gefitinib/erlotinib) versus 5.6 months with chemotherapy [56]. However, most patients with lung cancer are diagnosed at an advanced stage, and the prognosis remains poor despite novel therapeutics.
To improve PFS, combination treatments with 1st-gen ET and 2nd-gen ET have been evaluated in several clinical trials [57]. Bev inhibits angiogenesis by restricting oxygen and nutrient supplies to suppress tumor growth, and combination chemotherapy with Bev has been shown to be effective against CNS metastasis and pleural effusion [58,59], but the conclusion is controversial. Several clinical trials (JO25567, NEJ026) [22,46] included in this study also compared the efficacy of Erlo+Bev with that of Erlo, and the results showed that PFS was significantly prolonged in the Erlo+Bev group. A meta-analysis [60] investigating Erlo+Bev compared with monotherapy for NSCLC found no substantial benefits for the OS or PFS of all patients, but the combination treatment significantly enhanced OS for EGFR-mutant patients. The mechanism by which Erlo+Bev improves PFS is still unclear, but there are some hypotheses. Bev might normalize blood flow, thus improving drug delivery to tumor blood vessels [61,62]. Autocrine or paracrine signaling by the VEGF receptor might catalyze cancer cell proliferation and produce anti-apoptotic effects, which could be inhibited by Bev to restore apoptosis [63]. Resistance by Erlo+Bev to the VEGF-mediated pathway has been confirmed in basic research [64]. Previous evidence has demonstrated that the response induced by the 21 L858R mutation was inferior to that induced by the 19 del mutation, which could be improved by Erlo+Bev. However, in this study, Erlo+Bev was found to be the preferred treatment for both 21 L858R and 19 del mutations. Although the OS endpoint was or met and was not significantly different, further clinical validation is still needed. When combined with ET, Bev+Plat has shown promising efficacy among patients with liver metastasis harboring EGFR mutations, which was consistent with our findings. Accordingly, Erlo+ET+Plat also offers potential benefits for patients with advanced EGFR-mutant NSCLC.
Osi is an oral 3rd-gen ET that selectively inhibits ET sensitivity and EGFR T790M resistance (present in almost 60% of patients) and has been approved for the treatment of patients with metastatic T790M-positive NSCLC. Preclinical and clinical data (AURA3) [34,65] support the ability of Osi to cross the blood-brain barrier and penetrate the CNS; the PFS of advanced EGFR-mutant NSCLC in the FLAURA trial [42]  This study has several clinical implications and strengths that should be mentioned. To the best of our knowledge, this study is the most comprehensive work comparing treatment effectiveness for advanced EGFR-mutant NSCLC to date. Current national and international guidelines are mostly based on the results of single RCTs, as well as standard meta-analyses dedicated to the pairwise comparisons of two or three treatments. Moreover, with the large number of treatment options, meta-analyses of direct comparisons are inevitably limited by the relatively small number of studies assessing a particular pair of treatments. Metaanalyses on multiple treatments reduce this issue by creating indirect comparisons and allowing data synthesis that helps to identify the best treatment. This study is the first to address the efficacy of therapeutics for advanced EGFR-mutant NSCLC with state-of-the art Bayesian methods. The conclusions are also strengthened by the highest level of evidence. Finally, this study is based on multivariable, time-varying HRs that assumed proportional hazards, examined the relative treatment efficacy based on parameters of survival plots (shape and scale), and considered the influence of time [67]. Network studies regarding relative risks or odds ratios do not have these strengths.
There are also several limitations that should be acknowledged. First, one or two studies were omitted from the treatment-level and medication-level analyses for not contributing to a complete network loop.
Although the credibility of our conclusion might not be maximized, the omitted studies had little impact on our final results. Second, the results on ORR and DoR might not be as well reported because of the limited study number. Additionally, potential AEs could influence our judgment on the preferred treatment. Third, the OS outcome was not as comprehensively reported by the primary studies as PFS, and the results based on statistical data still need clinical validation.

CONCLUSIONS
This is the first Bayesian NMA to show promising, significant efficacy and safety for Erlo+Bev, Erlo+Bev+Plat and Osi over other available treatments for advanced EGFR-mutant NSCLC, considering the balance between PFS and OS. Erlo+Bev and Osi are the top-ranked regimens for patients with either 19 del or 21 L858R mutations. Accordingly, the three strategies can be reasonably recommended to these intended patients based on their effectiveness. However, their AEs should also be determined in real clinical practice. IT+aVEGFR+Plat is a potential superior regimen but still needs to be confirmed by more clinical trials. The current study not only provides evidence for the use of Erlo+Bev but also addresses the landscape of the use of Osi in advanced EGFR-mutant NSCLC. In the future, more evidence is needed to develop novel therapies and AGING to identify the best treatment options for patients according to their NSCLC subtype and for those with site-specific metastases.

MATERIALS AND METHODS
This article was based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension statement for network meta-analyses of health care interventions (Appendix Table 1 in the Supplementary Data). The protocol was registered with PROSPERO (CRD42019137033).

Search strategy
Related published trials were identified after a rigorous literature search in PubMed, EMBASE, Cochrane Library and Clinical Trials.gov from their inception to September 2019. The key terms used were "EGFR mutant", "non-small cell lung cancer", "NSCLC", and "randomized controlled trials" (Appendix Material and Methods in the Supplementary Data). No language restrictions were applied. Reference lists were searched manually for additional records.

Selection criteria
All published RCTs involving adult patients (≥18 years) whose Eastern Cooperative Oncology Group (ECOG) performance status was 0 or 1 and comparing any systematic interventions (pharmaceutical, surgical, radiological, combinations, etc.) for histologically or cytologically confirmed advanced (stage III/IV/ recurrent/distant metastasis (brain, liver, bone, etc.)) EGFR-mutant NSCLC was identified. There were no mandatory restrictions on first-line treatment settings or other-line settings. The included patients within the selected trials must have positive and clear advanced EGFR-mutant cancer diagnoses. The duration period of the eligible trials was no less than 6 months. No further restrictions were applied on other individual-level (i.e., age, sex) or program-level characteristics (i.e., start year, follow-up period). If a multi-arm trial compared one treatment to two or more different treatments, we extracted every arm/comparison. The most recent and informative publication was selected to avoid duplications. We excluded trials comparing different administration schemes with the same drug or combinations. Dose-expansion trials, reviews, and fundamental experiments were also excluded.

Definitions of outcomes and treatment arms
In this study, the primary outcomes were PFS and OS according to the Response Evaluation Criteria in Solid Tumors (RECIST version 1.1). The secondary outcomes were objective response rate (ORR), duration of response (DoR, month) and grade 3 or higher adverse events (AEs) (severe AEs). Eligible studies should report at least one of both clinical outcomes. EGFR mutations included exon 19 deletion (19 del), exon 21 Leu858Arg mutation (21 L858R) and other uncommon mutations (19 del and 21 L858R were the main focus) [47].
To organize the current treatment options in clinical trials into clinically meaningful arms, we used general prespecified criteria, as shown in Table 1. Cilengitide (Cil) and cetuximab (Cet) are seldom used in NSCLC, and for statistical convenience and network simplification, Cil is categorized in the aVEGFR class, and Cet is categorized in the ET class [48].

Data extraction and quality assessment
Relevant data were independently extracted by two investigators following our prespecified protocol. Any discrepancies were resolved by discussion with a third investigator. The extracted information included characteristics of the eligible trials (publication year, first author, trial name, follow-up period, number of arms, etc.), characteristics of the populations (mean age, number of enrolled patients, etc.), and characteristics of the program (types of systematic interventions, outcomes of intended endpoints, registration information, etc.). Outcome estimates were extracted using fully adjusted models. Additionally, we contacted the authors if there were any missing data. If we received no response, the analysis was performed without these data. Intent-to-treat data were used when available.
The risk of bias of the included RCTs was assessed using the modified Cochrane Collaboration's risk of bias tool [49]. The two coauthors performed a quality assessment on all the included RCTs. In the case of disagreements, the two authors rechecked the original articles, and a consensus was achieved after a discussion.

Statistical analysis
For PFS and OS, the hazard ratios (HRs) and confidence intervals (CIs) were directly extracted from the original studies or were calculated by methods provided by Tierney et al. [50]. We also tried to contact the authors if the study provided only figures without exact data. If the authors did not respond, the program Engauge Digitizer 4.1 (http://digitizer.sourceforge.net) was run to extract the exact data from the figures. Odd ratios (ORs) for ORR and grade 3 or higher AEs were manually calculated based on extracted information.

AGING
A Bayesian network meta-analysis (NMA) was performed with a random effects model to estimate the HR and 95% credible interval (95% CrI) for direct and indirect evidence on advanced EGFR-mutant NSCLC by combining multiple systematic arms across studies with all the information regarding PFS and OS. In the case of multi-arm trials (trials with three or more interventions), adjustments were made to preserve randomization and correlation within the multi-arm trials by converting log-HRs to log-hazards. ORs and 95% CrI in the random effects model were prepared for ORR and grade 3 or higher AEs for direct and indirect evidence; the mean difference (MD) and 95% CrI in random effects was conducted for DoR because DoR was regarded as a continuous variable. Following the Cochrane Handbook [49], the standard deviation (SD) was roughly computed by the (Xmax-Xmin)/range difference for further analysis.
The Markov chain Monte Carlo (MCMC) method was used to estimate the posterior distribution of each parameter, and the fit of the random effects model was assessed by the deviance information criteria (DIC) [51,52]. A hierarchical Bayesian model synthesizes comparisons between the treatment pairs and simultaneously summarizes all outcomes of interest by assuming a common heterogeneity parameter (a derived I 2 statistic > 50% or a P value for Cochran Q chi-square test <0.1 was regarded as indicating significant heterogeneity) [53,54]; the inconsistency of this model was evaluated by the edge-splitting method based on all direct and indirect evidence [54]. To confirm the robustness of our findings, sensitivity analyses were performed with studies restricted to phase III trials, studies excluding Reck et al. [45] and Soria et al. [42], respectively and Asian and non-Asian studies. The relative treatment rankings were graphically displayed with rankograms [55].
In the Bayesian context, the statistical significance of HRs and ORs was established when the 95% CrI did not contain 1, and that of MDs was established when it did not contain 0. Calculations were performed in R version 3.5.3 (https://www.r-project.org) using the gemtc and rjags packages, which are publicly available. The detailed statistical methods are provided in the Appendix Material and Methods in the Supplementary Data.

Data availability statement
Data sharing is not applicable to this article as no new data were created or analyzed in this study.

CONFLICTS OF INTEREST
The authors declare no conflict of interest.

Supplementary Materials and Methods
This article was according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension statement for network meta-analysis for health care. The protocol was registered with PROSPERO (CRD42019137033).

Search strategy
Relating published trials were identified after a rigorous literature search on PubMed, EMBASE, Cochrane Library and Clinical Trials.gov from inception to Sep 2019. The key items used were "EGFR mutant", "EGFR mutation", "non-small cell lung cancer", "NSCLC", "randomized controlled trials". No restrictions were applied on language. Reference lists were searched manually for additional records [1,2].
# Comprehensive searches were conducted in four electronic databases: ('crossover procedure':de OR 'double-blind procedure':de OR 'randomized controlled trial':de OR 'single-blind procedure':de OR (random* OR factorial* OR crossover* OR cross NEXT/1 over* OR placebo* OR doubl* NEAR/1 blind* OR singl* NEAR/1 blind* OR assign* OR allocat* OR volunteer*):de,ab,ti) Each of the four components of the search strategy was first searched upon individually, combining synonyms describing that concept with the Boolean operator OR. The four individual component search sets were then combined together using the Boolean operator AND.
Resulting citations were managed and duplicates removed using the Endnote citation management software program X8 (Thomson Reuters).

Selection criteria
All the published RCTs of adult patients (≥18 year) whose ECOG status was 0 or 1 that compared any systematic interventions (pharmaceutical, surgical, radiological, combinations etc.) for advanced EGFRmutant NSCLC were identified. No mandatory restrictions on first-line treatment settings or other-line settings. The included patients within selected trials must have positive and clear advanced EGFR mutant cancer diagnoses. Duration period of eligible trials should not be less than 1 year. No further restrictions were applied on other individual-level and programlevel characteristics. If a multi-arm trial compared one treatment to two or more different treatments, we extracted every arm/comparison respectively. The most recent and informative publication was selected for avoiding duplication. We excluded trials comparing different administration schemes with the same drug or combinations. Dose-expansion trials, reviews, fundamental experiments were also excluded.

Definitions of outcomes and treatment arms
In this study, the primary outcomes were PFS and overall survival (OS) according to the Response Evaluation Criteria in Solid Tumors (RECIST version 1.1). The secondary outcomes were objective response rate (ORR), duration of response (DoR, month) and grade 3 or higher adverse events (AEs) (severe AEs). Eligible studies should report at least one of the both clinical outcomes. EGFR mutations include regular exon 19 deletion (19 del) and exon 21 Leu858Arg mutation (21 L858R) and other uncommon mutations (19 del and 21 L858R were mainly focused).
Actually, cilengitide and cetuximab were rarely used in NSCLC, for the statistical convenience and the network simplification, cilengitide was regarded as aVEGFR class and cetuximab was outlined into ET class.

Data extraction and quality assessment
Relevant data were extracted by two independent investigators following our prespecified protocol. Any discrepancies would be resolved by a discussion with a third investigator. The extracted information included: characteristics of the eligible trials (publication year, the first author, trial name, follow-up period, number of arms etc.), characteristics of the populations (mean age, number of enrolled patients etc.), characteristics of the program (types of systematic interventions, outcomes of intended endpoints, registration information etc.). Outcome estimates were extracted in fully adjusted models. Additionally, we contacted the authors if there were any miss data. If we received no response, analysis was performed without these data. Intent-to-treat data were used when available.
Risk of bias of included RCTs was assessed using the modified Cochrane Collaboration's Risk-of-Bias Tool [3]. Two coauthors performed quality assessment on all the included RCTs. The Cochrane Risk of Bias Tool was adopted to assess risk of bias for each RCT. Seven items were used to evaluate heterogeneity in each trial: randomization sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting, and other biases. The quality of each study was categorized as high, low, or unclear. In case of disagreement, the two authors would recheck the original articles and a consensus would be achieved after a discussion.

Statistical analysis
For PFS and OS, the hazard ratios (HR) and confidence intervals (CI) were directly extracted from the original studies or were calculated by methods provided by Tierney et al [4]. We also tried to contact the authors if the study provided only figures without exact data. In case the authors did not respond, the program Engauage Digitizer 4.1 (http://digitizer.sourceforge.net) was run to extract the exact data from the figures. This program can calculate clear values by digitizing data points from an image file after the manual setting of the coordinate axis. Odd ratios (ORs) for ORR and grade 3 or higher AEs were manually calculated based on extracted information.
A Bayesian network meta-analysis (NMA) was performed with a random effects model to estimate the HR and 95% credible interval (95% CrI) for direct and indirect evidence on advanced EGFR-mutant NSCLC by combining multiple systematic arms across studies with all the information regarding PFS and OS. In the case of multi-arm trials (trials with three or more interventions), adjustments were made to preserve randomization and correlation within the multi-arm trials by converting log-HRs to log-hazards. ORs and 95% CrI in random effects model were prepared for ORR and grade 3 or higher AEs for direct and indirect evidence; mean difference (MD) and 95% CrI in random effects was conducted for DoR, because DoR was a continuous variable. Following Cochrane Handbook, standard deviation (SD) was roughly computed by (Xmax-Xmin)/range difference for further analysis. The Markov Chain Monte Carlo (MCMC) method was used to estimate the posterior distribution of each parameter, the fit of the random effects model was assessed by the deviance information criteria (DIC) [5][6][7][8]. A three-chain model with non-informative priors was run with an adaptation phase of 10000 iterations followed by 100000 model iterations. The thin ratio was set to 10. Non-convergence was assessed by the Gelman-Rubin statistic. Relative treatment rankings (probability for each treatment to be the most effective (first best regime), the second best, the third best and so on) were displayed graphically with rankograms, which indicated the probable best and worst therapies [9]. A hierarchical Bayesian model synthesizes comparisons between the treatment pairs and simultaneously summarizes all outcomes of interest by assuming a common heterogeneity parameter (a derived I 2 statistic > 50% or a P value for Cochran Q chi-square test <0.1 was regarded as indicating significant heterogeneity), the inconsistency of this model was evaluated by the edge-splitting method based on all direct and indirect evidence [3,8]. Trace, density and comsol estimations/plots were used to inspect the uncertainty of the MCMC model [5][6][7][8]. To confirm the robustness of our findings, sensitivity analyses were performed restricted on phase III trials, studies excluding In the Bayesian context, statistical significance of HRs and ORs was established when the 95% CrI did not contain 1, of MDs was established when did not contain 0. Calculations were performed in R version 3.5.3 (https://www.r-project.org) using the gemtc and jag etc. public packages. Tables   Please browse Full Text version to see the data of Supplementary Tables 1, 2