Single-agent maintenance therapy for advanced non-small cell lung cancer (NSCLC): a systematic review and Bayesian network meta-analysis of 26 randomized controlled trials

Search strategy

We conducted a comprehensive literature search of PubMed, EMBASE, and the Cochrane Central Register of Controlled Trials (CENTRAL) from inception to November 09, 2015. We administered a high sensitivity searched strategy with keywords set around “non-small cell lung cancer,” “maintenance therapy,” and “RCT.” No language restriction was administered. Details of the search strategy are presented in Supplemental Information 1. We also manually searched proceedings of the annual American Society of Clinical Oncology (ASCO) meetings and European Society of Medical Oncology (ESCO) congresses from 2000 to 2015 as a supplement. Citations of relevant reviews and trials were also screened. All results were input into Endnote X7 reference software (Thomson Reuters, Stamford, CT, US) for duplication exclusion and further reference management.

Selection of trials

Studies meeting the following inclusion criteria were eligible: (i) RCTs, (ii) patients were pathologically or cytologically-diagnosed with non-resectable stage III or IV NSCLC, (iii) comparisons had to be between single-agent maintenance therapy and placebo, observation, or another single-agent maintenance regimen, and (iv) sufficient data on OS or/and PFS. Trials with randomization conducted before induction therapy and trials of complementary medicine were excluded. When multiple publications reported on one trial, we selected the most recent report for data extraction.

Data extraction and risk of bias assessment

Data was independently extracted by two reviewers (Q Wang and H Huang) using standardized data compilation forms. Name of the first author, publication year, number of patients and population characteristics, induction and maintenance therapy regimens, survival statistics, adverse effects (AEs) were major aspects included. The hazard ratios (HRs) and 95% confidence intervals (CIs) were either obtained from the original articles or estimated from Kaplan–Meier curves using Tierney’s spreadsheet (Tierney et al., 2007). For each included trial, the following domains of bias were judged and ranked into “low risk,” “high risk,” or “unclear risk”: generation of random sequence, allocation concealment, blinding, incomplete outcome data, selective reporting of outcome, and other biases. Two investigators (Q Wang and H Huang) independently performed the assessment. All divergences during data extraction and assessment of risk of bias were solved by discussion with a third investigator (M Huang).

Statistical analysis

The NMA combined evidence from head-to-head comparisons into a network to obtain estimates of the relative efficacy of each treatment. Analyses were conducted using R 3.0.1 (R Development Core Team, 2013) with an interface to WinBUGS 1.4.3 (Medical Research Council Biostatistics Unit, Cambridge, UK). We built a network within the Bayesian framework and the posterior distribution of the treatment effect was estimated using Markov Chain Monte Carlo methods.

For the three-arm trial Perol 2012, log HRs (contrast statistics) were converted to log HRs (arm-specific statistics) according to the method introduced by Woods, Hawkins & Scott (2010).

All analyses were performed with 2 chains, and each had a sample of 200,000 simulations after discarding the results of a burn-in period of 40,000 simulations. We estimated the relative treatment effects based on the posterior distributions and ranked the probability for each treatment in descending order as the most efficacious regimen, the second, the third, and so on, according to OS and PFS separately. Since OS is the most concerned outcome in clinical trials of antitumor therapy, we set OS as the primary outcome in our analysis and draw conclusions based on OS mainly. The deviance information criterion (DIC) provided a measure of model fit that a lower value suggested a simpler model. Convergence of the model was assessed with the Brooks–Gelman–Rubin diagnostic methods in WinBUGS.

Both fixed and random effects models were administered in the primary analysis; posterior mean of the residual deviance (resdev), effective number of parameters (pD), and DIC results of the two models were compared in sensitivity analysis. We assessed the inconsistency between direct and indirect evidence using the method suggested by Veroniki et al. (2013). We also assessed the probability of publication bias with contour-enhanced funnel plots (Peters et al., 2008).

Quality assessment of evidence

We assessed the quality of evidence in two steps. First, we used the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) system to assess quality of direct evidence (Guyatt et al., 2011). GRADE focuses on a body of evidence rather than individual studies. RCTs were initially identified as high quality of evidence and identification of problems on limitations in trial design, inconsistency, indirectness, imprecision and publication bias decreased the evidence quality rating. Quality of evidence was rated as high, moderate, low or very low. Then, we used the iGRADE approach, which is a modification of the GRADE approach for mixed treatment comparisons proposed by Dumville et al. (2012), to evaluate the quality of NMA evidence.

Characteristics of eligible studies

Through online databases and meeting abstracts searches, a total of 1,368 records were identified. After rounds of assessment, 26 trials covering 7,839 patients met all the inclusion criteria, and comprised of 24 complete manuscripts (Ahn et al., 2013; Alfonso et al., 2014; Belani et al., 2003; Brodowicz et al., 2006; Butts et al., 2005; Butts et al., 2014; Cai et al., 2015; Cappuzzo et al., 2010; Carter et al., 2012; Ciuleanu et al., 2009; Fidias et al., 2009; Gaafar et al., 2011; Giaccone et al., 2015; Hanna et al., 2008; Hu et al., 2010; Johnson et al., 2008; Karayama et al., 2013; Kelly et al., 2008; Mubarak et al., 2012; O’Brien et al., 2015; Pérol et al., 2012; Paz-Ares et al., 2013; Westeel et al., 2005; Zhang et al., 2012) and 2 meeting abstracts (Belani et al., 2010; Socinski et al., 2014). Selection procedure is summarized in Fig. 1. Summary of characteristics of the 26 eligible studies and HR data of each individual study is shown in Table 1. With the exception of Perol 2012 which was a three-arm trial (continue-gemcitabine or switch-erlotinib vs. observation) and Karayama 2013 which compared two maintenance regimens directly (continue-pemetrexed vs. switch-docetaxel); the remaining 24 trials all compared single-agent maintenance therapy vs. no-maintenance control. The network of evidence constructed by the included RCTs is shown in Fig. 2. Risks of bias of the enrolled studies are depicted in Table S1.

OS and PFS analyses

In total, 24 trials were included in the OS analysis (Ahn et al., 2013; Alfonso et al., 2014; Belani et al., 2003; Belani et al., 2010; Brodowicz et al., 2006; Butts et al., 2005; Butts et al., 2014; Cappuzzo et al., 2010; Carter et al., 2012; Ciuleanu et al., 2009; Fidias et al., 2009; Gaafar et al., 2011; Giaccone et al., 2015; Hanna et al., 2008; Hu et al., 2010; Johnson et al., 2008; Karayama et al., 2013; Kelly et al., 2008; Mubarak et al., 2012; O’Brien et al., 2015; Pérol et al., 2012; Paz-Ares et al., 2013; Westeel et al., 2005; Zhang et al., 2012). No-maintenance control was set as the reference in all analyses. Based on assessment of model fit, results calculated by random effects models are presented in this section. The HRs for different maintenance regimens compared to no-maintenance are shown in Fig. 3A. Several maintenance therapy regimens yielded longer OS than no-maintenance, although differences were not statistically significant in some regimens. Switch-docetaxel, continue-paclitaxel, switch-sunitinib, switch-vandetanib, switch-carboxyaminoimidazole (CAI), and switch-vinorelbine did not improve OS. Switch-maintenance therapy with racotumomab-alum vaccine showed excellent efficacy compared to no-maintenance with a HR = 0.64 [95% credible intervals (CrI), 0.45–0.92] Pooled relative treatment effect estimates of all comparisons are presented in Table S2.

In PFS analysis, we included 23 trials (Ahn et al., 2013; Alfonso et al., 2014; Belani et al., 2010; Brodowicz et al., 2006; Butts et al., 2014; Cai et al., 2015; Cappuzzo et al., 2010; Carter et al., 2012; Ciuleanu et al., 2009; Fidias et al., 2009; Gaafar et al., 2011; Giaccone et al., 2015; Hanna et al., 2008; Johnson et al., 2008; Karayama et al., 2013; Kelly et al., 2008; Mubarak et al., 2012; O’Brien et al., 2015; Pérol et al., 2012; Paz-Ares et al., 2013; Socinski et al., 2014; Westeel et al., 2005; Zhang et al., 2012). The HRs for different maintenance regimens compared to no-maintenance regimens are shown in Fig. 3B. Continue-paclitaxel, switch-belagenpumatucel-L, or switch-CAI did not yield longer PFS than no-maintenance. Switch-pemetrexed and switch-gefitinib showed excellent efficacy compared to no-maintenance with HRs = 0.54 (95% CI [0.26–1.04]) and 0.60 (95% CI [0.40–0.90]). Pooled relative treatment effect estimates of all comparisons are presented in Table S2.

Ranking which indicated the probability of the best regimen in descending order, among all treatments is shown in Figs. 3C, 3D and Table 2. According to the rank order based on OS, switch-racotumomab-alum vaccine had the greatest probability as the best regimen (52%), with switch-pazopanib ranked second (32%), and switch-pemetrexed ranked third (6%). Based on PFS, switch-pemetrexed ranked first (34%), followed by switch-sunitinib (19%), with switch-pazopanib ranked third (12%).

Adverse events (AEs)

Maintenance chemotherapy (including pemetrexed, gemcitabine, docetaxel, paclitaxel, and vinorelbine) was commonly associated with hematologic events such as neutropenia, thrombocytopenia, and anemia. Maintenance tyrosine kinase inhibitor (TKI) (including EGFR-TKI and other TKIs) commonly caused more skin and gastrointestinal AEs, such as rash, nausea, and vomiting. Maintenance vaccine (including belagenpumatucel-L, racotumomab-alum, and L-BLP25) was commonly associated with injection site reaction and flu-like symptoms. The main AE of CAI was nausea.

Sensitivity analysis

The primary outcome OS was calculated using both fixed and random effects models. Resdev, pD and DIC were very similar for both models (−23.37, 15.8 and −7.5 in fixed effects model; −23.33, 17.3 and −6.0 in random effects model), which indicated the robustness of results.

Inconsistencies

The data did not suggest any inconsistency between direct and indirect evidence in the network (Fig. 4). In fact, direct evidence of the relative efficacy of different maintenance therapy regimens was rather few in the network. In the analysis of OS, only two closed loops were formed (none vs. continue-gemcitabine vs. switch-erlotinib; none vs. continue-pemetrexed vs. switch-docetaxel).

Publication bias

Symmetry of the ‘comparison-adjusted’ funnel plot suggested that efficacy of the regimens were no more exaggerated than their respective comparison-specific weighted average effect in small studies. The regimens also sorted from oldest to newest, and the resulting ‘comparison-adjusted’ funnel plot did not suggest any publication bias in the network (Fig. 5).

Quality assessment

GRADE and iGRADE analyses are presented in Table S3. Since direct data comparing different maintenance therapy regimens was available for only two couples of regimens, measurement of inconsistencies between direct and indirect data was limited. In general, the most common reasons for lowering the quality of evidence were limitations in trial design and imprecision in some studies. Data suggested that evidence on switch-docetaxel, continue-paclitaxel and switch-vinorelbine were rated as limited quality, while evidence on switch-pemetrexed, switch-belagenpumatucel-L and switch-racotumomab-alum was rated as higher quality.