Risk of immune‐related pneumonitis for PD1/PD‐L1 inhibitors: Systematic review and network meta‐analysis

Abstract Background Immune‐related pneumonitis is a clinically relevant and potentially life‐threatening adverse event. We performed a systematic review and network meta‐analysis to compare the risk of immune‐related pneumonitis among different PD1/PD‐L1 inhibitor‐related therapeutic regimens. Methods Randomized controlled trials with PD1/PD‐L1 inhibitors were identified through comprehensive searches of multiple databases. Both published and unpublished data were extracted. Bayesian NMA was performed using random‐effects models. All‐grade (Grade 1‐5) and high‐grade (Grade 3‐5) immune‐related pneumonitis were estimated using odds ratios (ORs). Results A total of 25 studies involving 16 005 patients were included. Compared with chemotherapy, the ORs of immune‐related all‐grade and high‐grade pneumonitis were significant for nivolumab (all‐grade: OR = 6.29, 95% CrI: 2.67‐16.75; high‐grade: OR = 5.95, 95% CrI: 2.35‐17.29), pembrolizumab (all‐grade: OR = 5.78, 95% CrI: 2.79‐13.24; high‐grade: OR = 5.33, 95% CrI: 2.49‐12.97), and nivolumab plus ipilimumab therapy (all‐grade: OR = 14.82, 95% CrI: 5.48‐47.97; high‐grade: OR = 15.26, 95% CrI: 5.05‐55.52). Compared with nivolumab, nivolumab plus ipilimumab therapy was associated with an increased risk of all‐grade pneumonitis (OR = 2.34, 95% CrI: 1.07‐5.77). Nivolumab plus ipilimumab therapy had the highest risk of both all‐grade and high‐grade pneumonitis among PD1/PD‐L1 inhibitor‐related therapeutic regimens. Conclusions This study demonstrates that compared with chemotherapy, PD‐1 inhibitor may result in a higher risk of immune‐related pneumonitis. Nivolumab plus ipilimumab therapy had the highest pneumonitis risk. These findings could be taken into account by the physicians in decision making.


| Study design
We conducted a systematic review with both pairwise metaanalysis and Bayesian NMA. The study was carried out according to the Cochrane handbook for systematic reviews of interventions. 17 We reported the study according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. 18,19 The study was registered in PROSPERO international prospective register of systematic reviews (CRD42018099163).

| Search strategy and selection criteria
We systematically searched PubMed, Embase, Cochrane Central Register of Controlled Trials and ClinicalTrials. gov to identify potentially eligible studies. The terms used for the search strategy included "neoplasm", "cancer", "atezolizumab", "avelumab", "durvalumab", "nivolumab", "pembrolizumab". There was no restriction on language or year of publication. We manually checked reference lists of related review articles to identify additional studies. The final date for the database running searches was June 19th, 2018.
Eligible studies had to be RCTs and should include either anti-PD-1 or anti-PD-L1 monoclonal antibody (ie atezolizumab, avelumab, durvalumab, nivolumab, pembrolizumab), alone or in combination with other types of treatment, in the intervention or control group. We evaluated the rates of immune-related pneumonitis reported. We excluded studies only in abstract form and studies of quality of life or cost effectiveness analyses.

| Study selection and data extraction
Two independent investigators (HY and FH) selected the potentially eligible studies and extracted the data from all the eligible studies. The titles, abstracts, and full-text records were evaluated sequentially. The following information was extracted: title, trial name, year, funding sources, line of treatment, blinding, age, sex, tumor type, length of follow up, types and dosage of drugs, number of patients in the treatment and control arms, number of patients with pneumonitis of all-grade (grade 1-5) and high-grade (grade [3][4][5] in the treatment and control arms. Both published and unpublished data were extracted. The unpublished data were extracted from ClinicalTrials.gov.

| Quality assessment
Two authors (HY and FH) independently assessed the risk of bias of included studies based on the Cochrane Collaboration's tool. 17 Disagreement was resolved by discussion with a third author (LN). Six domains were evaluated: sequence generation, allocation concealment, blinding, incomplete outcome data, selective reporting, and other sources of bias.
We used the Grading of Recommendations Assessment, Development and Evaluation system (GRADE) approach to rate the quality of evidence. 20 There were 4 levels of quality of evidence: high, moderate, low, and very low. The quality of evidence for each outcome was based on the fundamental study design and additional methodological factors. 20

| Outcome measures
The primary outcome of interest was all-grade (grade 1-5) pneumonitis. Secondary outcome was high-grade (grade 3-5) pneumonitis based on the National Cancer Institute Common Terminology Criteria for Adverse Events version 4.0.

| Statistical analysis
DerSimonian-Laird random effects model was used to perform traditional pairwise meta-analyses. Summary effect size was presented as odds ratio (OR) for binary outcomes along with corresponding 95% confidence intervals (CIs). A 2-sided P value of less than 0.05 or 95% CIs excluding one was regarded as statistically significant. Heterogeneity F I G U R E 1 Literature search and selection F I G U R E 2 Network of eligible comparisons for the Bayesian network meta-analysis. The size of the nodes is proportional to the number of patients (in parentheses) randomized to receive the treatment. The width of the lines is proportional to the number of comparisons (beside the line) comparing the connected treatment (nodes). A total of 25 randomized controlled trials included 30 comparisons were analyzed among studies was assessed using Cochrane Q statistic and quantified with I 2 statistic. Values over 50% indicated substantial heterogeneity. 17 Publication bias was examined using funnel plots. 21 Bayesian NMA allowing for indirect comparisons was performed to evaluate the risk of pneumonitis using a Markov Chain Monte Carlo (MCMC) simulation technique. We estimated the posterior distribution of all parameters using vague priors. We updated the MCMC model with 100000 simulated draws after a burn-ins of 10000 iterations and used a thinning interval of 10 for each chain. We tested the adequacy of burn-in and convergence using the Brooks-Gelman-Rubin statistic. 22 Relative effects of treatments were reported as OR for binary outcomes along with corresponding 95% credible intervals (CrIs). All analyses were performed under the random-effects model since they generally showed more conservative estimated effects and better goodness of fit. We calculated the posterior mean of the residual deviance to determine goodness of fit of the models. Ideally, each data point should contribute about one to the posterior mean of the residual deviance. Therefore, it can be compared with the number of data points for model fit checking. We used the loop-specific approach to evaluate the presence of inconsistency. 23 We calculated the values of 2 odds ratios (RoR) from direct and indirect evidence in the loop with 95% CI to assess the presence of inconsistency in each loop. Inconsistency was defined as disagreement between direct and indirect evidence with a 95% CI excluding zero.
Meta-regression analyses were performed by adding prespecified covariates (ie median age, percentage of male, line of treatment, study phase, whether double-blind was used) to the network meta-analysis models. Sensitivity analyses were performed to evaluate the robustness of results based on the events reported in published articles only. The data analyses were conducted using STATA version 14.0 and WinBUGs version 1.4.3.

| Eligible studies and patient characteristics
Initial search identified 5566 records. 4023 potentially eligible articles were retrieved for detailed assessment. Thirty articles reporting 25 studies were included for meta-analysis ( Figure 1, Table S1). The 25 studies covered 12 types of treatment and involved a total of 16 005 cancer patients ( Figure 2). The baseline characteristics of the included studies are listed in Tables 1 and 2 and Table  S2. The risk of bias summary for included studies is listed in Table S3.
Results of NMA for all-grade pneumonitis risk were displayed in Figure 3A and Figure S1. Compared with chemotherapy, nivolumab and pembrolizumab were associated with an increased risk of all-grade pneumonitis (nivolumab vs chemotherapy: OR = 6.29, 95% CrI: 2.67-16.75; pembrolizumab vs chemotherapy: OR = 5.78, 95% CrI: 2.79-13.24), with moderate quality of evidence respectively. Compared with chemotherapy, nivolumab plus ipilimumab therapy was associated with an increased risk of all-grade pneumonitis (OR = 14.82, 95% CrI: 5.48-47.97), with low quality of evidence. Compared with nivolumab, nivolumab plus ipilimumab therapy was also associated with an increased risk of all-grade pneumonitis (OR = 2.34, 95% CrI: 1.07-5.77), with high quality of evidence. Compared with nivolumab plus ipilimumab therapy, pembrolizumab plus chemotherapy was associated with a decreased risk of all-grade pneumonitis (OR = 0.18, 95% CrI: 0.04-0.89), with moderate quality of evidence.
Results of NMA for high-grade pneumonitis risk were displayed in Figure 3B and Figure S2. Compared with chemotherapy, nivolumab and pembrolizumab were associated with an increased risk of high-grade pneumonitis (

| Ranks and meta-regression analyses
The ranks of all treatments were presented in Table S6. Nivolumab plus ipilimumab therapy had the highest risk of both all-grade and high-grade pneumonitis among PD1/PD-L1 inhibitor-related therapeutic regimens. Meta-regression analyses did not reveal any prespecified factors that influenced the estimated effects significantly (Table S7).

| Model fit and inconsistence check
The model fit was evaluated using the posterior mean of the residual deviance, which was 42 and 38 for all-grade and high-grade pneumonitis, respectively. The model's overall fit was relatively satisfactory. According to the forest plots, the statistically inconsistency between direct and indirect comparisons was low for all-grade and high-grade pneumonitis outcomes. All loops were consistent ( Figure S3). Figure S4 present the adjusted funnel plot for the pneumonitis network. The funnel plots of all-grade and high-grade pneumonitis outcomes did not show asymmetry, suggesting no potential risk of reporting bias. Sensitivity analyses based on published data did not indicate any major influence on the outcomes ( Figure S5-S7).

| Summary of key findings
This study has 3 key findings: First, there was moderate quality of evidence that nivolumab, pembrolizumab and nivolumab plus ipilimumab therapy increased the risk of all-grade and high-grade immune-related pneumonitis, compared with chemotherapy. Second, nivolumab plus ipilimumab therapy was associated with an increased risk of all-grade pneumonitis compared with nivolumab, with high quality of evidence. Third, nivolumab plus ipilimumab therapy had the highest risk of both all-grade and high-grade pneumonitis among different types of PD1/PD-L1 inhibitorrelated therapeutic regimens.

| Comparison with other studies
Previous published systematic reviews and meta-analyses regarding the immune-related risk of pneumonitis have shown that PD1 inhibitors are associated with an increased risk of immune-related pneumonitis compared with chemotherapy. 8,15,24 The Bayesian network meta-analysis in our study allows us to compare the therapeutic regimens indirectly when no head-to-head trial existed. There was high to moderate quality evidence showing that nivolumab plus ipilimumab therapy was associated with an increased risk of pneumonitis, compared with chemotherapy and nivolumab respectively. Two studies have previously investigated the differences in the toxicities of PD1 and PD-L1 inhibitors. 24 reported a higher incidence of immune-related pneumonitis with use of PD-1 inhibitors compared with PD-L1 inhibitors in patients with nonsmall cell lung cancer. 24 The summary of incidence of all-grade and high-grade pneumonitis was also reported in Table S8 of our study. Pillai et al found a slight increase in pneumonitis risk with PD-1 inhibitors. 25 It should be noted that no RCT to date has directly compared the risk of immune-related pneumonitis between PD-1 and PD-L1 inhibitors. We still lack direct evidence to make such conclusion. In our study, indirect comparison showed an increased but not statistically significant risk of immune-related pneumonitis with use of nivolumab or pembrolizumab compared with durvalumab or atezolizumab, respectively.

| Strength and limitations of study
To our knowledge, this is the first systematic review and NMA which provides the most current and structured evidence of immune-related pneumonitis for PD1/PD-L1 inhibitor-related therapeutic regimens. Studies from ClinicalTrials. gov were searched carefully. Both published and unpublished data were extracted. We believe that all relevant RCTs were included in analyses and the publication bias was reduced as much as possible. The main limitation of this study is that indirect comparisons from NMA are very likely to suffer bias through confounding by study-level characteristics.
The results from indirect comparisons should be interpreted with caution as direct evidence is lacking. However, the trial populations and study characteristics were very comparable to the target population of this NMA. We further evaluated the potential confounding factors with meta-regression analyses, which showed no major influence to our primary results. Secondly, compared with chemotherapy, we did not find significant increase in pneumonitis risk for atezolizumab or durvalumab. It did not indicate that there was no risk of pneumonitis with the use of these 2 agents. Compared with chemotherapy, the direction of risks for durvalumab and atezolizumab were both positive (more than one), despite of lacking significance. One of the potential reasons of lacking significance may due to limited number of trials with positive results for these 2 agents. More RCTs are needed in the future to detect their potential risks further.

| Research and clinical implications
Traditionally, most assessments of safety of PD1/PD-L1 inhibitors come from comparisons with chemotherapy. Despite our current systematic review and NMA provides insight in these comparisons from indirect comparisons, evidence on head-to-head comparisons among different PD1/ PD-L1 inhibitor-related therapeutic regimens is still lacking. New trials comparing between different PD1/PD-L1 inhibitor-related therapeutic regimens are needed. Future trials could also be conducted to assess the safety of the combination of PD1/PD-L1 inhibitor and chemotherapy to enrich the evidence. Two clinical implications should be noted. First, nivolumab plus ipilimumab therapy had the highest pneumonitis risk among different PD1/PD-L1 inhibitor-related therapeutic regimens. The results of both all-grade and high-grade pneumonitis outcomes were stable in sensitivity analysis. It should be noted that ipilimumab blocks cytotoxic T-lymphocyte antigen-4 (CTLA-4) as well as augments T-cell immune response as an immunomodulator. CTLA-4 has both cell intrinsic activities and cell extrinsic activities. In contrast, immunoregulation by PD-1 is antigen specific and cell intrinsic. 26 Consistent with their mechanism of action, immune-related adverse event rates are more likely to be higher for the combination use of  PD1 and CTLA4 inhibitors, compared with using PD1 inhibitors alone. Moreover, immune-related all-grade and highgrade pneumonitis were also significant for both nivolumab and pembrolizumab therapy. Physicians may need to consider the increased pneumonitis risk when choosing these therapies for cancer patients. Second, the combination of chemotherapy and PD1/PD-L1 inhibitor had a decreased risk of immunerelated pneumonitis, compared with using nivolumab plus ipilimumab. These results may be taken into account by the physicians in decision making when choosing among the different combinations of therapies.

| CONCLUSIONS
This systematic review and network meta-analysis offers substantial evidence and demonstrates that PD-1 inhibitor is very likely to result in a higher risk of immune-related pneumonitis compared with chemotherapy. Nivolumab plus ipilimumab therapy had the highest pneumonitis risk. These findings may be taken into account by the physicians in decision making when choosing among different PD1/PD-L1 inhibitor-related therapeutic regimens for cancer patients.

ACKNOWLEDGMENT
We gratefully thank the 2 anonymous reviewers whose comments helped a lot in improving this manuscript.