Comparison of immune checkpoint inhibitors related to pulmonary adverse events: a retrospective analysis of clinical studies and network meta-analysis

We searched the PubMed, Embase, Cochrane Library, PsycINFO, CINAHL, Scopus, and Web of Science databases for relevant articles published from January 2011 to April 2023, with no language restrictions (see Additional File 1: Supplementary Methods for the specific search strategies employed). We identified additional studies by searching the ClinicalTrials.gov website (https://​clinicaltrials.​gov/​). Eligible studies had to be phase II and III randomized clinical trials (RCTs) and single-arm clinical trials. In addition, the following inclusion criteria had to be met: (1) patients: ≥ 18 years of age with a tumor; (2) intervention: at least one treatment arm received ICIs (e.g., ICI monotherapy, a combination of two ICIs, ICIs in combination with chemotherapy or targeted therapy); (3) comparison: blank control group/placebo, chemotherapy, targeted therapy drug or any intervention containing an ICI; and (4) results: PAEs (pneumonitis, interstitial lung disease) determined using Common Terminology Criteria for Adverse Events. We excluded non-randomized studies, cohort studies, and phase I trials.

Two independent investigators (BH and BD) extracted and assessed the data. BH used standardized data extraction forms to extract relevant information from the included trials, and BD independently reviewed the data. Any disagreements were resolved through discussion. The collected information included basic study characteristics (author, publication year, National Clinical Trial number, trial name, sample size, age, tumor type, line of treatment, disease stage, follow-up time, smoking history, and previous radiation therapy), specific medication regime, and PAEs.

PAEs relating to pulmonary toxicity include pneumonitis and interstitial lung disease, excluding infection-related pulmonary diseases. The outcome measures were treatment-related PAEs. Treatment-related PAEs refer to any PAEs for which treatment cannot be ruled out as a cause. The primary outcome of interest was treatment-related PAEs. The secondary outcomes were fatal PAEs (i.e., Grade 5 PAEs). Treatment-related PAEs were further divided into all-grade and grades 3–4. Fatal PAEs caused death from treatment-related pulmonary toxicities. We only extracted fatal events associated with treatment-related pulmonary toxicities, which were stated in the respective studies.

We independently assessed the quality of the RCTs blinded. BH and BD used the Cochrane Risk of Bias Tool v.2 to assess the risk of bias in each RCT in the NMA.

The incidence rates for all outcomes were calculated using the SPSS software version 26.0. We fit the raw data to a normal distribution after the logit transformation [15]. Further, to determine the corresponding 95% confidence intervals (CIs) for each incidence, we used mixed-effects logistic regression.

Our Bayesian NMA was performed using the Markov Chain Monte Carlo method [16] with R software version 3.5.3, which compared all included interventions. We used a random effect generalized model, running 100,000 inference iterations on each of the four chains. The first 50,000 iterations were discarded as burn-in to obtain the posterior distribution. The Gelman–Rubin method was used to check the convergence of the model by combining density and bundle maps [17]. Evidence relationships were summarized using a network diagram. The PAEs of the different interventions were ranked according to their surface under the cumulative ranking curve (SUCRA) values [18], and the ranking results were visualized using a heat map. The league table compared all treatment regimens. Based on whether the log odds ratio (LOR) passed the zero value, we judged whether there was a significant difference between different interventions [19]. We used bias information criteria to provide a measure of model fit to select a random effects model, a fixed effects model, or an inconsistent model [20]. We then tested the assumption of local consistency via node segmentation [20, 21]. To explore heterogeneity sources, we performed subgroup analyses according to the cancer type and different doses of drugs. To rule out small studies that produced different results, we explored the possibility of publication bias by visually comparing the adjusted funnel plots and performing Egger’s test [22, 23]. Differences were considered statistically significant at P < 0.05.

An initial literature search yielded 100,932 records. After excluding duplicate studies, further screening of the titles and abstracts yielded 185 potentially eligible studies. Of these, 167 were included in this meta-analysis (Additional File 1: Figure S1), with 83,181 participants. Additional File 1: Table S1 shows the PRISMA flow diagram, and Additional File 1: Table S2 presents the baseline characteristics of the patients. Additional File 1: Table S3 shows the results of individual studies.

The quality of the trials was considered acceptable. Many studies had an unclear risk of bias in outcome measurements (37.8%) and missing outcome data (41.3%) (Additional File 1: Figure S2). Moreover, the funnel plot in this NMA showed rough symmetry (Additional File 1: Figure S3). Egger’s test did not reveal significant publication bias in most outcomes, showing P < 0.05 only in fatal PAEs and all-grade PAEs of respiratory system cancer.

The incidence of grade 3–4 PAEs was 1.06%. The incidence of all-grade PAEs was 2.81%. The incidence of fatal PAEs was 0.13% (Fig. 1a). The incidence of PAEs was highest in patients receiving triple therapy (all-grade: 5.15%; grades 3–4: 2.15%) and lowest in patients receiving single-ICIs (all-grade: 3.33%) and ICI plus targeted therapy drug (grade 3–4:1.21%) (Fig. 1b). From the perspective of the treatment regimen, patients receiving anti-PDL1 plus anti-CTLA4 and plus chemotherapy (all-grade: 6.88%) and anti-PD1 plus anti-CTLA4 (grades 3–4: 3.51%) had the most elevated incidence rate of PAEs; patients receiving targeted therapy drug plus chemotherapy (all-grade: 0.49%) and anti-PD1 plus anti-PDL1 (grades 3–4: 0%) had the lowest incidence of PAEs (Fig. 1c and d).

Subgroup analysis of PAEs was performed according to the tumor location, such as the respiratory, genitourinary, skin, head and neck, and digestive systems. Additional File 1: Figure S4a–e show the network plots, and Additional File 1: Figure S5 shows the SUCRA values. Additional File 1: Tables S5 and S6 are league tables. Among patients with respiratory tumors, those who received anti-PD1 plus anti-CTLA4 and plus chemotherapy showed the greatest risk of all-grade PAEs, which was significantly different from the other interventions. The regimen with the greatest risk of grade 3–4 PAEs was anti-PD1 plus anti-CTLA4 and plus chemotherapy, which differed significantly from the other interventions except for targeted therapy drug and anti-PDL1 plus targeted therapy drug. Patients with urogenital tumors treated with anti-PDL1 plus targeted therapy drug had the greatest risk of developing all-grade PAEs, and it differed significantly from the risks associated with anti-PDL1 plus chemotherapy, anti-PD1 plus chemotherapy, targeted therapy drug, and placebo. The greatest risk of grade 3–4 PAEs in patients with urogenital cancer was related to anti-PDL1 plus targeted therapy drug, which had significant differences with the other interventions except for anti-PD1, anti-PD1 plus targeted therapy drug, and anti-PDL1 plus anti-CTLA4. The greatest risk of all-grade PAEs in patients with skin cancer was related to anti-PDL1 plus targeted therapy drug, which differed significantly from the other interventions except for targeted therapy drug and anti-PD1 plus anti-CTLA4. The greatest risk of grade 3–4 PAEs in patients with skin cancer was from anti-PDL1 plus targeted therapy drug, which differed significantly from the other interventions except for anti-CTLA4 plus chemotherapy and anti-PD1 plus anti-CTLA4. The greatest risk of all-grade PAEs in patients with head and neck cancer was linked to anti-PD1 plus targeted therapy drug, which had significant differences with the other interventions except for anti-PDL1 plus chemotherapy. The greatest risk of grade 3–4 PAEs in patients with head and neck cancer was linked to anti-PD1 plus targeted therapy drug, which had significant differences with the other interventions except for anti-PD1 plus chemotherapy. Furthermore, the greatest risk of all-grade PAEs in patients with digestive system tumors was linked to anti-PDL1 plus anti-CTLA4, which was significantly different from the other interventions except for anti-PD1-based interventions. Finally, the greatest risk of grade 3–4 PAEs in patients with digestive system tumors was associated with anti-PD1 plus anti-CTLA4, which differed significantly from the other interventions except for anti-PDL1 plus anti-CTLA4, and anti-PD1 plus targeted therapy drug.

We also performed subgroup analyses based on ICI doses (Additional File 1: Figure S6). Compared with placebo, nivolumab 360 mg q3w showed a significant difference in the risk of PAEs in all-grade, whereas nivolumab 240 mg q2w did not. When the intervention group was atezolizumab plus targeted therapy drug and the control group was targeted therapy drug, the 840 mg q2w atezolizumab regimen was significantly different in the risk of all-grade PAEs compared with targeted therapy drug. There was no significant difference between the 1200 mg q3w atezolizumab treatment and targeted therapy drug. Compared with chemotherapy, the pembrolizumab 2 mg/kg q3w regimen had a statistically significant difference in the risk of grade 3–4 PAEs, and the treatment with pembrolizumab 10 mg/kg q3w treatment did not differ significantly from the control group.

According to the DIC values, the stochastic consistency model is the preferred model with a good complexity trade-off between model fitting and model (Additional File 1: Table S7). Bayesian network analysis revealed that the shrinkage factors of Brooks–Gelman–Rubin diagnostic graphs were < 1.2 (Additional File 1: Table S7), indicating that the research model had good convergence. Values with I² > 50% were not observed; consequently, significant heterogeneity was not observed in PAEs. The node splitting method demonstrated no significant inconsistencies in the results (Additional File 1: Table S8). Therefore, the overall results were considered relatively robust.

For ICI monotherapy, anti-PD1 had the highest incidence of all-grade PAEs. However, there was no significant difference between anti-PD1 and anti-PDL1 in the incidence of all-grade PAEs. In addition, the incidence of ICI monotherapy varied across studies owing to the different dose regimens. Here, the risk of all-grade PAEs was dose-dependent for nivolumab, pembrolizumab, and atezolizumab, whereas no significant difference in PAEs was observed between dose groups for the other ICIs. By contrast, Wu et al. hypothesized that the incidence and intensity of PAEs induced by anti-PD1 are independent of drug dose [26]. However, this difference may be due to the inclusion of more recent studies in our analysis and consideration of treatment-related PAE differentiation. In addition, our control group was administered the same drug. Wu et al. only included studies published until 2016, and the types of drugs in the control group were not consistent. Therefore, it is not surprising that the incidence of PAEs varied with regimens. Higher doses of nivolumab and a low dose and high frequency of atezolizumab administration in combination with targeted therapy may also be a factor affecting the risk of all-grade PAEs. Therefore, to prevent further aggravation of PAEs in patients using anti-PD1/anti-PDL1 with different usage and dosage, clinicians must focus on minor changes in these patients, including timely diagnosis and intervention.

Recently, the FDA and EMA have approved several ICIs in combination with chemotherapy or targeted therapies. In addition, some cancers already use combination therapy as the standard of care. For combination treatment regimens, we have found the following points. First, anti-PD1 plus anti-CTLA4 and plus chemotherapy had the greatest risk of PAEs, which differed significantly from the other regimens. This may be due to the increased risk of PAEs associated with drug combinations with different mechanisms of action. Association of anti-CTLA4 and anti-PD1 enhances T cell activation and proliferation. This increases the production of proinflammatory cytokines [27]. Second, the incidence of PAEs was much higher with anti-PD1 plus anti-CTLA4 than with ICI monotherapy, but there was no significant difference in the risk of PAEs. By contrast, Nishino et al. suggested that the incidence of developing pneumonitis is higher with combination therapy than with ICI monotherapy. This may be because their study was limited to three ICI molecules, the anti-PD1 nivolumab, pembrolizumab, and the anti-CTLA4 agent ipilimumab. However, we also included atezolizumab, avelumab, durvalumab, tremelimumab, camrelizumab, cemiplimab, sintilimab, and tislelizumab. Third, the risk of all-grade PAEs was significantly lower with ICI plus chemotherapy than with dual-ICIs, and the risk of all-grade PAEs was significantly lower with anti-PD1 plus chemotherapy than with anti-PD1 monotherapy. These observations are in agreement with those of Chen et al. [28]. A possible reason is that the traditional chemotherapy regimen primarily consists of cytotoxic drugs, which induce immunosuppression and lead to decreased immune function, thereby reducing the risk of immune-related adverse reactions [29, 30]. Another possible factor is that the combination of pre-treated glucocorticoids in the chemotherapy regimen reduces the risk of PAEs. Glucocorticoids can not only suppress the immune system [31, 32] but also have certain therapeutic effects on some lung diseases, including ICI-related pneumonitis [33, 34]. However, the mechanism of how cytotoxic drugs and corticosteroids combined with ICI regulate the immune system remains unclear. More prospective studies of the mechanisms described above are essential to maximize anti-cancer benefits while minimizing the risks of PAEs. Moreover, we included the latest drug therapy combination, anti-PD1 plus anti-PDL1. Interestingly the incidence of all-grade PAEs is not low, but the risk of grade 3–4 PAEs is low. Although the underlying mechanism is still unclear, our results provide clinicians with the latest reference for PAEs of drug combinations.

Our study has several strengths as we rigorously and comprehensively searched the literature for patients with ICI-treated cancer. We included cohorts of patients worldwide, making our findings more generalizable. Additionally, several studies with relatively large sample sizes were included, which increased the statistical power of our meta-analysis. However, our study also had a few limitations, which may be methodological flaws. First, the studies we included were all RCTs; therefore, the baseline conditions of the patients were highly selected and may not reflect the real-world situation. Second, PAEs include a variety of pulmonary toxicity outcomes, and the measurement criteria of different outcomes and the number of trials by various authors lacked some transparency in each study. Finally, we did not investigate the performance of different Bayesian meta-analytic methods in rare events [19].