A, Incidences of the most common all-grade adverse events. B, Incidences of the most common grade 3 or higher adverse events. C, Incidences of the most common all-grade irAEs. D, Incidences of the most common grade 3 or higher irAEs. Vertical lines in A and C indicate the overall mean incidence of all-grade adverse events (1.66%). Vertical lines in B and D indicate the overall mean incidence of grade 3 or higher adverse events (0.11%). Values to the left of the line are lower than the mean, to the right, higher. ALT indicates alanine aminotransferase; AST, aspartate aminotransferase; GGT, γ-glutamyltransferase.
A, Mean incidences of all grade adverse events by cancer type; vertical line indicates the overall mean incidence of all-grade adverse events (1.66%). B, Mean incidences of grade 3 or higher adverse events by cancer type. The vertical line indicates the overall mean incidence of grade 3 or higher adverse events (0.11%). For both panels, values to the left of the line are lower than the mean, to the right, higher.
A, Mean incidences of all-grade adverse events by drug and dose; vertical line indicates the overall mean incidence of all-grade adverse events (1.66%). B, Mean incidences of grade 3 or higher adverse events by drug and dose. C, Comparisons of mean incidences of adverse events between different drugs. The vertical line indicates the overall mean incidence of grade 3 or higher adverse events (0.11%). For A and B, values to the left of the line are lower than mean, to the right, higher. For C, values to the left of the line are higher for second drug in the comparison, to the right, for first drug in comparison. PD-1 indicates programmed cell death; PD-L1, programmed cell death ligand 1.
eTable 1. List of the Studies Included in This Meta-Analysis
eTable 2. Relative Ratios of Grade 3 or Higher Adverse Event Incidences to Respective All-grade Adverse Event Incidences
eTable 3. List of Studies That Reported At Least One Treatment-related Death
eFigure 1. Incidences of All Adverse Events, (A) Incidences of All-grade Adverse Events and (B) Incidences of Grade 3 or Higher Adverse Events
eFigure 2. Mean Incidences of Adverse Events by Study, (A) Mean Incidences of All-grade Adverse Events by Study and (B) Mean Incidences of Grade 3 or Higher Adverse Events by Study
eFigure 3. Heterogeneity of Adverse Event Incidence in This Meta-Analysis
eFigure 4. Mean Incidences of All-grade Adverse Events by Study if Ignoring Missing Data
Customize your JAMA Network experience by selecting one or more topics from the list below.
Wang Y, Zhou S, Yang F, et al. Treatment-Related Adverse Events of PD-1 and PD-L1 Inhibitors in Clinical Trials: A Systematic Review and Meta-analysis. JAMA Oncol. 2019;5(7):1008–1019. doi:10.1001/jamaoncol.2019.0393
What are the incidences of treatment-related adverse events of PD-1 and PD-L1 inhibitors, and do they differ between different drugs and cancer types?
In this systematic review and meta-analysis of 125 clinical trials involving 20 128 patients, the overall incidences of all-grade adverse events were 66.0% and of grade 3 or higher adverse events were 14.0%. The overall mean adverse event incidences were similar across different cancer types but varied between different drugs.
A comprehensive summary of treatment-related adverse events for PD-1 and PD-L1 inhibitors in clinical trials may be an important guide for clinical practice.
Programmed cell death (PD-1) and programmed cell death ligand 1 (PD-L1) inhibitors have been increasingly used in cancer therapy. Understanding the treatment-related adverse events of these drugs is critical for clinical practice.
To evaluate the incidences of treatment-related adverse events of PD-1 and PD-L1 inhibitors and the differences between different drugs and cancer types.
PubMed, Web of Science, Embase, and Scopus were searched from October 1, 2017, through December 15, 2018.
Published clinical trials on single-agent PD-1 and PD-L1 inhibitors with tabulated data on treatment-related adverse events were included.
Data Extraction and Synthesis
Trial name, phase, cancer type, PD-1 and PD-L1 inhibitor used, dose escalation, dosing schedule, number of patients, number of all adverse events, and criteria for adverse event reporting data were extracted from each included study, and bayesian multilevel regression models were applied for data analysis.
Main Outcomes and Measures
Incidences of treatment-related adverse events and differences between different drugs and cancer types.
This systematic review and meta-analysis included 125 clinical trials involving 20 128 patients; 12 277 (66.0%) of 18 610 patients from 106 studies developed at least 1 adverse event of any grade (severity), and 2627 (14.0%) of 18 715 patients from 110 studies developed at least 1 adverse event of grade 3 or higher severity. The most common all-grade adverse events were fatigue (18.26%; 95% CI, 16.49%-20.11%), pruritus (10.61%; 95% CI, 9.46%-11.83%), and diarrhea (9.47%; 95% CI, 8.43%-10.58%). The most common grade 3 or higher adverse events were fatigue (0.89%; 95% CI, 0.69%-1.14%), anemia (0.78%; 95% CI, 0.59%-1.02%), and aspartate aminotransferase increase (0.75%; 95% CI, 0.56%-0.99%). Hypothyroidism (6.07%; 95% CI, 5.35%-6.85%) and hyperthyroidism (2.82%; 95% CI, 2.40%-3.29%) were the most frequent all-grade endocrine immune-related adverse events. Nivolumab was associated with higher mean incidences of all-grade adverse events compared with pembrolizumab (odds ratio [OR], 1.28; 95% CI, 0.97-1.79) and grade 3 or higher adverse events (OR, 1.30; 95% CI, 0.89-2.00). PD-1 inhibitors were associated with a higher mean incidence of grade 3 or higher adverse events compared with PD-L1 inhibitors (OR, 1.58; 95% CI, 1.00-2.54).
Conclusions and Relevance
Different PD-1 and PD-L1 inhibitors appear to have varying treatment-related adverse events; a comprehensive summary of the incidences of treatment-related adverse events in clinical trials provides an important guide for clinicians.
Programmed cell death (PD-1) and programmed cell death ligand 1 (PD-L1) inhibitors have revolutionized cancer therapy.1,2 To date, 2 PD-1 inhibitors (nivolumab and pembrolizumab) and 3 PD-L1 inhibitors (atezolizumab, avelumab, and durvalumab) have been approved by the US Food and Drug Administration for various indications. These drugs work by blocking the PD-1 or PD-L1 immune checkpoint pathway to reactivate T cell–mediated antitumor immunity.2 With reactivation of cellular immunity, these checkpoint inhibitors have been reported to cause autoimmune-like disorders.2,3 Given the increasing use of PD-1 and PD-L1 inhibitors, understanding their toxicologic profile is crucial.
Clinical trials of PD-1 and PD-L1 inhibitors report treatment-related adverse events according to standard guidelines, such as the National Cancer Institute Common Terminology Criteria for Adverse Events, and represent an ideal resource for comprehensive analysis of incidences of treatment-related adverse events. However, substantial variations exist in cancer type, drug and dosing schedule, and adverse event reporting criteria in the publication. Ignoring these variations and missing data patterns in adverse event reporting can lead to inaccurate estimation of the true incidences of treatment-related adverse events associated with PD-1 and PD-L1 inhibitors.
We performed a systematic review and meta-analysis of treatment-related adverse events of the Food and Drug Administration–approved PD-1 and PD-L1 inhibitors in published clinical trials. Using a novel bayesian approach to derive exact inferences based on patient-level data, we investigated the incidences of different treatment-related adverse events associated with these drugs, and we quantified the potential differences in adverse event incidences among a variety of cancer types, drugs, and dosing schedules.
A systematic search of the literature was conducted to identify published clinical trials of PD-1 and PD-L1 inhibitors that reported treatment-related adverse events. The search was done in PubMed, Web of Science, Embase, and Scopus using the terms nivolumab, pembrolizumab, atezolizumab, avelumab, durvalumab, PD-1 inhibitor, and PD-L1 inhibitor. The search was conducted from October 1, 2017, through final search for updates on December 15, 2018. The references of relevant published trials and review articles were also searched for additional eligible studies. Studies eligible for inclusion met all of the following criteria: (1) cancer therapy clinical trial, (2) participants were treated with a single-agent PD-1 or PD-L1 inhibitor, (3) reported tabulated data on treatment-related adverse events, and (4) published in English. Studies published online ahead of print were eligible, but meeting abstracts were excluded. When multiple publications reporting on the same study population were identified, the one with the most updated and/or comprehensive adverse event data was selected. The literature search, study selection, and data extraction were performed independently by 2 of us (F.Y. and X.W.), and discrepancies were reviewed by another investigator on the team (Y.W.) and resolved by consensus. This study followed the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) guideline.
The trial name, phase, cancer type, PD-1 and PD-L1 inhibitor used, dose escalation, dosing schedule, number of patients, number of all adverse events, and criteria for adverse event reporting in the publication were obtained from each included study. All-grade (severity) adverse event and grade 3 or higher (severity) adverse event data were both extracted.
The response variable is the number of reported all-grade or grade 3 or higher adverse events, assumed to follow a binomial distribution. To explain the between-study variation in the meta-analysis, we adjusted the incidence probability of adverse events by study-level moderators, including the therapeutic regimen and dosing schedule, cancer type, and adverse event.
We applied bayesian multilevel regression models for data analysis. Because many less frequently observed adverse events were not reported given a predetermined study-specific cutoff value, binomial distribution was proposed for fully reported adverse events and cumulative binomial probabilities were proposed for left-censored adverse events, in the likelihood of coherent parameter estimation. With a logit transformation (logit(z) = log(z)-log(1-z)) on the incidence probability, we assumed normal distributions for the additive effects of study-level moderators to adjust for study-specific effects. A noninformative prior distribution was proposed for the mean parameters of normal distributions, and weakly informative Cauchy prior distributions with mode at 0 and scale at 25 was proposed for the SD parameters.4,5 The same statistical model was separately applied to all-grade and grade 3 or higher adverse events.
For all bayesian analyses, we found the joint posterior distributions of the model parameters using a Markov chain Monte Carlo algorithm. Because closed forms of the full-conditional distributions are not available, we generated these distributions using Gibbs sampling and the Metropolis-Hastings algorithm. For data analysis, we used statistical software R, version 3.4.3 (with packages rjags_v4-6, coda_v0.19-1 and ggplot2_v2.2.1; R Foundation for Statistical Computing), and JAGS, version 4.3.0 (GNU General Public License). For both all-grade and grade 3 or higher adverse events, we plotted the incidences and their 95% probability intervals (bayesian credible intervals [CrIs]) by study, therapeutic regimen, cancer type, and adverse event using forest plots. Odds ratios (ORs), risk ratios (RRs) of grade 3 or higher adverse events to all-grade adverse events, and their CrIs were estimated from the medians and the 2.5 percentile and 97.5 percentile of the posterior distributions. For bayesian inferences, values of the posterior probabilities greater than .90 may be considered to show weakly significant positive association; greater than .95, significant positive association; and greater than .99, highly significant positive association of regimen with incidence. Values less than .10 correspond to weakly significant negative association; less than .05, significant negative association; or less than .01, highly significant negative association. Values near .50 correspond to no association.
Literature search and review of reference lists identified 5179 relevant publications. After screening and eligibility assessment, we included in the meta-analysis a total of 125 clinical trials involving 20 128 patients (Figure 1; eTable 1 in the Supplement).6-130 The PD-1 and PD-L1 inhibitors used included nivolumab (n = 46), pembrolizumab (n = 49), atezolizumab (n = 15), avelumab (n = 9), and durvalumab (n = 6). The trials involved the treatment of melanoma (n = 16), lung cancer (n = 26), gastrointestinal cancer (n = 14), genitourinary cancer (n = 22), hematologic malignant neoplasm (n = 8), other cancers (n = 31), and mixed cancer types (n = 10). One study had both melanoma and lung cancer arms,60 and another study included all cancer types but reported genitourinary cancer data separately.127
Collectively, the 125 studies reported more than 300 different types of adverse events. Overall, 12 277 (66.0%) of 18 610 patients from 106 studies developed at least 1 adverse event of any grade, and 2627 (14.0%) of 18 715 patients from 110 studies developed at least 1 grade 3 or higher adverse event.
For the meta-analysis, we focused on adverse events that either were reported by at least 10% of the studies or were likely immune-related adverse events (irAEs). Using these criteria, we narrowed down to 75 adverse events, which included most clinically relevant adverse events that are commonly seen in practice. A comprehensive list of the incidences of each adverse event is provided in eFigure 1 in the Supplement. The overall mean incidence of all-grade adverse events was 1.66% (95% CI, 1.47%-1.86%), and the mean incidence of grade 3 or higher adverse events was 0.11% (95% CI, 0.08%-0.14%). The mean incidences of all-grade and grade 3 or higher adverse events across different studies are shown in eFigure 2 in the Supplement.
As shown in Figure 2A, the most common all-grade adverse events were fatigue (18.26%; 95% CI, 16.49%-20.11%), pruritus (10.61%; 95% CI, 9.46%-11.83%), and diarrhea (9.47%; 95% CI, 8.43%-10.58%). The most common grade 3 or higher adverse events were fatigue (0.89%; 95% CI, 0.69%-1.14%), anemia (0.78%; 95% CI, 0.59%-1.02%), and aspartate aminotransferase (AST) increase (0.75%; 95% CI, 0.56%-0.99%) (Figure 2B).
PD-1 and PD-L1 inhibitors block the immune checkpoint pathway and reactivate cellular immunity and can cause autoimmune-mediated adverse events. These irAEs are of particular clinical interest and importance. Commonly recognized irAEs include various endocrine dysfunctions and other autoimmune-like disorders. We analyzed the incidences of adverse events that are likely immune-related.
Among the endocrine dysfunctions, the most frequent all-grade adverse events were hypothyroidism (6.07%; 95% CI, 5.35%-6.85%) and hyperthyroidism (2.82%; 95% CI, 2.40%-3.29%), followed by hyperglycemia (1.20%; 95% CI, 0.91%-1.55%), thyroiditis (0.75%; 95% CI, 0.52%-1.04%), and adrenal insufficiency (0.69%; 95% CI, 0.50%-0.93%) (Figure 2C). The most common grade 3 or higher adverse events were hyperglycemia (0.24%; 95% CI, 0.13%-0.38%), adrenal insufficiency (0.18%; 95% CI, 0.10%-0.30%), type 1 diabetes (0.18%; 95% CI, 0.10%-0.30%), hypophysitis (0.16%; 95% CI, 0.09%-0.27%), and hypothyroidism (0.08%; 95% CI, 0.04%-0.13%) (Figure 2D).
The most common other all-grade irAEs were diarrhea (9.47%; 95% CI, 8.43%-10.58%), AST increase (3.39%; 95% CI, 2.94%-3.89%), vitiligo (3.26%; 95% CI, 2.80%-3.79%), alanine aminotransferase (ALT) increase (3.14%; 95% CI, 2.71%-3.62%), pneumonitis (2.79%; 95% CI, 2.39%-3.23%), and colitis (1.24%; 95% CI, 0.99%-1.54%) (Figure 2C). For grade 3 or higher irAEs, AST increase (0.75%; 95% CI, 0.56%-0.99%) was most common, followed by ALT increase (0.70%; 95% CI, 0.52%-0.93%), pneumonitis (0.67%; 95% CI, 0.50%-0.89%), diarrhea (0.59%; 95% CI, 0.45%-0.77%), and colitis (0.47%; 95% CI, 0.34%-0.65%) (Figure 2D). Because diarrhea is likely a sign of colitis and ALT or AST increase a sign of hepatitis, incidences of autoimmune pneumonitis, colitis, and hepatitis are clinically significant.
Some adverse events were more likely to be severe if observed in a patient, as reflected by a high RR of grade 3 or higher adverse event incidence to the respective all-grade adverse event incidence. Notable among these adverse events were hepatitis (RR = 50.59%), lipase increase (RR = 42.01%), γ-glutamyltransferase increase (RR = 41.96%), type 1 diabetes (RR = 41.86%), and colitis (RR = 37.90%) (eTable 2 in the Supplement). The most common all-grade adverse events, such as fatigue, pruritus, diarrhea, rash, and nausea, had relatively lower RRs. The RRs of some irAEs of clinical interest were pneumonitis (RR = 24.01%), ALT increase (RR = 22.29%), AST increase (RR = 22.12%), hypothyroidism (RR = 1.32%), hyperthyroidism (RR = 1.42%), adrenal insufficiency (RR = 26.09%), hypophysitis (RR = 26.67%), and hypopituitarism (RR = 26.92%). These data suggest that, although hypothyroidism and hyperthyroidism tended to be mild, other irAEs, including pneumonitis, hepatitis, colitis, and other endocrine dysfunctions, were more likely to be severe.
Among the 125 studies, 112 (89.6%) reported whether any treatment-related deaths occurred. Among these, 40 studies reported at least 1 treatment-related death, with a total of 82 such deaths reported (eTable 3 in the Supplement). The overall incidence of treatment-related death was 0.45% (82 of 18 353).
As shown in the Table, the most common cause of treatment-related death (n = 82) was pneumonitis (23 [28.0%]). Other common causes were pneumonia (5 [6.1%]), sepsis (7 [8.5%]), respiratory failure (5 [6.1%]), and cardiovascular failure (3 [3.7%]). Respiratory causes (39 [48.0%]) accounted for almost half of the treatment-related deaths. Cardiovascular (8 [9.8%]), infectious (7 [8.5%]), hematologic (5 [6.1%]), and hepatic (3 [3.7%]) diseases were other common causes.
Based on the type of cancer treated in the clinical trials, we classified the 125 studies into 7 different categories: melanoma, lung cancer, gastrointestinal cancer, genitourinary cancer, hematologic malignant neoplasm, other cancers, and mixed cancer types. As shown in Figure 3A, the highest mean all-grade adverse events incidence was observed in melanoma (1.72%; 95% CI, 1.45%-2.27%), which was not much different from the lowest that was observed in lung cancer (1.55%; 95% CI, 1.23%-1.81%). Similarly, no statistically significant difference was found in the mean incidence of grade 3 or higher adverse events between any 2 categories (Figure 3B). These data suggest that the mean incidences of all-grade and grade 3 or higher adverse events were similar across different cancer types.
We compared the mean incidences of adverse events between different dosing schedules of the same drug as well as between different drugs. As shown in Figure 4A and B, no statistically significant differences were found in the mean incidences of all-grade or grade 3 or higher adverse events between different dosing schedules for nivolumab. The same was true for pembrolizumab and atezolizumab.
Nivolumab (3 mg/kg every 2 weeks [Q2W] dose) had higher mean incidences of all-grade adverse events (OR, 1.28; 95% CI, 0.97-1.79) and grade 3 or higher adverse events (OR, 1.30; 95% CI, 0.89-2.00) compared with pembrolizumab (10 mg/kg Q2W dose). Nivolumab also had a higher mean incidence of grade 3 or higher adverse events (OR, 1.81; 95% CI, 1.04-3.01) compared with PD-L1 inhibitors. The overall mean incidence of grade 3 or higher adverse events for PD-1 inhibitors was higher compared with PD-L1 inhibitors (OR, 1.58; 95% CI, 1.00-2.54) (Figure 4C).
The heterogeneity among studies was statistically quantified using a bayesian multilevel regression model and decomposed into several additive components attributed to various factors, including cancer type, adverse event, regimen and dosing schedule, and the residual heterogeneity owing to underlying clinical baseline variations between the patients enrolled in each study. Because of the nonlinearity of the logit transformation and a much higher mean incidence, all-grade adverse events actually had a larger variation in incidence compared with grade 3 or higher adverse events, even though in the models, grade 3 or higher adverse events showed larger SDs compared with all-grade adverse events for all study moderators. Therefore, we only compared the SDs across various factors within each model. Because between-dose variation by drug was very similar, the estimates were pooled together rather than estimated separately in the models.
eFigure 3 in the Supplement illustrates the posterior median and CIs of the SDs regarding either all-grade or grade 3 or higher adverse events. For both all-grade and grade 3 or higher adverse events, the largest variation came from the heterogeneity between different adverse event categories, for which the SD was statistically significantly larger than that of the residual effect (SD ratio for all-grade adverse events, 1.51; 95% CI, 1.37-1.66; for grade 3 or higher adverse events, 1.52; 95% CI, 1.26-1.83). Here, we consider the heterogeneity of the residual effect, denoting the information that cannot be explained by any of the study moderators, as a benchmark in comparison. The other factors, including drug, dose, and cancer type, all showed a small variation compared with the residual effect, with median SD ratios between 0.11 and 0.51.
We performed a systematic review of PD-1 and PD-L1 inhibitor–associated adverse events using a collection of sparse binomial data from published studies. Unlike meta-analyses using continuous summary statistics based on the large-sample theory, this meta-analysis used the number of each treatment-related adverse event to derive exact statistical inferences that were close or even identical to results from individual-level data (if available). Obtaining and merging original individual-level patient data are difficult, but this meta-analysis provides an alternative for estimating the study moderator effects without loss of relative efficiency.131,132 To our knowledge, this is the largest and most comprehensive meta-analysis of treatment-related adverse events for immune checkpoint inhibitors. Previous meta-analyses included fewer studies and primarily focused on certain adverse events, such as pneumonitis, endocrine dysfunction, or selected irAEs.133-136 A comprehensive analysis of all common treatment-related adverse events reported in clinical trials is critical, as the results constitute an important reference for clinicians. Such a global overview of immune checkpoint inhibitor adverse event incidences is complementary to American Society of Clinical Oncology and National Comprehensive Cancer Network guidelines on management of irAEs137 and informs clinical practice guidelines.
From the standpoint of patient counseling, several results from this meta-analysis are important. Approximately 2 in 3 patients treated with PD-1 or PD-L1 inhibitors in clinical trials had at least 1 adverse event, and 1 in 7 patients experienced at least 1 grade 3 or higher adverse event. These numbers can be important to share with patients before they begin treatment with a PD-1 or PD-L1 inhibitor. Fatigue was the most common all-grade adverse event (18.26%) and the most common grade 3 or higher adverse event (0.89%). Although less likely to be severe at presentation (about 5% chance), fatigue has a relatively high incidence (approximately 1 in 5) that is worth disclosing to patients. Pruritus, diarrhea, and rash are the next most common all-grade adverse events (approximately 1 in 10), but the likelihood of patients experiencing serious manifestations of these adverse events is low.
This meta-analysis showed that most of the most common grade 3 or higher adverse events were likely immune-related, including pneumonitis and dyspnea, diarrhea and colitis, ALT or AST increase and hepatitis, and lipase increase (suggestive of pancreatitis). Close monitoring and early recognition of pertinent symptoms and signs may help enable their proper management, such as prompt initiation of steroids. Dyspnea can be early signs of pneumonitis, diarrhea a sign of colitis, and ALT or AST increase a sign of hepatitis. If not detected early, these autoimmune-mediated disorders tend to present with higher severity and may even be fatal. Our results indicated that 24.01% of pneumonitis cases were grade 3 or higher in severity, and pneumonitis was the most common cause of treatment-related death in patients treated with PD-1 and PD-L1 inhibitors. In addition, hepatitis was the adverse event found most likely to be serious if it occurred, with 50.59% of hepatitis being grade 3 or higher. Diarrhea was the third most common all-grade adverse event, and clinical vigilance is necessary for early recognition and intervention to prevent severe colitis.
Among the irAEs manifesting as endocrine dysfunctions, hypothyroidism (6.07%) and hyperthyroidism (2.82%) were most common. Hyperglycemia, thyroiditis, adrenal insufficiency, hypophysitis, type 1 diabetes, and hypopituitarism were less common. However, these adverse events were more likely to be severe, with approximately 20% to 35% likelihood of being grade 3 or higher, as opposed to about 2% for hypothyroidism and hyperthyroidism. This difference may be partly attributed to frequent monitoring of thyroid-stimulating hormone in clinical trials, which allows for detection of thyroid dysfunction at an earlier stage. Hyperglycemia can be detected easily through routine laboratory work, but interpretation requires vigilance for possible pancreatic dysfunction. Routine monitoring of adrenal and pituitary function is not yet prevalent in clinical practice, likely owing to relatively low incidences of dysfunction. In this setting, a careful interview is important for early detection of pertinent symptoms.
The results of this meta-analysis indicated that the overall mean incidence of all-grade and grade 3 or higher adverse events did not differ between different cancer types. We did not further investigate whether specific adverse events were more common in particular cancer types (for example, pneumonitis in lung cancer or colitis in gastrointestinal cancer), which is a potential focus for future analyses. In addition, we found that nivolumab appeared to have a higher mean incidence of all-grade and grade 3 or higher adverse events, compared with pembrolizumab, but the mechanism and clinical significance are unclear. PD-L1 inhibitors appeared to be associated with lower mean incidence of grade 3 or higher adverse events, compared with nivolumab and pembrolizumab, possibly owing to the presence of the other PD-1 ligand, PD-L2, which may maintain some level of checkpoint signaling. No head-to-head comparison trials have been conducted, and interpretation of these results should be made with caution.
A major strength of this meta-analysis is the coherent estimation of adverse event incidences with accommodation of both fully reported and censored data. The missing data problem is pivotal in meta-analysis, because published studies do not always provide all of the necessary information, which is especially true for treatment-related adverse events as the primary outcome. Usually, only the prevalent adverse events were reported for each study, and most information regarding less frequently observed adverse events was censored using a predetermined study-specific cutoff value. Furthermore, the larger the scope of the study, the higher the cutoff value (usually a percentage of the total sample size), which brought more uncertainty regarding the censored information. This type of missing information, if treated as missing completely at random, will result in overestimation of the incidence probability of the corresponding adverse events. Therefore, we took an innovative approach by introducing additional cumulative binomial probabilities in the likelihood function to accommodate the censored data. The between-study heterogeneity was simultaneously quantified by study-level moderators using bayesian multilevel modeling, with exact inference avoiding continuity correction for sparse binomial data.138
This meta-analysis has limitations. Small-study effects were observed when studies with smaller sample sizes had different incidences and wider CIs for both all-grade and grade 3 or higher adverse events. Although the forest plots showed no asymmetry favoring low adverse events incidence studies, this meta-analysis is subject to publication bias given that all of our analyses were based on publications. In addition, this study is subject to any biases or errors of the original investigators, and the results are generalizable only to patient groups eligible for these trials.
This meta-analysis, which used an innovative bayesian multilevel regression model, has defined the incidences of all common treatment-related adverse events of PD-1 and PD-L1 inhibitors. The incidences of adverse events are independent of cancer types, but different PD-1 and PD-L1 inhibitors may be associated with different incidences of adverse events. This global overview of the adverse events of PD-1 and PD-L1 inhibitors can be used as a reference by clinicians and may guide clinical practice.
Accepted for Publication: February 6, 2019.
Corresponding Author: Michael L. Wang, MD, Department of Lymphoma/Myeloma, The University of Texas MD Anderson Cancer Center, 1515 Holcombe Blvd, Houston, TX 77030 (email@example.com).
Published Online: April 25, 2019. doi:10.1001/jamaoncol.2019.0393
Author Contributions: Drs Y. Wang, Zhou, and Yang contributed equally to this work. Drs Y. Wang and M. L. Wang had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.
Concept and design: Y. Wang, Zhou, Guan, Duma, Price, Molina, Halfdanarson, Grothey, Nowakowski.
Acquisition, analysis, or interpretation of data: Y. Wang, Zhou, Yang, Qi, X. Wang, Shen, Duma, Vera Aguilera, Chintakuntlawar, Molina, Pagliaro, Grothey, Nowakowski.
Drafting of the manuscript: Y. Wang, Zhou, Yang, Qi, X. Wang, Shen, Chintakuntlawar, Molina, Nowakowski.
Critical revision of the manuscript for important intellectual content: Y. Wang, Qi, Guan, Duma, Vera Aguilera, Chintakuntlawar, Price, Molina, Pagliaro, Halfdanarson, Grothey.
Statistical analysis: Zhou, Yang, Qi, X. Wang, Shen.
Administrative, technical, or material support: Nowakowski.
Supervision: Zhou, Vera Aguilera, Chintakuntlawar, Molina, Halfdanarson, Grothey.
Conflict of Interest Disclosures: Dr Halfdanarson reported grants from Ipsen, Agios, ArQule, and Thermo Fisher Scientific as well as other from Advanced Accelerator Applications outside the submitted work. Dr Grothey reported grants and nonfinancial support from Bayer, Genentech, and ARRAY as well as grants from Boston Biomedical and Daiichi outside the submitted work. Dr Nowakowski reported grants from Celgene, Morphosys, Bayer, Genetech, Curis, and Nanostring outside the submitted work. Dr Ansell reported other from Bristol-Myers Squibb and Merck outside the submitted work. Dr M. L. Wang reported grants, personal fees, and nonfinancial support from Janssen, AstraZeneca, AcertaPharma, and Celgene; grants and nonfinancial support from Pharmacyclics; grants from BeiGene, Kite Pharma, Juno Therapeutics, and Oncternal; nonfinancial support from MoreHealth and Pulse Biosciences; and personal fees and nonfinancial support from PeerView Institute for Medical Education and OMI outside the submitted work. No other disclosures were reported.