2.1 Search methods and study selection

This systematic review and network meta-analysis were conducted following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement (Supporting Information: Table S1) [11]. The protocol was registered in the Prospective Register of Systematic Reviews (PROSPERO CRD42022339062). We searched the Cochrane, Embase, and PubMed databases for relevant studies by using the following search terms: (“PD-1” OR “PD-L1” OR “CTLA-4”) AND (“phase III” OR “phase 3”) AND “adjuvant therapy”. Papers published as of May 20, 2022 were included. Relevant papers from the references of identified studies were also examined. The full search criteria are presented in Supporting Information: Table S5.

Two independent reviewers (R. Xie and J. Wu) performed the study screening. Disagreements were resolved by a third reviewer (J. Shou). The inclusion criteria for trial selection were as follows: (1) phase III randomized clinical trials (RCTs) in which participants received adjuvant PD-1, PD-L1, or CTLA4 inhibitor monotherapy; (2) the patients then received tumor resection or radiotherapy; (3) tabulated data on TRAEs and deaths were available; and (4) published in English. Studies that did not satisfy the inclusion criteria were excluded. Other exclusion criteria were as follows: (1) data from conferences before 2022; (2) articles in which participants were treated with sequential monotherapies; and (3) trials in which treatments other than ICIs were involved. The full literature was retrieved for further qualitative review.

2.2 Data extraction

General characteristics including start year, study ID, patient age, sample size, intervention arm, patient sex, and the control arm were extracted. Data for reported AEs (AEs of any cause, TRAEs, and immune-related adverse events [IRAEs]) and deaths (death due to AEs and deaths due to TRAEs) were extracted. AEs in all-grade and grade ≥3 were defined as grades 1–5 and 3–5, respectively. Any noted unintentional or unfavorable clinical signs or symptoms, including complications of miscarriage in all-causes were denoted as AEs [12]. TRAEs are any AEs that in the investigators’ opinion may have been caused by the study medication with reasonable possibility. IRAEs are autoimmune conditions caused by unintended effects of the ICI-mediated activation of the immune system and may occur in any organ system [13]. The safety profiles of the most commonly reported all-grade AEs were also analyzed for categorization on the basis of organ systems: renal, hepatic, pulmonary, gastrointestinal, skin, endocrine, cardiac, and blood/lymphatic. All AEs were classified in accordance with the Common Terminology Criteria for Adverse Events (CTCAE) version 5.0 [14]. Discontinuation events due to TRAE were also extracted. The risk of bias in each trial was assessed using the Cochrane Risk of Bias Tool and was annotated as high, unclear, or low risk of bias [15]. The following categories were scored: allocation concealment, random sequence generation, blinding of participants and personnel, incomplete outcome data, blinding of outcome assessment, selective reporting, and other biases (Supporting Information: Figure S3). Any disagreements in the study selection, data extraction, and quality assessment were resolved by discussion to achieve a consensus.

2.3 Data synthesis and statistical analysis

Odds ratios (ORs) with 95% confidence intervals (95% CIs) are used to describe AE rates. If significant heterogeneity existed, network meta-analyses of AEs were performed with the Bayesian random-effects consistency model, which is a conservative approach to dealing with between-study heterogeneity; otherwise, we used the fixed effects model [16]. Bayesian network modeling confers the advantage of accommodating complex situations by offering a straightforward method for probabilistic statements and predictions on the treatment effects [17]. We then used inconsistency model analysis to show the inconsistency of evidence [18]. The ranking probability of the treatments for AEs was calculated by the surface under the cumulative ranking curve (SUCRA) analysis [17]. Heterogeneity among trials was assessed by the I² values of the consistency model if more than one trial existed. Ι² values greater than 25%, 50%, and 75% indicated low, moderate, and high heterogeneity, respectively [19]. The summary of incidence was calculated by the ratio of events that occurred and total patients. Network meta-analyses with low incidence and poor convergence were removed and reported as incidence with 95% CIs. The 95% CIs were estimated together with the incidence through binomial probability. We assessed the study inclusion reliability by calculating the estimated sample size (Supporting Information: Table S4) [20]. All of the enrolled studies had large enough sample sizes except for the comparison of nivolumab and ipilimumab (2259 enrolled, 3529 required). However, all published phase III trials to date with valid data for ICI adjuvant therapy were enrolled in this network meta-analysis (9243 patients in total).

To illustrate the sample size and number of trials, we used the “rjags” and “GeMtc” packages of R 4.0.2 (https://www.r-project.org/) to generate the Bayesian network modeling of AEs [21]. Incidence rates with 95% CIs were calculated with the binconf() function from the “Hmisc” package. Analyses of heterogeneity and ranking probability were also run in R. To further determine the heterogeneity effects, the number of adaptations was set to 5000, and the sample iteration parameter was adjusted to 20,000.

3.1 Characteristics of the enrolled trials and systematic review

In total, 9267 records from the database were identified and screened for titles and abstracts (Figure 1). Ten phase III RCTs were included in the analysis [3, 4, 8-10, 22-26]. To ensure the reliability of reported results and quality control, phase II clinical trials were not included in the study. The ICIs are identified as PD-1 inhibitors (pembrolizumab and nivolumab), PD-L1 inhibitor (atezolizumab), and CTLA-4 inhibitor (ipilimumab). Among the 10 studies encompassing 9243 patients, 7 reported deaths due to AEs, 10 reported deaths due to TRAEs, 10 reported AEs of any cause, 8 reported TRAEs, 6 reported IRAEs, and 6 reported discontinuations due to TRAEs. The general characteristics of these studies, including the group interventions and patient populations for safety assessment, are summarized in Table 1. The efficacy of ICIs (DFS and OS) is summarized in Supporting Information: Table S6. All of the enrolled studies showed a survival benefit of ICIs except for the IMvigor010 study.

3.2 Overall incidence of deaths caused by ICI adjuvant therapy

To fully describe the landscape of death events, we separately calculated the incidences of deaths due to AEs and TRAEs (Figure 2a). Among the 33 death events reported in the seven trials, the death rate of AEs of any cause was 0.96% (95% CI: 0.66–1.34). Patients who received adjuvant ipilimumab (1.49%, 95% CI: 0.60–3.04) and atezolizumab (1.24%, 95% CI: 0.62–2.21) had overall mortality rates higher than 1%. However, three enrolled studies (KEYNOTE-054, CheckMate 274, and CheckMate 238) did not report all-cause death events. We further investigated the death events due to TRAEs, of which 21 death events were noted, with an overall incidence of 0.40% (95% CI: 0.26–0.61). In contrast to other ICIs, pembrolizumab caused the fewest treatment-related deaths (0.24%, 95% CI: 0.10–0.56), whereas treatment-related death rates of both ipilimumab (0.76%, 95% CI: 0.31–1.55) and atezolizumab (0.56%, 95% CI: 0.18–1.31) were higher than the average. The causes of death for patients treated with ICI adjuvant therapy were complex. All deaths due to TRAEs are summarized in Table 2. The KEYNOTE-716 and KEYNOTE-564 studies conferred no treatment-related deaths. In total, 17 deaths with detailed causes were extracted and noted, among which colitis was prominent in ipilimumab, with four cases of death. Despite the incomplete description of relevant causes of death in the KEYNOTE-091 study, the current profile suggests that pembrolizumab caused only one death event from myositis in the KEYNOTE-054 study. Next, we categorized the 17 death events with reported causes on the basis of organ systems. The most frequent causes of death by organ system were gastrointestinal (0.10%, 95% CI: 0.04–0.24), followed by pulmonary (0.08%, 95% CI: 0.03–0.21), cardiac (0.06%, 95% CI: 0.02–0.18), and blood/lymphatic and multiple organ failure (0.04%, 95% CI: 0.01–0.17; Figure 2b). There was one treatment-related death from myositis that was classified as “other.”

3.3 Network meta-analysis with the consistency model

Figure 3a shows the network plots for AEs of different causes from the 10 studies of the four ICI treatments. The arms of placebo, best supportive care, and observation were stratified into a control arm due to not receiving any treatment interventions. As shown in Figure 3b, the ORs of ipilimumab versus the control arm were 6.36 (95% CI: 3.00–14.60) and 3.62 (95% CI: 2.64–5.09) for AEs of any cause in all-grades and of grade ≥3, respectively. Notably, pembrolizumab had similar ORs when compared with the control arm in these two comparisons (OR: 1.85, 95% CI: 1.34–2.83 and OR: 1.83, 95% CI 1.49–2.27). Regarding TRAEs, we found that ipilimumab had high ORs in any grade (OR: 11.46, 95% CI: 2.96–43.23) and grade ≥3 (OR: 14.11, 95% CI: 0.70–255.63), whereas nivolumab caused much fewer TRAEs in both categories (any grade OR: 2.83, 95% CI: 1.37–6.05 and grade ≥3 OR: 2.73, 95% CI: 0.49–15.85). In terms of grade ≥3 IRAEs, the OR of atezolizumab versus the control arm (OR: 12.43, 95% CI: 2.42–78.48) was significantly lower than those of the other ICIs. The ranking probability for each ICI is presented in Supporting Information: Figure S1. The consistency analysis of discontinuations due to AEs indicated that ipilimumab had the most discontinuation events (OR: 20.52, 95% CI: 6.68–58.18; Supporting Information: Figure S2a). Supporting Information: Figure S2b shows the incidences of discontinuation for all ICIs, in which nivolumab had the fewest discontinuation events (13.26%, 95% CI: 11.55–15.19).

By examining the complete data of the 10 enrolled studies, we found that organ system AEs were reported either as IRAEs or TRAEs in each study. As shown in Figure 4, we analyzed IRAEs or TRAEs of all-grades on the basis of organ systems. For renal- and urinary-related AEs, only nivolumab and ipilimumab were included, as there were only valid data for these, and they showed ORs versus the control arm of 2.10 (95% CI: 0.83–5.47) and 2.45 (95% CI: 0.49–14.42), respectively. Nivolumab had the lowest OR for hepatic AEs (OR: 1.78, 95% CI: 0.59–5.07), while the OR of pembrolizumab was much higher (OR: 12.12, 95% CI: 1.09–461.73). Regarding pulmonary events, the comparisons indicated nivolumab had the fewest AEs, with an OR of 3.64 (95% CI: 0.68–19.71). Gastrointestinal system AEs from pembrolizumab were notable (OR: 5.60, 95% CI: 1.36–30.39), whereas nivolumab is likely to be the safest agent for gastrointestinal AEs (OR: 1.40, 95% CI: 0.61–3.20). For AEs of the skin, ipilimumab caused more events than the other ICIs (OR: 6.23, 95% CI: 2.36–15.63). Unlike the results for other systems, the analysis of endocrine AEs suggested that nivolumab had the highest OR value compared to the control arm (OR: 7.33, 95% CI: 2.37–25.79).

IRAEs of special interest that are frequently reported for ICI therapies were also analyzed (Figure 5). The consistency model analyses with convergence values near 1.00 are denoted as ORs (Figure 5a), while other IRAEs are presented as incidence rates with 95% CIs (Figure 5b). For immune-related colitis, ipilimumab was observed to have the highest OR versus the control arm (OR: 12.93, 95% CI: 3.55–52.38), while nivolumab led to the least risk of colitis (OR: 2.30, 95% CI: 0.59–10.53). In terms of thyroid disorders, pembrolizumab and atezolizumab caused the most hyperthyroidism events (OR: 21.59, 95% CI: 3.18–224.47) and hypothyroidism (OR: 70.66, 95% CI: 19.04–533.88), respectively. Compared with nivolumab and pembrolizumab, atezolizumab had a higher OR for pneumonitis (OR: 13.70, 95% CI: 1.90–361.64). Next, we analyzed incidences of adrenal insufficiency, among which atezolizumab had the least IRAEs (0.79%, 95% CI: 0.32–1.62). For diabetes mellitus, we found that pembrolizumab correlated with the highest incidence rate (1.49%, 95% CI: 0.93–2.24). Notably, there were two agents with incidence rates >10% for specific AEs, that is, hepatitis for atezolizumab (14.80%, 95% CI: 12.53–17.32) and hypophysitis for ipilimumab (13.53%, 95% CI: 11.38–15.90).

3.4 Assessments of inconsistency and heterogeneity

The variance of the consistency and inconsistency model are presented in Supporting Information: Table S2. The heterogeneity of general AEs was estimated and shown in Supporting Information: Table S3. The Ι² value suggested moderate heterogeneity in the analysis of TRAEs (Ι² = 68.2 for TRAEs of any grade, Ι² = 73.1 for grade ≥ 3 TRAEs), whereas we found low heterogeneity in all-cause AEs and IRAEs. Supporting Information: Figure S3 shows the quality assessment, as scored by the Cochrane risk of bias tool. The KEYNOTE-091 study was reported as a conference abstract, thus the risk of bias in each aspect was beyond moderate. Other studies were assessed to have an acceptable low risk of bias.