PubMed, Web of Science, Embase, and Cochrane databases were searched to identify relevant studies published from the database inception to June 9, 2020, with language confined to English. The key retrieval terms in the search strategy included cancer, tumor, neoplasm, immune checkpoint inhibitors (anti-PD-1, anti-PD-L1, anti-CTLA-4), specific ICI names (nivolumab, pembrolizumab, atezolizumab, durvalumab, avelumab, ipilimumab, cemiplimab), and some terms relevant to “rechallenge” (retreat, readministrate, restart, reinitiate, resume). The detailed search strategy is provided in Supplementary Material 2: Table S1 . References of selected papers were also searched to identify additional studies.

Inclusion and exclusion criteria of studies were established before the literature search. Studies have to fulfill the following criteria for eligibility: enrolled adult patients (aged over 18) and enrolled cancer patients who rechallenged ICIs after the initial irAEs. Studies not adhering to the inclusion criteria were excluded. Other exclusion criteria were as follows: patients concurrently treated with ICIs and other treatments (e.g., radical resection, radiation therapy, chemotherapy, or targeted therapy); no detailed information of irAEs or treatment outcomes of ICIs; non-clinical studies, review, systematic review, or conference abstract without exhaustive data; non-English articles; and no full-text original articles. Two researchers (QZ and LX) independently screened titles and abstracts of every search output to identify all studies that potentially met the inclusion criteria. Then, the full texts of all potentially eligible studies were read for further discrimination. The two researchers (QZ and LX) solved any discrepancies on study selection via discussion, and a third researcher (JZ) was consulted when necessary.

All data were collected by two researchers (QZ and LX) independently in accordance with a predefined procedure. The following detailed characteristics of the study (cohort study, case series, and case report) were extracted: author, publication year, study design, cancer type, types of initial and rechallenge ICIs, rechallenge ratios, time interval between initial irAEs and ICI rechallenge, types and incidence of initial and rechallenged irAEs, ORR, and disease control rate (DCR) after rechallenge. Rechallenged irAEs included flared and novel irAEs after ICI rechallenge. The same two independent researchers (QZ and LX) assessed the methodological quality of all included studies using the Newcastle-Ottawa Scale (NOS) criteria (18), weighted as selection, comparability, and outcome. The scale ranges from 0 (poor methodological quality) to 9 (optimal methodological quality) points. Any discrepancies were solved via discussion or consultation with the third researcher (JZ).

Safety assessment included incidence of all-grade rechallenged irAEs and incidence of high-grade rechallenged irAEs. The severity of irAEs was recorded as grade 1 to 5 based on version 5 of the Common Terminology Criteria for Adverse Events (CTCAE) of the National Cancer Institute (Bethesda, MD, USA). Grade ≥3 was considered as high-grade irAEs, while grade 1 or 2 was low-grade irAEs. Efficacy assessment included ORR and DCR after ICI rechallenge. ORR was defined as the rate of patients who had a complete response or partial response, while DCR was defined as the rate of patients who had a complete response, partial response, or stable disease.

We employed Review Manager 5.3 (Cochrane Community, London, UK) and SPSS 18.0 (SPSS Inc., Chicago, IL, USA) for statistical analyses and plotting. Synthesis of all-grade and high-grade rechallenged irAEs, ORR, and DCR was conducted via meta-analysis using pooled odds ratios (OR), with 95% confidence intervals (CIs) calculated via the Mantel–Haenszel model (19). The pooled incidence of all-grade and high-grade rechallenged irAEs and other available dividing factors was calculated via a meta-analysis of proportions. Since included studies in the meta-analysis were all retrospective studies, random-effects model with the Mantel–Haenszel model (19) was applied considering the significant heterogeneity, which was then proved by the I-squared (I ²) test. Heterogeneity was indicated as low (I ² = 0% to 40%), moderate (I ² = 40% to 70%), and substantial (I ² = 70% to 100%). Predefined subgroup analysis was mainly conducted for accessible data including types of initial irAEs and cancer types. Moreover, we pooled individual-level cases for clinical factors of patients including age, gender, types and grade of initial irAEs, corticosteroid dosage, cancer type, types of initial and rechallenged ICIs, and time interval between initial irAEs and ICI rechallenge, additionally complementing analysis on the safety and efficacy of ICI rechallenge. Univariate and multivariate OR with 95% CIs were computed using a logistic regression model. Candidate factors with p-values <0.1 in the univariate analysis were included in the multivariate model. p-values were computed using an unpaired two-tailed Wald test. For sensitivity analysis, one study was sequentially omitted to judge the stability of the pooled results.

Our literature search found 1,921 articles, and 21 additional articles were retrieved by searching the references of included studies. We ultimately reviewed the full texts of 236 articles after removing duplicates and screening titles and abstracts; of these, 77 studies comprising 788 individuals were enrolled for the present study (see Figure 1 ). The reasons for excluding the other 159 articles are listed in Figure 1 .

Eighteen cohort studies comprising 691 patients were enrolled in the meta-analysis of safety and efficacy (see Table 1 and Supplementary Material 2: Table S2 ). The median (range) number of patients enrolled was 31 (10–167) for safety and 19 (8–80) for efficacy. The included cancers were multiple cancer (seven studies, n = 372), melanoma (five studies, n = 189), non-small cell lung cancer (NSCLC) (five studies, n = 105), and colorectal cancer (one study, n = 25). Among the initial ICI treatment, anti-PD-1/PD-L1 monotherapy (191/416, 45.9%) was the most reported ICI therapy, followed by anti-CTLA-4 monotherapy (55/416, 13.2%) and combination therapy (170/416, 40.9%). Ten studies focused on global irAEs, and eight studies focused on specific irAEs, consisting of two for colitis/diarrhea, two for pneumonitis, and one each for neurological toxicity, pancreatic toxicity, hepatotoxicity, and acute kidney injury. Five hundred thirty-three patients reported initial irAE grades: low-grade (274/533, 51.4%) and high-grade (259/533, 48.6%). Anti-PD-1/PD-L1 monotherapy (449/691, 65.0%) was rechallenged more frequently than anti-CTLA-4 monotherapy (41/691, 5.9%) or combination therapy (27/691, 3.9%).

Five case series and 54 case reports comprising 97 patients were enrolled for the analysis of factors associated with the safety and efficacy of ICI rechallenge (see Supplementary Material 2: Table S3 ). Melanoma, lung cancer, renal cancer, hematologic cancer, and other cancer types were reported in 41 (42.3%), 38 (39.2%), 2 (2.0%), 8 (8.2%), and 8 (8.2%) patients, respectively. Sixty-eight (70.1%) patients were initially treated with anti-PD-1/PD-L1 antibodies, 8 (8.2%) with anti-CTLA-4 antibodies, and 21 (21.6%) with combination. Anti-PD-1/PD-L1 monotherapy (83/97, 85.6%) was rechallenged more frequently than anti-CTLA-4 monotherapy (7/97, 7.2%) or combination therapy (7/97, 7.2%). Detailed demographic information was reported in 74 patients. In total, the median (range) age was 62 (30–87) years, and 47 (63.5%) were male. The initial irAEs were mainly respiratory [15 (20.3%)], gastrointestinal [12 (16.2%)], and hematologic [10 (13.5%)] irAEs. There were 27 (36.5%) low-grade, 36 (48.6%) high-grade, and 11 (14.9%) unknown grade irAEs. Nine (12.2%) patients were treated with low-/moderate-dose steroids, 49 (66.2%) with high-dose steroids, 7 (9.5%) with unknown-dose steroids, and 9 (12.2%) with unknown or no treatment.

Fifteen cohort studies were included in the analysis of safety (7, 11, 14, 20–31). The recurrence rate of all-grade and high-grade irAEs were 34.2% and 11.7%, separately. ICI rechallenge was associated with a significantly higher incidence of all-grade irAEs than initial ICIs (OR, 3.81; 95% CI, 2.15–6.74; p < 0.0001; I ² = 58.6%) (see Figure 2 ); however, no significant difference was noted for high-grade irAEs (OR, 1.63; 95% CI, 0.86–3.11; p = 0.136; I ² = 19.4%) (see Figure 2 ).

The results of subgroup analyses for types of initial irAEs and cancer types are displayed in Figure 3 . Initial pneumonitis was associated with a higher all-grade recurrence (OR, 4.09; 95% CI, 1.76–9.51; p = 0.001; I ² = 0.0%), but not with a higher high-grade recurrence (OR, 1.26; 95% CI, 0.17–9.48; p = 0.826; I ² = 0.0%). Initial global irAEs were also associated with a higher all-grade recurrence (OR, 3.62; 95% CI, 1.01–12.90; p = 0.047; I ² = 85.3%); however, no significant difference was noted for high-grade recurrence (OR, 1.53; 95% CI, 0.50–4.75; p = 0.458; I ² = 62.7%). Patients with NSCLC had a higher incidence of all-grade rechallenged irAEs (OR, 5.72; 95% CI, 3.46–9.45; p < 0.0001; I ² = 0.0%), but no significant difference existed in high-grade irAEs (OR, 1.50; 95% CI, 0.34–6.64; p = 0.591; I ² = 50.9%). Cohorts enrolled with multiple cancers also showed a significantly higher incidence of all-grade irAEs in ICI rechallenge (OR, 4.48; 95% CI, 1.95–10.31; p < 0.0001; I ² = 0.0%).

Further analysis revealed that the recurrence rate of all-grade irAEs was not significantly different in different rechallenged ICIs (χ ² = 0.800, p = 0.670, df = 2). Specifically, for patients initially treated with anti-PD-1/PD-L1 antibodies, anti-CTLA-4 antibodies rechallenge had a significantly higher incidence of all-grade irAEs than anti-PD-1/PD-L1 antibodies rechallenge (p = 0.040). However, for patients initially treated with anti-CTLA-4 antibodies or combination, no significant difference existed in the incidence of all-grade irAEs in different rechallenged ICIs (initial anti-CTLA-4 antibodies: χ ² = 0.248, p = 0.618, df = 1; initial combination: χ ² = 0.391, p = 0.532, df = 1) (see Supplementary Material 2: Tables S4, S5 ).

Eight cohort studies were included in the analysis of efficacy (9–11, 14, 15, 21, 25, 29). The pooled ORR and DCR of ICI rechallenge were 43.1% and 73.6%, respectively. No significant difference was noted between initial ICI treatment and ICI rechallenge for ORR (OR, 0.45; 95% CI, 0.15–1.35; p = 0.155; I ² = 66.6%) (see Figure 4 ). Similarly, no significant difference was noted between initial ICI treatment and ICI rechallenge for DCR (OR, 0.70; 95% CI, 0.24–2.06; p = 0.521; I ² = 41.9%) (see Figure 4 ). Further pooled analysis revealed that compared with initial ICI treatment, ICI rechallenge in patients with NSCLC has no significant difference for ORR (OR 0.36; 95% CI, 0.11–1.20; p = 0.097; I ² = 71.6%) and DCR (OR, 0.65; 95% CI, 0.16–2.65, p = 0.543; I ² = 60.3%) (see Figure 5 ).

Table 2 shows the factors associated with the occurrence of all-grade and high-grade rechallenged irAEs. For high-grade rechallenged irAEs, univariate analysis showed that gastrointestinal irAEs (OR, 6.81; 95% CI, 1.58–29.26; p = 0.010) and time interval between initial irAEs and ICI rechallenge (OR, 1.02; 95% CI, 1.00–1.05; p = 0.031) were associated with a higher recurrence, whereas initial anti-PD-1/PD-L1 antibodies were associated with a lower recurrence (OR, 0.26; 95% CI, 0.07–0.99; p = 0.049). For all-grade rechallenged irAEs, anti-PD-1/PD-L1 antibodies rechallenge was associated with a lower recurrence (OR, 0.28; 95% CI, 0.08–0.94; p = 0.039). Factors selected from the univariate analysis (p < 0.1) for multivariate analysis showed no significance for both high-grade and all-grade rechallenged irAEs (p > 0.05).

Univariate analysis for ORR and DCR after ICI rechallenge showed that no clinical factors of patients were found significant (p < 0.05) and multivariate analysis was not applicable (see Supplementary Material 2: Table S6 ).

In the sensitivity analysis, the pooled results for all-grade irAEs, high-grade irAEs, ORR, and DCR remained stable, regardless of which study was deleted, which indicates the robust association (see Supplementary Material 2: Figure S1 ).