This systematic review adhered to the guideline of systematic review of prognostic factor studies by Riley et al and reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses statement. A detailed protocol has been registered prospectively at PROSPERO (CRD42020202142).
Search strategy and selection criteria
We conducted a comprehensive search on PubMed, Scopus, MEDLINE (via EBSCO), Cumulative Index to Nursing and Allied Health Literature (CINAHL), Cochrane Controlled Register of Trials (CENTRAL), and the World Health Organization (WHO) COVID-19 research databases, searching for relevant studies published from inception up to 14 September 2020 with keywords listed in Supplementary Table S1. Furthermore, we also searched grey literature (i.e. Google Scholar, ProQuest, MedRxiv, BioRxiv, and Social Science Research Network) databases, in addition to manually searching reference lists from the included studies and similar reviews. Lastly, we retrieved similar records of the included studies with the PubMed’s ‘similar articles’ algorithm and subsequently deduplicated and screened them against the pre-specified eligibility criteria. No language restrictions were applied during the search.
Literature searches were performed by two independent investigators, with any discrepancies resolved by the blind assessment of a third investigator. The retrieved records were screened against the following inclusion criteria: (1) study design, primary studies including case series or letter to editors with at least 10 patients; (2) population and exposure, studies enrolling COVID-19-infected cancer patients with and without prior exposure to ICIs; and (3) outcomes, including mortality, severity, and any other prognosis-related outcomes. In the case of studies only mentioning immunotherapy as an exposure to COVID-19 patients, the corresponding authors were contacted to confirm their study settings, and the studies were subsequently excluded when no response was provided (see Additional methods in Supplementary Material for more details). Contrarywise, records were excluded if the full-text articles were non-English or irretrievable.
Data extraction and risk of bias assessment
The following information were extracted from each included studies: (1) author and year of publication; (2) study characteristics, including recruitment period, study design, settings, location and sample size; (3) patient characteristics, including mean/median age, proportion of male patients, comorbidities, cancer types, adjuvant therapies, and characteristics of ICI (i.e. last exposure to treatment, drug class); and (4) outcomes. The primary outcomes in this review were the risk of poor prognosis (i.e. mortality and severity) among COVID-19-infected cancer patients. Whenever possible, outcomes were further investigated per criterion according to the WHO interim guidance, viz. rate of hospitalization, intensive care unit (ICU) admission, invasive ventilation, acute respiratory distress syndrome (ARDS), and shock. Data extraction were performed by one review author (GL) using a pre-specified form in MS Excel® for Office 365 MSO ver. 2002 (Microsoft Corporation, Redmond, WA, 2018). A second investigator (RAB) checked the accuracy of these data extractions, and any disagreements resolved by the consensus between the authors.
Any reported effect size types (i.e. hazard ratio [HR], odds ratio [OR], relative risk [RR]) were incorporated in this study. When only binary data were provided, unadjusted odds ratio were calculated from the frequency of events and sample sizes. In the case of studies reporting multiple adjustment models, we extracted the adjusted model incorporating the greatest number of covariates.
The included studies were further assessed for risk of bias by using the Quality in Prognosis Studies (QUIPS) tool, where the overall bias risk was judged to be low, moderate, and high risk. Risk of bias assessments were conducted by two independent reviewers, and any discrepancies were resolved by a third adjudicator in a blinded fashion. Details on the QUIPS checklist can be found on Supplementary Table S2.
Data analysis and synthesis
Data analyses were performed by using the Review Manager 5.3 (The Nordic Cochrane Centre, The Cochrane Collaboration, 2014, Copenhagen) and MetaXL software version 5.3. (EpiGear International, Queensland, Australia). In the case of studies involving overlapping populations, analyses were prioritized to the largest-sized study. Outcomes were pooled as ORs, RRs, or HRs separately along with their 95% confidence intervals (CIs) by using the generic inverse variance methods. Both unadjusted and adjusted outcomes were extracted and synthesized in this study; however, adjusted estimates were prioritized for reporting and interpretation whenever available. Due to the likeliness of unexplained heterogeneity, a DerSimonian-Laird random-effects model was used. Heterogeneity between studies was investigated with Cochran’s Q test and I2 statistics. According to I2 values, heterogeneity was classified as negligible (0-25%), low (25-50%), moderate (50-75%), or high (>75%), while the significance level for Q statistics were set at 10%.
Subgroup analyses were performed only for outcomes with at least two studies in each subset. A priori, we defined subgroups according to study design, location, sample size, and risk of bias; while additional subsets based on the presence of adjuvant therapy (i.e. ICI monotherapy vs ICI plus chemotherapy), cancer type (i.e. lung vs non-lung cancer), and comparator group (i.e. no treatment, chemotherapy, targeted therapy, radiotherapy, and surgery) were determined posteriori. On the other hand, sensitivity analysis was conducted by leave-one-out analysis and simultaneous exclusion of studies with high-bias risk. A P-value of <0.05 indicated statistical significance. When the number of studies was adequate (n10), potential publication bias was investigated visually by contour-enhanced funnel plot and quantitatively with Egger’s and Begg’s test.
Lastly, the overall quality of evidence was assessed with the modified version of Grading of Recommendations Assessment, Development and Evaluation (GRADE) framework for prognostic reviews, where the certainty of evidence was rated as high, moderate, low, or very low according to the judgements of these following domains: phase of investigation, study limitation, inconsistency, indirectness, imprecision, publication bias, moderate/large effect sizes, and exposure-response gradients.