Are patients with cancer at higher risk of COVID-19-related death? A systematic review and critical appraisal of the early evidence


 Background
 Early reports suggested that COVID-19 patients with cancer were at higher risk of COVID-19-related death. We conducted a systematic review with risk of bias assessment and synthesis of the early evidence on the risk of COVID-19-related death for COVID-19 patients with and without cancer.
 
 Methods and Findings
 We searched Medline/Embase/BioRxiv/MedRxiv/SSRN databases to 1 July 2020. We included cohort or case-control studies published in English that reported on the risk of dying after developing COVID-19 for people with a pre-existing diagnosis of any cancer, lung cancer, or haematological cancers. We assessed risk of bias using tools adapted from the Newcastle-Ottawa Scale. We used the generic inverse-variance random-effects method for meta-analysis. Pooled odds ratios (ORs) and hazard ratios (HRs) were calculated separately.
 Of 96 included studies, 54 had sufficient non-overlapping data to be included in meta-analyses (>500,000 people with COVID-19, >8,000 with cancer; 52 studies of any cancer, three of lung and six of haematological cancers). All studies had high risk of bias. Accounting for at least age consistently led to lower estimated ORs and HRs for COVID-19-related death in cancer patients (e.g. any cancer versus no cancer; six studies, unadjusted OR=3.30,95%CI:2.59-4.20, adjusted OR=1.37,95%CI:1.16-1.61). Adjusted effect estimates were not reported for people with lung or haematological cancers. Of 18 studies that adjusted for at least age, 17 reported positive associations between pre-existing cancer diagnosis and COVID-19-related death (e.g. any cancer versus no cancer; nine studies, adjusted OR=1.66,95%CI:1.33-2.08; five studies, adjusted HR=1.19,95%CI:1.02-1.38).
 
 Conclusions
 The initial evidence (published to 1 July 2020) on COVID-19-related death in people with cancer is characterised by multiple sources of bias and substantial overlap between data included in different studies. Pooled analyses of non-overlapping early data with adjustment for at least age indicated a significantly increased risk of COVID-19-related death for those with a pre-existing cancer diagnosis.
 
 Data availability
 All the original data of this study were available upon reasonable request to the corresponding authors (KC or JS).



Introduction
The World Health Organisation declared COVID-19 (a disease caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2)) a pandemic on 11 March 2020. The pandemic has led to considerable disease burden worldwide, with >182 million people infected and >3.9 million attributed deaths by 1 July 2021 [1]. The early days of the pandemic were characterised by high uncertainty and an urgent need to understand who is most at risk of severe disease, to enable targeted shielding and precautionary measures. Initial reports suggested people with pre-existing conditions including cancer were at higher risk of death, with mechanistic hypotheses including effects of a compromised immune system due to cancer itself and/or cancer treatment [2]. Based on early studies, it was further proposed that lung cancer may increase risk due to concomitant lung damage, while haematological cancer may increase risk due to immunocompromise (secondary to the myelosuppressive nature of treatments and or impact of progression) [3][4][5][6].
The pressing need for evidence to inform clinical practice and public health policy at the onset of the pandemic led to the rapid conduct, analysis and publication of studies under challenging circumstances. These were subsequently fast tracked for publication with or without formal peer review to facilitate real-time impact [7,8]. However, multiple expressions of concern regarding the methodological quality of these studies were raised, including lack of adjustment for confounders, inadequate ascertainment of cancer status and other comorbidities [8][9][10]. Moreover, several studies were based on overlapping samples, leading to difficulties in determining which studies provided independent evidence [11].
The early reports have influenced clinical and policy decisions with respect to the cancer care pathway (e.g. delays in the delivery of or modification of cancer treatments) and increased anxiety for people with cancer and those who support them [12][13][14][15][16][17][18][19][20]. These effects could increase the burden of cancer, further to other pandemic-related impacts such as suspension of HPV vaccination, cancer screening programmes, as well as delays in diagnosis and treatment due to overwhelmed healthcare systems and redeployment of services [19,21]. At different stages of the pandemic, decisions on the J o u r n a l P r e -p r o o f 4 prioritisation of vaccine provision and the emergence of SARS-CoV-2 variants continue to require timely production and synthesis of high-quality evidence.
As the initial evidence played a critical role in decision-making during the pandemic, it is crucial to examine the strengths and limitations of that evidence and to gain insights for ongoing evidence reviews. Consequently, we carried out a systematic review of the early studies (to 1 July 2020) that provide information on the question, -Do COVID-19 patients with cancer have a higher risk of COVID-19-related death than those without cancer?‖. We also carried out a separate systematic review, reported in a companion article, to examine whether people with cancer have higher risk of developing COVID-19. We identified, critically appraised and synthesised the results of early studies, focussing on sources of bias and methodological limitations, and the impact on the results.

Methods
The protocol for this systematic review was registered on PROSPERO (CRD42020191922).

Eligibility criteria
While we were particularly interested in COVID-19 mortality in cancer patients, because of potential limitations of cause-of-death coding and reporting in the first months of the pandemic, we broadened our inclusion criteria to include studies reporting COVID-19-specific or all-cause mortality after COVID-19 diagnosis (-COVID-19-related‖ death). Cohort and case-control studies were included if they reported COVID-19-related mortality for people with a previous diagnosis of cancer, compared to those who did not have a previous cancer diagnosis, or anyone with a COVID-19 diagnosis. Eligible exposures were previous diagnosis of any cancer, active cancer (cancer diagnosed or treated in the last year, or described as active), or specifically lung cancer or haematological cancer (based on biological hypotheses suggesting higher risks for people with these cancer types). Where the exposure was described as cancer with no further details provided, we classified this as -any cancer‖.
Studies restricted to populations with non-cancer-specific health conditions were excluded.

Information sources and search strategy
J o u r n a l P r e -p r o o f 5   Medline and Embase databases were searched on 3 July 2020 for English-language articles   published 1 January-1 July 2020 by combining database-specific subject headings and text terms for   COVID-19 and cancer or comorbidities (Supplementary Table 1). Reference lists of relevant systematic reviews and full-text articles were checked for additional potentially relevant studies. All COVID-19-related pre-prints posted until 1 July 2020 on BioRxiv and MedRxiv (https://connect.biorxiv.org/relate/content/181) and the SSRN website (https://www.ssrn.com/index.cfm/en/coronavirus/) were also scanned.

Selection process
Two reviewers (CC or DC) screened titles and abstracts of identified published articles against prespecified inclusion criteria with 10% assessed by both reviewers to ensure concordance. Titles and abstracts of pre-prints were screened by a single reviewer (SH). Full texts of potentially relevant articles were independently assessed for inclusion by two reviewers, with disagreements resolved by a third reviewer. Reasons for exclusion were recorded for all excluded full-text articles.

Data collection
Pairs of reviewers (chosen from CC, DC, VF, SH, HH, SY) independently extracted study characteristics and results for each included study, with disagreements resolved by third reviewer adjudication. The following information was extracted: publication status, study design, country, population characteristics, source of study population, study period, method of COVID-19 diagnosis, cancer definition and numbers, comparator definition and numbers, minimum possible length of follow-up, outcome definition, number of people with the outcome for those with and without preexisting cancer and, where reported, the effect estimate and 95% confidence interval (95%CI) and any covariates included in analyses.

Risk of bias assessment
For each study included in the meta-analyses, risk of bias was independently assessed by a pair of reviewers (chosen from SA, CC, DC, VF, SH, DO'C, JS, SE), using modified versions of the Newcastle-Ottawa Scale designed specifically to assess the risk of bias in observational aetiological cohort and case-control studies (Supplementary Tables 2-3) [22]. Differences were resolved by J o u r n a l P r e -p r o o f 6 consensus or adjudication by a third reviewer. The risk of bias was rated low, moderate or high for each of: selection of exposed and unexposed cohorts, co-interventions, exposure status ascertainment, reverse causation, outcome ascertainment, completeness and differences in followup, exclusions due to missing exposure or covariate data, adjustment for important confounders or over-adjustment, and the reliability of covariate data. As the full list of important confounders remains to be established, by definition, low risk of bias for this domain was not possible. Therefore, overall ratings for studies were limited to moderate risk of bias (low or moderate risk across all domains) or high risk (high risk for at least one domain). Studies were considered to have high risk of bias due to over-adjustment if they adjusted for an intermediate variable on the causal pathway between having cancer and death, e.g. the number of comorbidities including cancer or clinical indicators of COVID-19 severity.

Effect measures
Most studies reported the association between any pre-existing cancer diagnosis and death as adjusted and/or unadjusted odds ratios (ORs), risk ratios (RRs) or hazard ratios (HRs). Unadjusted rate ratios were calculated for studies with a general population comparator. If unadjusted effect estimates were not reported, we calculated ORs and 95%CIs from exposure-outcome cross tabulations with 0.5 added to each cell when there were zero cells [23].

Selection of studies for meta-analyses
To assess the impact of different comparisons, effect estimates, and adjustment for confounders on results, all analyses were conducted separately by combinations of effect measure, exposure measured, comparator, and study-type. Where a study reported the same effect estimate adjusted in more than one way, the effect estimate adjusted for the most covariates was selected unless there was a concern about over-adjustment.
To avoid data duplication, studies with overlapping samples were identified, and only the study with the largest number of people with cancer was included in a meta-analysis. Studies with insufficient or inconsistent data were excluded from the meta-analyses.

Meta-analyses
Pooled effect estimates and 95%CIs from generic inverse-variance random-effects analyses were calculated using Stata 14. ORs and RRs were pooled together in the same meta-analysis as the risk of death was <10% in both the cancer and comparison groups in all relevant studies [24]. If both ORs and RRs were available for a study, ORs were used as these were reported more often. HRs were pooled separately. To assess the effect of adjustment for confounders, we informally compared adjusted and unadjusted OR/RRs and HRs for studies where both were available (a statistical test was not possible as the estimates were obtained from the same studies).

Assessment of heterogeneity
Heterogeneity was assessed with the χ 2 test and I 2 statistic (see Supplementary materials).

Subgroup analysis and investigation of heterogeneity
For meta-analyses with sufficient numbers of studies, pre-specified subgroup analyses were performed for country, source of study population (general, hospitalised, hospitalised in ICU or with severe/critical disease), publication status (original journal article, pre-print) and covariates included in adjustment (age and sex only, >2 variables, over-adjusted). Possible subgroup differences were assessed using χ 2 tests.

Supplementary analyses
To assess the sensitivity of our primary results to our choice of analytical method, we repeated the two analyses with the most studies using fixed-effect rather than random-effects methods.

Reporting bias assessment
None of the meta-analyses of adjusted effect estimates included 10 or more studies, so we did not conduct pre-planned assessments of publication bias using visual inspection of funnel plot asymmetry and Egger's statistical test [25].

Results
In total, 12,225 records were identified and 96 studies satisfied the inclusion criteria ( Figure 1). The main reasons for exclusion were study design other than cohort or case-control study, or letter/comment without relevant data (Supplementary Table 4). Fifty-four studies [2,[4][5][6] were included in the meta-analyses after omitting 37 studies with overlapping samples and five studies with insufficient/inconsistent data to calculate effect estimates (Supplementary Table 5).
Characteristics of the 54 studies included in the meta-analyses are summarised in Table 1. These studies included over 500,000 people who developed COVID-19, of whom >8,000 had a pre-existing cancer diagnosis. Most studies were of hospital inpatients whose COVID-19 diagnosis was based on a SARS-CoV-2 PCR assay, and the minimum follow-up period was 0-30 days (0 days for the majority of studies). Four studies specifically reported deaths from COVID-19 or due to acute respiratory distress, while all other studies reported overall mortality only.
Of the studies included in meta-analyses, 19 provided information on cancer status (e.g. active or not) and 11 specifically restricted analyses to active cancer. Fifty-two studies either did not specify cancer type or only reported cancer type for sub analyses (three reported on lung cancer and four on haematological cancers) ( Table 1). Two studies included only haematological cancers. Details and results for the 17 meta-analyses conducted are shown in Table 2.
All 54 studies had high risk of bias (Tables 3-4). The main sources of bias were unclear or inadequate ascertainment of cancer status, potential differences in treatment or management of COVID-19 patients with and without cancer, limited or lacking ascertainment of confounders, and insufficient control for important confounders. Adjustment for important confounders was assessed to be at moderate risk of bias for the 15 of 18 studies that controlled for at least age (adjustments used in individual studies are listed in Supplementary Table 6).
In a comparison of unadjusted and adjusted ORs or RRs in six studies (Figure 2a) Table 2).
Given these differences, we focused the interpretation of results on studies that adjusted or controlled None of the studies of haematological or lung cancers adjusted for age ( Table 2 shows  There were no significant differences when stratified by country (p=0.28), covariates included in the adjustment (p=0.08), or source of study population (p=0.06). Only one of these nine studies reported estimates for COVID-19 as cause of death (as opposed to overall mortality), so we could not assess heterogeneity due to specific cause of death. We verified that the pooled estimate for these nine studies was similar when using a fixed-effect instead of random effect meta-analysis (pooled fixedeffect OR/RR=1.58, 95%CI:1.37-1.81; Supplementary Figure 18).

J o u r n a l P r e -p r o o f
10 Our systematic review synthesised early evidence on the risk of COVID-19-related death for people with cancer from 54 studies reporting on >500,000 COVID-19 cases with >60,000 deaths (including >1800 deaths for people with pre-existing cancer). Of the 96 studies that satisfied inclusion criteria, 37 studies included patients who were also included in other larger studies, complicating the identification of independent evidence. Many studies had short follow-up periods, a small number of people with cancer, and unclear definitions of cancer status. All 54 studies included in meta-analyses had high risk of bias, the majority with multiple sources of bias, leading to uncertainty regarding the strength of the association between pre-existing cancer and risk of COVID-19-related death. Only 18 studies adjusted effect estimates for at least age. Even minimally adjusted effect estimates were consistently smaller than the corresponding unadjusted estimates. Nonetheless, pooled adjusted estimates indicated a significantly increased risk of death for those with pre-existing cancer.
In the early stages of the pandemic, based on the precautionary principle, concerns regarding the risk of severe COVID-19 and death for people with cancer who are exposed to SARS-CoV-2 led to cancer treatment protocol changes in different countries and settings. However, treatment changes may also lead to cancer progression and death, so early high-quality evidence on the magnitude of risk is imperative. We found the early evidence was largely limited, and early systematic reviews were confined to unadjusted effect estimates and small sample sizes, and did not assess risk of bias for individual studies [76][77][78][79][80][81][82][83]. Our systematic review has provided in-depth critical assessment of the evidence generated early in the COVID-19 pandemic that can inform the design of future studies and ongoing reviews, but has some limitations. The titles and abstracts were not screened independently by two reviewers, although there was high agreement for the subset screened in duplicate. We did not contact the authors of included studies to clarify missing or unclear details, or to obtain additional information. For the risk of bias assessment, we could not identify all important confounders, as definitive evidence for conditions and characteristics associated with COVID-19-related death is yet to emerge. Given the heterogeneity in definitions of disease severity, we focussed on death as the outcome of interest, rather than more broadly and inconsistently defined severe COVID-19. However, our work also has several strengths including a comprehensive assessment of the early evidence, a focus on effect estimates adjusted at least for age, and in-depth risk of bias assessment.
We note that most of the early studies reported deaths from any cause after COVID-19 diagnosis, rather than COVID-19 specific deaths. Although cancer deaths could contribute to the elevated risk of J o u r n a l P r e -p r o o f death for COVID-19 patients with a pre-existing cancer diagnosis compared to those without, the contribution is likely be small given the short follow-up periods (usually two months or less) in the early studies. The World Health Organisation guidelines further specify that -A death due to COVID-19 is defined … as a death resulting from a clinically compatible illness, in a probable or confirmed COVID-19 case, unless there is a clear alternative cause of death that cannot be related to COVID disease (e.g. trauma). … A death due to COVID-19 may not be attributed to another disease (e.g. cancer) and should be counted independently of pre-existing conditions that are suspected of triggering a severe course of COVID-19‖ [84]. This is consistent with our analysis strategy in which all deaths of people with COVID-19 were treated as COVID-19-related. Later studies also considered COVID-19-related death as any death with COVID-19 mentioned on the death certificate (e.g. OpenSAFELY) [85], or confirmed/suspected COVID-19 deaths as described on the death certificate together with all deaths occurring in individuals with confirmed SARS-CoV-2 infection in the initial study period (e.g. QCOVID) [86].
While the early studies did not carry out in-depth analyses of time since cancer diagnosis or receipt of specific cancer treatments, several larger studies published after 1 July 2020 (which were also able to adjust for multiple potential confounders) suggest that both factors may increase the risk of COVID-19-related death. For example, the OpenSAFELY study, with >17 million general practice patients in the UK and >10,000 COVID-19 deaths [85], reported that diagnosis of non-haematological cancers closer to COVID-19 development was associated with higher risk of death, with no association for cancers diagnosed 5+ years prior to developing COVID- 19   12 neutropenia (CTCv4) or lymphopenia of <10%, 10-50% and >50%, respectively) [86]. This study also reported a positive association between risk of death from COVID-19 and radiotherapy receipt in the previous 6 months (women: adjusted HR 2.11 (95%CI:1.30-3.41); men: adjusted HR 2.09 (95%CI:1.48-2.96)).
As our systematic review focused on a critical appraisal of early evidence, the methodological insights gained from this review will inform the COVID-19 and Cancer Global Modelling Consortium (CCGMC; CCGMC.org) Observatory platform for the ongoing review and analysis of emerging data on risk of COVID-19-related death for people with cancer, supporting efforts for timely identification and synthesis of high-quality evidence. As part of the CCGMC efforts, the Observatory will help provide informed advice to support international decision-making in cancer control both during and after the pandemic, and we will aim to keep this review up-to-date as a living review. The Observatory will also include modelled estimates of COVID-19 impact, and the evidence from this review provisionally supports taking into account potential accelerated mortality in cancer patients with COVID-19, although more data are needed by cancer type and stage.
Recently, concerns regarding the risk of severe COVID-19 and death for people with cancer who are exposed to SARS-CoV-2 led multiple organisations to call for people with cancer to be prioritised for vaccination, including the American Association for Cancer Research, the National Comprehensive Cancer Network, and the European Society for Medical Oncology [87][88][89][90]. As the application of the precautionary principle for all high-risk groups only provides limited information, ideally, decisions on prioritisation would be based on a nuanced understanding of risks for different subgroups of the population, including risks for people with cancer depending on time since diagnosis, treatment type and time since treatment. Comparisons to other individuals at high risk would also be needed to evaluate the trade-offs in prioritising specific groups.
In the future, vaccination efforts in many countries may reduce the risks posed by COVID-19, but the emergence of several SARS-CoV-2 variants of concern will require ongoing monitoring of disease risk [91]. At the time of writing, there was little direct evidence on vaccine effectiveness in specific subpopulations of people with cancer. To enable assessments of vaccine effectiveness for people with cancer, the availability of immunisation registers and linkage to cancer registries, medical and death records will be important. As for analyses of COVID-19 outcomes, this would ideally also J o u r n a l P r e -p r o o f 13 include well-powered subgroup analyses by cancer type, treatment type, presence of other/specific comorbidities, and time since diagnosis and treatment. Linkage to comprehensive medical records would also facilitate adjustment for important confounders such as age and comorbidities, acknowledging that the recording of key covariates may still be incomplete, and some important factors (e.g. ethnicity) may be difficult to ascertain from routinely collected data. Thus, enhanced data collection for suitable surveillance cohorts is important. As provision of real-time information remains a challenge for many population-wide registries, to enable rapid and evidence-based responses to emerging variants, investments in infrastructure are needed to ensure high-quality near-time record linkage and accurate yet timely assessments of health impacts.
In conclusion, the early literature on risk of COVID-19-related death for people with cancer was characterised by pervasive biases and analytical limitations. Data from analyses adjusted at least for age suggest a higher risk of COVID-19-related death for people with cancer. Fine-grained analyses of surveillance cohorts of cancer patients and population-wide record linkage including cancer and immunisation registries, and real-time availability of clinical information will be important to inform the ongoing public health response to the COVID-19 pandemic.

Author contributions
KC and D'OC conceived the study. KC, D'OC, JS, SE, SH, VF, CC, DC designed the study. VF, SH, CC, DC, SE, HH, SY, SD analysed the data. VF, SH, CC, DC accessed and verified the data. VF, SH, CC, DC, SE, D'OC and JS wrote the manuscript. All authors contributed to data interpretation, reviewed, revised, and approved the manuscript, and accept responsibility to submit for publication.

Data availability
All the original data of this study were available upon reasonable request to the corresponding authors (KC or JS).

Funding
No specific funding was received for this study.

Competing interests
Prof Karen Canfell reports she is co-PI of an investigator-initiated trial of cervical screening, "Compass", run by the VCS Foundation Australia, which is a government-funded not-for-profit charity.
The VCS Foundation has received equipment and a funding contribution from Roche Molecular Diagnostics. She is also co-PI on a major implementation program "Elimination of Cervical Cancer in the Western Pacific" which will receive support from the Minderoo Foundation and the Frazer Family    Table 2). c) Meta-analysis of 5 studies that reported adjusted HRs (Analysis 4 in Table 2).
Whiskers represent 95% CIs. Estimates >1 represent higher risk of COVID-19-related death for people with a pre-existing cancer diagnosis.  Table 2). Whiskers represent 95% CIs. Estimates >1 represent higher risk of COVID-19-related death for people with active cancer.   Table 2, grouped by publication status): -pre-print‖ denotes pre-print articles not published by 1 July 2020. Estimates >1 represent higher risk of COVID-19-related death for people with a pre-existing cancer diagnosis.