Mortality in hospitalized patients with cancer and coronavirus disease 2019: A systematic review and meta-analysis of cohort studies

Search Strategy

We performed a search of peer-reviewed literature in the English language, conference abstracts, and trial registrations from the Medline (Ovid), Embase (Elsevier), and CENTRAL (Cochrane Library) databases; the World Health Organization Novel Coronavirus website¹³; and major international oncology meetings from 2019 to 2020 (last date of search, June 8, 2020) with the help of a health sciences librarian (L.O.). We developed the search strategy initially on Medline and then translated it to other databases (for the full search syntax and details, see Supporting Table 1). We prespecified the searches to studies written in or translated into English. The search and analysis strategy were registered on the PROSPERO international prospective register of systematic reviews (Center for Reviews and Dissemination, University or York) before data extraction (registration number CRD42020189350).¹⁴ Covidence was used for removal of duplication, storing, and managing citations across all databases.¹⁵ We used the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines to conduct our literature search and reporting.¹⁶

Study Selection

Two authors (A.D. and R.G.) independently screened all titles and abstracts after initial de-duplication. The inclusion criteria included studies reporting on adults aged >18 years who had an existing diagnosis of cancer and were subsequently diagnosed with COVID-19 infection by either: 1) polymerase chain reaction (PCR) or quantitative real-time PCR, or 2) clinical diagnosis based on presentation. We excluded studies with <10 patients diagnosed with both cancer and COVID-19, or without specific outcomes information on cancer patients, or nonoriginal studies (review articles, viewpoints, editorials, or commentaries). Then, 2 authors (A.D. and R.G.) independently performed full-text reviews to determine final study inclusion for data extraction. Specifically, studies were excluded if they had an incorrect study design, did not include a cancer-specific population, did not contain cancer-specific outcomes, were duplicate publications from larger studies (ie, a single-center study published separately from a multicenter study that included the same center), or had <10 patients in the study (see Supporting Table 2). All discrepancies were resolved by consensus and were further adjudicated by a third author (A.L.).

Data Extraction

Three authors (A.D., R.G., and S.A.) independently extracted all eligible studies. Important fields extracted included study characteristics, setting (inpatient vs inpatient and outpatient), patient demographics and comorbidities, cancer types (solid vs hematologic as well as individual cancer types), cancer-relevant details (including ongoing treatment vs not and the presence of existing comorbidities), COVID-19–relevant details, and outcomes of interest. All-cause mortality was our primary outcome and was defined as death from any cause within 30 days of initial COVID-19 diagnosis or in-hospital mortality if adequate follow-up had been achieved. For studies that only reported mortality during the study period with inadequate follow-up, we contacted the corresponding authors for updated data (see Supporting Table 3); if these data were not available or we did not receive a reply, the lack of follow-up information was marked as a bias under missing information on the quality-assessment form. Finally, we requested individual patient-level data from each study, and, if provided or available, we integrated the updated data or publication into the final data analysis. All discrepancies in data extraction were adjudicated by a fourth author (A.L.).

Quality Assessment

We used a modified version of the Risk of Bias in Nonrandomized Studies-of Interventions (ROBINS-I) risk-assessment tool from the Cochrane Method group.^{17, 18} Because most studies were retrospective case series and cohort studies without specific exposure or intervention, we focused quality assessment on 5 domains: bias in selection, bias because of missing data, bias in measurement of outcomes, bias in selection of reported results, and confounding if risk factors were assessed. A study was considered to have high overall bias if 2 of the 5 domains were considered to be high-risk. Finally, publication bias was examined by using funnel plots of study effect size relative to the standard error as a visual aid for assessing systematic heterogeneity or bias.

Statistical Analysis

The primary goal of the current study was to estimate the overall mortality among patients with cancer and COVID-19 infection and to assess drivers of heterogeneity between studies. For the main meta-analysis, we included studies that were conducted predominantly among hospitalized patients who had cancer and COVID-19 to avoid an admixture of different populations (ie, nonhospitalized patients who were inherently less sick and less likely to die).

We calculated the pooled proportion of mortality for hospitalized patients across studies using a random-effects model. An aggregate participant meta-analysis of proportion was conducted using the metaprop and meta packages from Stata version 16.1 (StataCorp).¹⁹ Specifically, CIs of studies were estimated using the Clopper-Pearson exact binomial method.²⁰ Effect size was estimated using the DerSimonian and Laird random-effects models after Freeman-Tukey double arcsine transformation of proportion to stabilize the variances in the pooled proportion.^{21, 22} Forest plots were generated to visualize pooled point estimates and CIs. Heterogeneity (between-study variance) was quantified using the I² statistic, in which a higher percentage suggests higher heterogeneity.²³ A subgroup analysis was performed for binary study characteristics (eg, solid vs hematologic cancer type) after pooling the results from each subgroup. Metaregression with a random-effects model was performed for continuous study characteristics (eg, age, percentage male, percentage receiving active treatment) to identify additional moderating factors that could explain the effect size inconsistencies. The R² percentage (variance explained by the model with included covariates/total variance without covariates) was used determined the best fit of the derived models.

Systematic Review

In total, we identified 3033 studies from 5 data sources (Medline, n = 759; Embase, n = 1267; CENTRAL, n = 24; the World Health Organization website, n = 978; and national oncology meeting abstracts, n = 5) (Fig. 1). After de-duplication, 1764 studies underwent title and abstract screening, and 80 full-text articles were assessed for eligibility. Eighteen studies were included in the final qualitative review. The excluded references and the reasons for exclusion are provided in Supporting Table 2.

Details of the 18 included studies are provided in Table 1.^{11, 12, 24-40} Two studies reported by Kuderer et al¹¹ and Rivera et al²⁸ both provided data from the COVID-19 and Cancer Consortium (CCC19); therefore, they were considered as 1 study for inclusion. Furthermore, because the latter publication provided detailed patient-level data as a supplement, we used those specifically relevant to patients with moderate-to-severe COVID-19 illness as a surrogate for hospitalized patients in the current review. Overall, there was significant heterogeneity in the inclusion criteria and variability in study follow-up periods. Five studies included both hospitalized patients and outpatients diagnosed with COVID-19 or failed to specify a particular study setting. Thirteen studies (including the CCC19 patient-level subgroup update and the Thoracic Cancers International COVID-19 Collaboration [TERAVOLT] hospitalized cohort³³) predominantly recruited patients admitted to the hospital and formed the core group for the final meta-analysis.

These studies were largely conducted and had representative population from the United States (5 studies), the United Kingdom (3 studies), Spain (3 studies), France (2 studies), Italy (2 studies), China (2 studies), and multiple European sites (1 study). The sample sizes of these studies ranged from 25 to 1149 patients. The median age of patients in all studies ranged from 58 to 77 years. Four studies reported on patients with hematologic malignancies only,^{24, 25, 27, 29} whereas others had an admixture of both solid and hematologic malignancies. Among these studies, there was wide variation in the percentage of patients receiving active cancer treatment (usually defined as cancer treatment received in the past 1 month), ranging from 26% to 100%. Although all-cause mortality was 1 of the major outcomes described consistently, the follow-up period and outcome reporting varied among different studies. Some studies defined 14-day or 30-day mortality, whereas others defined mortality during the study period or at hospital discharge. Given the difference in the mortality definitions and patient characteristics, there was wide variation among the mortality rates from 7% to 42%.

Although many studies attempted to assess the association between patient characteristics and mortality, most had a very high risk of bias because of confounding or missing data. Among the 4 studies that included a sample size >200 and had a relatively low risk of bias, commonly reported risk factors on univariable analysis included older age (3 studies), male sex (3 studies), higher number of comorbidity, including hypertension or cardiovascular disease (3 studies), COVID-19 severity score or intensive care unit admission (3 studies), and active or advanced cancer (2 studies) (see Supporting Table 4). Multivariable models from each study were created with different assumptions and parameters and thus could not be directly compared. Nonetheless, there were conflicting data on whether or not recent anticancer therapy (specifically chemotherapy) was harmful.^{12, 34}

Quality Assessment

Overall, 10 of the 18 studies were found to have low overall bias (Fig. 2).^{11, 12, 24-40} Bias because of missing data was 1 of the major flaws for some studies that lacked appropriate follow-up intervals^{27, 32} and uniform reporting of mortality period.^{31, 39} To decrease this bias, we contacted the corresponding authors of all the included studies and requested further clarification and updated mortality reports, if available. A few studies were also found to have bias in the selection of participants by selecting patients only from a specific subgroup, such as those with hematologic malignancies,²⁹ undergoing radiation treatment,³⁸ receiving immunotherapy,³⁷ using International Statistical Classification of Diseases and Related Health Problems, 10th revision (ICD-10) codes only,⁴⁰ or missing selection criteria.²⁵ Most studies had a high bias because of confounding for predictors of mortality given small sample sizes and lack of multiplicity or multivariable adjustment.^{24, 25, 27, 31, 32, 35, 36, 39} There was asymmetry in the scatter of small studies for hospitalized patients, demonstrating publication bias with under-reporting for smaller studies with decreased mortality (see Supporting Fig. 1).

Meta-Analysis

The 13 studies that included predominantly hospitalized patients formed the core of the meta-analysis (Fig. 3).^{11, 12, 24-40} Among the 2922 patients who had COVID-19 and cancer, the pooled 30-day mortality rate was 30% (95% CI, 25%-35%; I², 82%). This estimate was significantly different from the 5 studies comprising 624 patients with a mixture of inpatient and outpatient population; those studies had an estimated 30-day mortality rate of 15% (95% CI, 9%-22%; I², 73%). Because of the extreme differences in patient selection and outcomes, a pooled overall estimate was not performed, and further analysis was only dedicated to the inpatient group.

Among the hospitalized patients, heterogeneity remained high (I², 82%). We first examined cancer type (solid vs hematologic) as a prespecified moderator. For this analysis, we separated the 2 studies that reported cancer type-specific mortalities into the appropriate subgroups.^{12, 30} In total, 2539 patients from 9 studies formed the solid tumor subgroup (range, 81%-100%); in contrast, 383 patients from 6 studies formed the hematologic malignancy subgroup (100%). The pooled 30-day mortality rate was significantly higher in the hematologic subgroup (40%; 95% CI, 33%-47%) compared with the solid tumor subgroup (26%; 95% CI, 22%-31%) (Fig. 4).^{11, 35} Nonetheless, residual heterogeneity remained high, and patients in the hematologic subgroup were older and received more active therapy.

To further explore heterogeneity, we performed metaregression on continuous study characteristics in addition to cancer type, including the number of participants, median age, percentage of men, and percentage receiving recent active anticancer therapy. In univariable metaregression, the size of the study was not associated with the effect size (reported mortality percentage) (see Supporting Fig. 2A). In contrast, increasing age, male sex, and a higher proportion receiving recent active treatment were significantly associated with increasing effect size (see Supporting Fig. 2B-D). In multivariable meta-regression of the studies, the model that included the interaction between age and recent active therapy, along with male sex as a separate covariate, resulted in an optimally fitted model with an R² of 96%. Specifically, the association between a higher mortality rate and a higher percentage receiving active therapy was observed in studies with an older median age, but not in those with a younger median age, whereas male sex remained an independent predictor of a higher mortality rate (Fig. 5). After accounting for these patient-specific characteristics, cancer type was no longer associated with increased mortality.