Firstly, the authors appreciate the reviewer for the time spent providing comments on this paper, and we offer our responses here to improve the current version of our draft. We have responded to the queries raised and edited the paper based on the suggested recommendations where we can. However, there are areas where we could not effect changes due to the constraints of using secondary data collected under the Clinical Characterisation Protocol designed by the International Severe Acute Respiratory and Emerging Infections Consortium (ISARIC) and the World Health Organization (WHO). ^[1] We hope that this revised version addresses the concerns, within the limitations of the data, and improves the clarity of the findings, as well as providing an improved version and making it more informative within the scientific literature.

REVIEWER COMMENTS:

This study compared the outcomes of patients with and without cancers who were hospitalized for COVID-19 using a larger dataset than most previous studies. It found that patients with cancer had higher 30-day in-hospital mortality, longer hospital stays, were less often admitted to intensive care units (ICU), and less received invasive mechanical ventilation (IMV), compared to non-cancer patients. However, the comparisons of outcomes other than 30-day in-hospital mortality may be misunderstood.

1. Page 2, Results. The period used for counting ICU and IMV admissions is unclear. Was it within the 30-day hospitalization or the entire follow-up period? Would it be better to revise “The hazard ratio of dying from cancer, …” to “The hazard ratio of 30-day in-hospital mortality between patients with and without cancers,

Authors’ response: This query has been resolved and edited fully in the paper, as the aim was to compare the hazard ratio of 30 days in hospital mortality between COVID-19 patients with or without cancer.

2. Page 5, Analysis and statistical method. Please replace “A survival analysis model” with the more precise term “A Cox proportional hazard model”, and describe how the authors addressed missing values of covariates?

Authors’ response:  This query has been resolved; a Cox proportional hazard model has been incorporated in the analysis and statistical methods, and patients with missing covariates were excluded. Patients with complete variables were analysed, and the results are presented in the tables.

3. Page 9, Figure 1. Did the authors aim to demonstrate that a larger length of hospitalization (LOH) indicates severe COVID-19? However, cancer patients might experience a larger LOH due to their weakened conditions rather than COVID-19. Without adjusting for confounders such as comorbidities, the result could be biased.

Authors’ response: We aimed to describe the distribution of LOH by cancer status without indicating the severity of the disease. We agree that there are additional factors which may contribute to the LOH among cancer patients – but these are outside of the aim for presenting the information in Figure 1. This figure can inform differences in health care resource use between these groups of patients. Furthermore, the finding in our paper agrees that patients with cancer might have a longer hospital stay due to other comorbidities.

4. Page 11, Table 4. Would it be better to revise “Hazards ratio of mortality among those with cancer, …” to “Hazard ratio of 30-day in-hospital mortality between patients with and without cancers, …”?

Authors’ response: This query has been resolved and the paper was edited according to recommendation.

5. Page 11, Table 5. The authors excluded 458,363 missing observations from the analysis, which constituted more than 80% of the data. Did the authors consider using more popular multiple imputation method to handle these missing values?

Authors’ response: More than >50% of patients had missingness for smoking, dementia, Chronic neurological disorder, Chronic rheumatological disorder, and obesity Understanding the underlying mechanism of missingness for some of the variables can be challenging. The standard implementation of multiple imputation assumes that missingness arises through a missing at random (MAR) mechanism. However, such tacit assumption can be questionable for some of the covariates. For example, it can be argued that missingness for dementia can be thought of arising from a not at random (MNAR) process – where the standard implementation of multiple imputation can be questioned. A rigorous investigation of the missingness mechanism and considerations for using principled statistical approaches is considered beyond the current scope of work. However, we have re-run the analysis in a further sensitivity analysis by restricting the data to covariates that have <10% missingness. Reassuringly, the obtained hazards ratio from the new multivariable model remains relatively robust.

Table 3 included in the analysis (n=102,184 patients, 16,105 events, and 458,363 missing observations excluded from the multivariable analysis).

Unadjusted hazards ratio (95% confidence interval)

(from Table 3)

Adjusted hazards ratio

(95% confidence interval)

Adjusted hazards ratio (95% confidence interval)

[Excluding predictors >10% missing]

Cancer

(reference: no)

1.16 (1.13-1.18)

1.20 (1.16-1.26)

1.17 (1.14-1.21)

60 years and above (reference: 0-59y)

2.01 (1.98-2.04)

2.63 (2.50-2.77)

2.26(2.22-2.29)

Diabetes mellitus

(reference: no)

1.4 (1.38-1.42)

1.20 (1.17-1.24)

1.29 (1.27-1.32)

Chronic pulmonary disease (reference: no)

1.31 (1.28-1.33)

1.33 (1.28-1.38)

1.30 (1.27-1.33)

Male

(reference: female)

1.11 (1.1-1.12)

1.24 (1.20-1.28)

1.13 (1.11-1.15)

Chronic cardiac disease

(reference: no)

1.2 (1.19-1.22)

1.26 (1.21-1.30)

1.18 (1.15-1.20

Obesity

(reference: no)

0.97 (0.95-0.99)

1.10 (1.06-1.15)

Excluded

Hypertension

(reference: no)

1.37 (1.35-1.38)

1.10 (1.01-1.14)

1.12 (1.11-1.14)

Dementia

(reference: no)

1.51 (1.48-1.55)

1.16 (1.10-1.22)

Excluded

Smoking

(reference: no)

1.04 (1.02-1.06)

1.04 (1.00-1.08)

Excluded

Asthma

(reference: no)

0.93 (0.91-0.95)

1.03 (0.98-1.08)

0.93 (0.92-0.96)

Chronic neurological disorder (reference: no)

1.13 (1.1-1.16)

0.95 (0.91-0.99)

Excluded

^[1] https://isaric.org/research/covid-19-clinical-research-resources/covid-19-crf/

Jalloh AT, Merson L, Nair D  et al. Association of cancer and outcomes of patients hospitalized for COVID-19 between 2020 and 2023 [version 1; peer review: 1 not approved].  F1000Research 2024,  13:673 ( https://doi.org/10.12688/f1000research.150761.1)

REBUTTAL

Firstly the authors would like to thank the reviewer for the time spent providing comments on this paper and the provision of two extremely helpful references. We note there is a significant delay between the receipt of the review and this response. This is partly due to the first two authors being deployed to Rwanda as part of international efforts to support its response to the Marburg outbreak in 2024. We provide our replies below and have edited the paper in line with the recommendations where we can. We have also explained where we can’t make changes largely due to the constraints of using secondary data collected under the Clinical Characterisation Protocol designed by the International Severe Acute Respiratory and Emerging Infections Consortium (ISARIC) and the World Health Organisation (WHO). ^[1] Our hope is that this revised version addresses the concerns within the limitations of the data and improves the clarity of the findings as well as providing an improved version that may attract a second positive review to enable this paper to be fully published and indexed. 

PR1. It is unclear to me the extent to which individuals were hospitalised for a specific reason (e.g. due to their cancer ) and identified as having Covid 19 or were hospitalised for respiratory symptoms and identified as having another diagnosis such as cancer. This is particularly important as there are conclusions about different care between these groups and the demographic characteristics of those with cancer are substantially different from others admitted and this may be driven by the reason for admission. Additionally, the implications for the difference in demographic characteristics are not fully explored in the discussion.

Response

Our study uses secondary data collected under the Clinical Characterisation Protocol designed by the International Severe Acute Respiratory and Emerging Infections Consortium (ISARIC) and the World Health Organisation (WHO). This is a strength of our study as it allows analysis of a large, international dataset. However, we also acknowledge the limitations in that, although the target population for the protocol is patients admitted due to COVID-19 illness, the data recorded did not include the primary reason for presentation at the hospital and subsequent admission as a validation of this population. In the Conclusion we acknowledge the lack of certainty of the reason for presenting to hospital and we have edited the paper to highlight this in the Strengths and Limitations section. However, we feel our Results are still relevant, and in line with a number of similar, published studies, as they report the treatment subsequent to admission of patients with confirmed COVID-19 with respect to the presence or absence of cancer. The results of this analysis can help to inform resource planning and non communicable disease management like cancer in future outbreaks of respiratory disease. We do agree that it is unknown if a treatment regime might have differed due to demographics or characteristics at admission and have added this to the Limitations and Discussion of Results.

PR2. The major results are the comparison of the outcomes of those with cancer versus those without cancer. The authors point out there are other individuals with conditions with increased risk in their comparison group. As such by combining all other groups the approach is likely to underestimate the increase in risk of those with cancer. This is possibly the output in figure 3, however this is not explicitly stated and the reporting of adjusted results for other variables (such as age and sex) appear to be without adjusting for this. In other results this is not addressed, e.g.  Data in Table 2 seems to be a direct comparison between those with and without cancer only.

Response

Thank you for pointing this out. The adjusted hazards ratio for cancer was obtained from a multivariable model that included age, sex, and country income level with no explicit adjustment made for further co-morbidities. We agree data presented in table 2 are crude that are not adjusted for any patient characteristics, however we sought to represent the data in relation to our study question.

Following your comments we have undertaken two further sensitivity analyses and present the findings below and in the revised paper. First by adjusting for the comorbidity status in addition to age, sex, and income status in the multivariable model to estimate the hazard ratio of cancer on mortality (sensitivity analysis table1). Second, a multivariable model with all the predictors listed in Table 2 included in the analysis (sensitivity analysis table 2 below). The findings from these two sensitivity analyses indicates that the qualitative association between any of the predictors and outcomes remain relatively stable with some or minor differences in the estimated hazards ratio, apart from chronic neurological disorder. However, it has to be cautioned that such a multivariable model with all the predictors included is subject to large missingness (as indicated in sensitivity analysis 2 table) that makes the analysis susceptible to bias due to missing data.

These results from the two sensitivity analyses are presented as Table 4 and Table 5 in the revised manuscript.

Sensitivity analysis table 1: Hazards ratio of mortality among those with cancer, adjusted for comorbidities

Adjustment set

Hazards ratio [95% CI]

Results presented in main manuscript

Not adjusted for any variables

(from Table 3)

1.16 (1.13-1.18)

Adjusted for age, sex, and income levels

(from Table 3)

1.18 (1.15-1.20)

Sensitivity analyses

Adjustment set:

age, sex, income level,

hypertension, diabetes, COPD, obesity, chronic cardiac diseases, dementia, asthma, neurological disorder, rheumatological disorder

1.18 (1.14-1.21)

Sensitivity analysis table 2: Multivariable model with all the predictors listed in Table 2 included in the analysis (n=102,184 patients, 16,105 events, and 458,363 missing observations excluded from the multivariable analysis).

Unadjusted hazards ratio

(from Table 3)

Adjusted hazards ratio (95% CI)

Cancer (reference: no)

1.16 (1.13-1.18)

1.20 (1.16-1.26)

60 years and above (reference: 0-59y)

2.01 (1.98-2.04)

2.63 (2.50-2.77)

Diabetes mellitus (ref: no)

1.4 (1.38-1.42)

1.20 (1.17-1.24)

Chronic pulmonary disease (ref: no)

1.31 (1.28-1.33)

1.33 (1.28-1.38)

Male (ref: female)

1.11 (1.1-1.12)

1.24 (1.20-1.28)

Chronic cardiac disease (ref: no)

1.2 (1.19-1.22)

1.26 (1.21-1.30)

Obesity (ref: no)

0.97 (0.95-0.99)

1.10 (1.06-1.15)

Hypertension (ref: no)

1.37 (1.35-1.38)

1.10 (1.01-1.14)

Dementia (ref: no)

1.51 (1.48-1.55)

1.16 (1.10-1.22)

Smoking (ref: no)

1.04 (1.02-1.06)

1.04 (1.00-1.08)

Asthma (ref: no)

0.93 (0.91-0.95)

1.03 (0.98-1.08)

Chronic neurological disorder (ref: no)

1.13 (1.1-1.16)

0.95 (0.91-0.99)

Chronic rheumatological disorder (ref: no)

1.13 (1.1-1.16)

0.96 (0.93-1.00)

PR3. The analysis does not include an examination of change over time of risk associated with cancer. It also does not look at the impact of vaccination on risk (which we become more prevalent over time).  There is literature on changing risk over the course of the pandemic of individuals with cancer - e.g. Ref 1 

and the impact of vaccination and vaccine response Ref 2 .

Response

Our study includes patients from January 30, 2020, to January 10, 2023. During this period, COVID-19 underwent significant changes in genomics, treatment, and epidemiology, with vaccines introduced at varying times across countries. However, our dataset lacks genotyping and reliable vaccination information, which are crucial for analyzing temporal changes accurately.

Without data on these key factors, especially vaccination status, we cannot provide a robust analysis of changes over time. The impact of evolving vaccination rates on outcomes is likely substantial but impossible to calculate with our current data.

We acknowledge this limitation more explicitly in this revision and our findings have already informed improvements to ISARIC's case report forms for future outbreaks to address these data gaps.

PR4. It is unclear on the rationale of covariates included in the analysis and why others have been omitted. For example, a key strength is the international nature of this dataset, which allows exploration of differences between countries. It is stated in the paper the authors are unaware of any other analysis looking at outcomes using and international data set. There is between countries differences in provision of health care, even of those with similar income level. However, the rationale for why country income level was chosen and its implications is not discussed or why a more granular variable reflecting country could not be used. Rather it is stated that due to the predominance of cases from high income countries they were unable draw inferences, suggesting that using world bank income classification is not an appropriate variable to use. The authors need to reconsider what is used to assess international differences and include this in the analysis. Other examples, including those mentioned above, vaccine status, other clinical comorbidities, likely variant (or time as a proxy), severity at admission (or length of hospitalization)  etc, 

Response

Despite its faults, our original analysis plan included an analysis of differences in patient outcomes between countries with different World Bank classifications as a proxy for the quality of healthcare systems between these two economic realities. Other analyses of this database have used this approach due to the sensitivities and inaccuracies of labelling a select collection of hospitals in participating countries as country-level results. We have added this explanation to the methodology. While we agree that there are huge differences between health care institutions within these national income brackets, we feel that presenting a descriptive analysis is useful to identify a signal of difference that should be explored with a targeted study or datasets that include the details needed. The impact of comorbidities has been explored in our analysis. Unfortunately, we did not have data on vaccination status.

(There was no PR5 in the peer review report)

PR6. The presentation of results needs review to ensure it is clear. For example, the Figure 2 Kaplan Meier plot references inclusion of people between 2020-2023 - this is the only reference to a time period, the x axis is labelled as time and is in days. I assume this is the risk from the date of admission, however that is not explicitly stated - as such the presentation is not clear.

Note,  If it is the number of days hospitalized, rather than the days since admission, part of the downward trend may be driven by those no longer hospitalized being removed from the analysis as time continues, leaving only those hospitalized for longer being included in survival risk at later timepoints . Without clear description in is challenging to determine if this is a relevant point.

Other examples include Figure 3 - while the graphical representation is helpful - it is not possible to ascertain the exact hazard ratio and confidence intervals are omitted

Response

The study population statement clarifies that patients were enrolled between 30 ^th January 2020 and 10 ^th January 2023 and were followed for up to 30 days from admission.  We agree the labelling and legends of the tables and graphs should be improved to provide more context as standalone figures and tables and have edited the paper accordingly. We have additionally removed Figure 3 as these results are presented with 95% CI in Table 3, and further in Tables 4 and 5.

PR7. It is stated the presence of cancer was self-reported, It may be that more detail of this process would clarify potential biases, however given findings that the cancer group were, for example, admitted less to ICUs it is conceivable that those presenting with more severe symptoms at the time of recruitment were unable to self-report if they had cancer. Greater clarity on either the process of collecting relevant clinical details or discussion of potential implications of data collection approaches are needed.

Response

Cancer status was obtained from patients who were able to report or family members for many who could not. Those who could not report and did not have family were excluded from the analysis, and hence the potential for bias to arise due to lack of patients’ ability of self-report is likely mitigated in this analysis.

However, we appreciate that this approach can lead to bias if those with missing self-reported cancer status have different characteristics than those who were included in the analysis. Appropriate analysis requires investigation of the mechanism that led to the missingness of the self-reported cancer status and this was beyond the scope of the current work.

We have clarified this in the Methods section to make this more explicit and reduce the concern for bias.

NB we note reference 14 to IDDO is incomplete and the link to the CRF as ref 14 in the Introduction is incorrect it should link to ref 15.

We have corrected this in our revision. Thank you.

^[1] https://isaric.org/research/covid-19-clinical-research-resources/covid-19-crf/