key: cord-0882308-lqoky08e
authors: Mak, J. K. L.; Kuja-Halkola, R.; Wang, Y.; Hagg, S.; Jylhava, J.
title: Frailty and comorbidity in predicting community COVID-19 mortality in the UK Biobank: the effect of sampling
date: 2020-10-27
journal: nan
DOI: 10.1101/2020.10.22.20217489
sha: 9f5b3b0832b67e84a4d80ba186dd9b06f3fe472e
doc_id: 882308
cord_uid: lqoky08e

Frailty has been linked to an increased risk of coronavirus disease 2019 (COVID-19)-associated mortality, but evidence has been inconclusive and limited to hospitalized older individuals. Using data from the UK Biobank, we assessed whether frailty and comorbidity predict COVID-19 mortality in the overall community population (n=437,555) and in a selected COVID-19 positive sample (n=2,059). Frailty was assessed using the Rockwood Frailty Index (FI) and the Hospital Frailty Risk Score (HFRS), whereas comorbidity was assessed by the Charlson Comorbidity Index (CCI). Overall, 408 individuals died of COVID-19, as ascertained from the death register data. In the full sample, HFRS (odds ratio [OR] 1.07; 95% confidence interval [CI] 1.06-1.07) and CCI (OR 1.14; 95% CI 1.08-1.20) were associated with increased risk of COVID-19 mortality, while FI was not statistically significantly different from null in the multivariable logistic regression model. Adding HFRS or CCI to a model with only age and sex resulted in significantly larger areas under the receiver operating characteristic curves. Nevertheless, when restricting the analyses to COVID-19 positive cases, which is a sample with over-representation of frail individuals, neither of the frailty measures or CCI added meaningful predictive accuracy on top of age and sex. Besides, we observed stronger associations between HFRS categories and COVID-19 mortality in relatively younger (<75 years) than older individuals ([≥]75 years). Our results suggest that HFRS and CCI, which could be readily derived from medical records, may be useful for COVID-19 mortality risk stratification in the community.

Coronavirus disease 2019 , caused by the severe acute respiratory syndrome 39 coronavirus 2 (SARS-CoV-2), has led to a global pandemic, affecting more than 41 million 40 individuals and causing ~1.1 million deaths worldwide as of 22 nd October 2020 [1] . 41

Accumulating evidence has shown that older age, male sex, comorbidities (e.g. diabetes, 42 hypertension), social deprivation, black ethnicity, and laboratory indicators such as elevated 43 levels of d-dimer and interleukin 6 are risk factors for mortality associated with COVID-19 [2-44 6]. However, there are relatively few data available for risk stratification in community 45 samples compared to hospitalized patients. 46

Frailty, characterized as a state of increased vulnerability due to cumulative decline in 47 multiple physiological systems [7] , has consistently shown to be a strong predictor of mortality 48 in the general population [8-10]. Various tools have been developed for measuring frailty. 49

Some of which require assessment by physicians, such as the Clinical Frailty Scale (CFS) [11] , 50 which is more suitable in clinical settings. Another widely used measure is the Rockwood 51 frailty index (FI), which is defined as a ratio of accumulated deficits over the total number of 52 deficits considered [12] . The Hospital Frailty Risk Score (HFRS), constructed based on the 53

(ICD-10) coding system [13] , was developed for frailty risk stratification among older 55 hospitalized patients and has been validated for its ability to predict adverse outcomes in 56 various settings [14, 15] . 57

Centre Research Ethics Committee. All participants provided written informed consent for 85 data collection, analysis, and record linkage. 86 We excluded participants who died before 1 st March 2020, requested to withdraw 87 from the study prior to August 2020, and had missing data on frailty and comorbidity measures. 88

This resulted in a sample size of n=437,555, which we referred to as the "full sample". The 89 subgroup of "COVID-19 positive sample" (n=2,059) consisted of those being tested positive, 90 diagnosed as COVID-19 patients in hospitals, and/or died of COVID-19. Analyses were 91 performed in both samples (Fig. 1) those died of COVID-19 but without positive test record (n=89), and found that both groups 111 were generally comparable except that individuals in the latter group were more likely to be 112 older, with lower income and with higher HFRS (Appendix Table S1 ). 113

Frailty and comorbidity measures 115

Frailty was assessed using the FI and HFRS, and comorbidity was measured using the CCI 116 (timeline of data collection is shown in Appendix Fig. S1 ). The FI has previously been created 117 and validated by us for the UK Biobank participants, using 49 self-reported frailty items 118 assessed at baseline during 2006-2010 that cover a wide range of items for physical and 119 mental well-being (Appendix Table S2 ) [10] . The FI was calculated as the sum of the items 120 (deficits) present in an individual divided by the total number of deficits, for instance, an 121 individual with 7 deficits from 49 items would receive an FI of 7/49=0.14. The FI was used as 122 both continuous and categorical variable, the latter being categorized into four groups: 123 relatively fit (≤0.03), less fit (>0.03-0.1), least fit (>0.1-0.21) and frail (>0.21) [28] . HFRS and 124 CCI were computed based on the ICD-10 codes from hospital records [13, 29] . Only medical 125 records before 1 st March 2020 were included so that diagnoses due to or associated with 126 COVID-19 would not bias the results. The HFRS was derived based on 109 frailty-related ICD-127 10 codes, as previously described by Gibert et al (Appendix Table S3 ) [13] . While it was 128 originally developed for older individuals (≥75 years) who had been admitted to hospital 129 during the prior 2 years, we utilized all available ICD-10 codes for each individual for 130 calculation. Each of the 109 codes were assigned a weight ranging from 0.1 to 7.1, depending 131 on its strength of association with frailty. HFRS was then calculated by summing all the 132 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in was derived by summing weighted ICD-10 codes, based on 17 comorbidities with weights from 135 1 to 6 depending on disease severity and mortality risk (Appendix Table S4 ) [29] , and was 136 treated as a continuous variable in all analyses. Individuals who had missing hospital data were 137 those who had not been hospitalized or resided outside England (these data were only 138 available for England). To maximize data utilization, we first excluded individuals who 139 attended baseline assessment in Wales or Scotland and with missing hospital data, and then 140 coded the remaining individuals with missing hospital data as 0 for HFRS and CCI. As a 141 sensitivity analysis, we assessed whether including diagnoses from long ago would affect the 142 results by calculating 2-year HFRS and CCI scores using diagnoses assigned only during the past 143 two years (i.e., between 1 st March 2018 and 29 th February 2020). The FI, HFRS and CCI 144 correlated moderately with each other (Appendix Table S5 preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted October 27, 2020. ; https://doi.org/10.1101/2020.10.22.20217489 doi: medRxiv preprint ownership, and household overcrowding; higher scores correspond to higher level of 156 socioeconomic deprivation. 157

Descriptive statistics were calculated for the full sample and COVID-19 positive sample. Due 160 to the apparent over-representation of frail individuals in the COVID-19 positive sample, we 161 performed logistic regression to formally ascertain if frailty and comorbidity were 162 determinants for being COVID-19 positive. 163

In both samples, multivariable logistic regression models were applied to investigate 164 the associations of frailty and comorbidity (FI, HFRS, CCI, as continuous measures) with COVID-165 19 mortality, adjusted for age (as linear effect, after confirming that the age-mortality 166 relationship was approximately linear) and sex. Ethnicity, smoking status, and socioeconomic 167 variables were subsequently added into the models to test whether they had an effect on the 168 associations. Areas under the receiver operating characteristic curves (AUROC) were used to 169 assess the predictiveness of the different measures. Because the HFRS was originally designed 170 for individuals older than 75 years and previous studies have reported age-varying risks for 171 frailty [9,10,25], we additionally stratified the analysis by age <75 and ≥75 years, as well as 172 performed an analysis with an interaction term between HFRS (continuous) and age group. 173

We further performed a series of sensitivity analyses to assess the robustness of our 174 findings, which included (i) using categorical instead of continuous FI and HFRS variables; (ii) 175 using 2-year HFRS and CCI, constructed by ICD-10 codes from the past 2 years only; and (iii) 176 performing multinomial logistic regression models to account for non-COVID-19 deaths as 177 competing risk, where mortality due to COVID-19 or other causes than COVID-19 were 178 compared to those who were alive as of 24 August 2020. 179 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in CI 0.83-0.87) was similar to the age and sex-adjusted univariable models of HFRS and CCI (Fig.  209   2A) . After restricting the sample to COVID-19 positive individuals, all of these associations 210 were attenuated, and the predictive accuracies were decreased. In the fully adjusted model 211 (model 5), only CCI was marginally associated with COVID-19 mortality (OR 1.09; 95% CI 1.01-212 1.14). Frailty and comorbidity did not add predictive value on top of age and sex, as indicated 213 by similar AUROCs across all models (Fig. 2B) . We subsequently adjusted for ethnicity, smoking 214 and socioeconomic variables in both samples; associations of frailty and comorbidity with 215 COVID-19 mortality were not affected by these variables (Appendix Table S7 ). 216

In both samples, compared with older individuals (≥75 years), relatively younger 217 individuals (<75 years) had higher ORs for COVID-19 mortality across HFRS categories (Fig. 3) ; 218 there was also significant interaction between HFRS and age in the COVID-19 positive sample 219 (Pinteraction<0.001), but not in the full sample (Appendix Table S8 ). We also tested the 220 interaction between HFRS and sex, yet it was not statistically significant and thus we did not 221 further perform subgroup analysis by sex. 222 223

The predictive abilities of frailty and comorbidity for COVID-19 mortality were largely similar 225 compared to the main analyses when (i) using FI and HFRS as categorical instead of continuous 226 variables (Appendix Table S9 ), (ii) using the 2-year HFRS and CCI variables instead of the 227 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted October 27, 2020. ; https://doi.org/10.1101/2020.10.22.20217489 doi: medRxiv preprint original scores (Appendix Table S10 ), and (iii) accounting for competing risk by deaths due to 228 other causes than COVID-19 (Appendix Table S11 ). 229

Using data from the UK Biobank, we found that HFRS and CCI, measures of frailty and 232 comorbidity respectively, were viable predictors of COVID-19 mortality and added predictive 233 value on top of age and sex in the overall community population. The associations persisted 234 even after adjusting for ethnicity, smoking and socioeconomic variables. Nonetheless, among 235 COVID-19 positive individuals who were already more likely to be frail, HFRS and CCI did not 236 improve predictive accuracy for COVID-19 mortality in addition to age and sex. Stronger 237 associations between HFRS and COVID-19 mortality were seen among younger (<75 years) 238 than older individuals (≥75 years), indicating that the HFRS may be applicable for predicting 239 mortality risk in younger adults as well. 240

To the best of our knowledge, this is the first study that has assessed the associations 241 between frailty and COVID-19 mortality in the community population. We showed that frailty 242 was associated with an elevated risk of COVID-19 mortality, and that a HFRS constructed 243 based on ICD-10 codes was a stronger predictor than an FI calculated using self-reported data 244 at baseline. The CCI, a measure of comorbidity computed by ICD-10 codes, likewise predicted 245 COVID-19 mortality, which is in line with prior research showing a positive association 246 between comorbidity and COVID-19 deaths [18,21,31]. Together, our results imply that frailty 247 and comorbidity measures available in routinely collected medical records may be applied for 248 risk stratification of COVID-19 mortality in the overall community population. 249

However, the predictive accuracies of frailty and comorbidity for COVID-19 mortality 250 were reduced after restricting the sample to only those with the disease. It has been argued 251 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in 

In conclusion, HFRS and CCI, measures of frailty and comorbidity respectively, constructed 293 using routinely collected medical records, predicted COVID-19 mortality in the overall 294 community sample and added predictive value on top age and sex. However, similar effects 295

were not seen in those who already have the disease. Our findings suggest that identification 296 of frail individuals in the general population may be a viable strategy for COVID-19 mortality 297 risk stratification. 298 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted October 27, 2020. ; https://doi.org/10.1101/2020.10.22.20217489 doi: medRxiv preprint Tables and Figures   Table 1 Characteristics of UK Biobank participants in the full sample and COVID-19 positive sample Table 2 Logistic regression models for the associations of age, sex, frailty and comorbidity with COVID-19 mortality in the full sample and COVID-19 positive sample perpetuity.

preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in

The copyright holder for this this version posted October 27, 2020. Fig. 3 Associations between Hospital Frailty Risk Score and COVID-19 mortality stratified by age into younger (<75 years) and older (≥75 years) individuals in full sample (n=437,555) and in COVID-19 positive sample (n=2,059). Error bars indicate 95% confidence intervals. Models were adjusted for sex. Note: The point estimates for the stratified analysis can be found in Appendix Table S8 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted October 27, 2020. ; https://doi.org/10.1101/2020.10.22.20217489 doi: medRxiv preprint

An interactive web-based dashboard to track COVID-19 in real 324 time

Clinical course and risk factors for mortality of 326 adult inpatients with COVID-19 in Wuhan, China: a retrospective cohort study

Covid-19: People in most deprived areas of England and Wales twice as likely to 329 die

Risk factors for mortality among COVID-19

A comparison of mortality-related risk 333 factors of COVID-19, SARS, and MERS: A systematic review and meta-analysis: Mortality-334 related risk factors of COVID-19, SARS, and MERS

The association of 336 race and COVID-19 mortality

Frailty in elderly people

Predicts All-Cause and Cause-Specific Mortality in Community Dwelling Older Adults

342 lifespan: Evidence from the Canadian National Population Health Survey

A new method of classifying prognostic 397 comorbidity in longitudinal studies: Development and validation

Controlling the False Discovery Rate: A Practical and Powerful 400 Approach to Multiple Testing

Comorbidities Predicting COVID-19 and Mortality in the UK Biobank Community Cohort

COVID-19 405 and associations with frailty and multimorbidity: a prospective analysis of UK Biobank 406 participants

Prediction for Progression Risk in Patients 408

With COVID-19 Pneumonia: The CALL Score

Development and validation 410 of an electronic frailty index using routine primary care electronic health record data

Sociodemographic and Health-Related Characteristics of UK Biobank Participants with Those 414 Full sample (n=437,555) COVID-19 positive sample

By category, n (%): Low risk (<5)

Charlson comorbidity index; CI, confidence interval; FI, Frailty Index; HFRS, Hospital Frailty Risk Score; OR, odds ratio; ROC, receiver operating characteristic curve

This research was conducted using the UK Biobank resource, as part of the registered project 300 22224. 301