key: cord-0874500-vx4ijhej
authors: Arch, B. N.; Kovacs, D.; Scott, J. T.; Jones, A. P.; Harrison, E. M.; Rosala-Hallas, A.; Gamble, C. G.; Openshaw, P. J.; Baillie, J. K.; Semple, M. G.
title: Evaluation of the effectiveness of remdesivir in treating severe COVID-19 using data from the ISARIC WHO Clinical Characterisation Protocol UK: a prospective, national cohort study.
date: 2021-06-21
journal: nan
DOI: 10.1101/2021.06.18.21259072
sha: a1a44fc81cd022d6ea155c469276ad5f0ba22d44
doc_id: 874500
cord_uid: vx4ijhej

Background Remdesivir was given UK early-access approval for use in COVID-19 in people aged 12 years and older on 26th May 2020 on the basis of unmet clinical need. Evidence on the side effects, complications of therapy and effectiveness of this therapy is lacking or conflicting. Methods Adults with severe COVID-19 treated with remdesivir were compared with propensity-score matched controls, identified from the ISARIC WHO Clinical Characterisation Protocol study of UK hospitalised patients with COVID-19. Remdesivir patients were matched to controls according to baseline underlying 14-day mortality risk. The effect of remdesivir on short-term outcomes was investigated (primary outcome: 14-day mortality). Effect sizes were estimated and adjusted for potential confounders using multivariable modelling. Results 1,549 patients given remdesivir and 4,964 matched controls were identified satisfying inclusion and exclusion criteria. The balance diagnostic threshold was achieved. Patients had symptoms for a median of 6 days prior to baseline; 62% were male, with mean (SD) age 63.1 (15.6) years, and 80% categorised as White ethnicity. Fourteen-day mortality was not statistically significantly associated with treatment (9.3% remdesivir vs. 11.9% controls, odds-ratio 0.80, [95% CI 0.60-1.07], p=0.116, adjusted for age, sex, number of key comorbidities, dexamethasone use, and diagnosis of viral pneumonia. Findings Treatment with remdesivir was not associated with a reduction in mortality in our primary endpoint at 14 days. Interpretation Remdesivir did not significantly improve mortality in this study. The findings are subject to the limitations of an observational study. Balance was achieved for measured baseline factors, but unmeasured confounders may account for observed treatment effect sizes. Funding Medical Research Council UK & National Institute of Health Research

At the time of designing this study, two clinical trials measuring the efficacy of remdesivir as a therapeutic in treating SARS COV-2 had published results: ACTT-1 and SOLIDARITY. ACTT-1 suggested that for those who required supplemental oxygen but not ventilation at baseline, remdesivir 

This study presents real-world data on the effectiveness of remdesivir use during a non-surge phase of the pandemic in the UK, specifically looking at patients for whom the ACTT-1 trial suggested would be most likely to benefit from remdesivir.

We show that during the pandemic, remdesivir was given to a wide demographic of patients in the UK (on average older than those in clinical trials). At 14-days post baseline no reduction in absolute mortality was observed. Propensity score matching achieved balance for measured baseline variables.

However as with all observational studies, differences between the groups in unmeasured variables that may influence clinicians but were not recorded in our study, are plausible.

Several therapeutic drugs licensed for use in other conditions have been trialled in the treatment of severe COVID-19. Remdesivir (Gilead Sciences, Inc.), was approved for use in people aged 12yrs and older affected with severe COVID-19 by the United Kingdom's (UK) Medicines and Healthcare products Regulatory Agency (MHRA) under their Early Access to Medicine Scheme (EAMS) from 26 th May 2020. EAMS status was withdrawn in July 2020, because remdesivir was commissioned for routine use in severe COVID-19 following an evidence review 1 by the National Institute for Health and Care Excellence (NICE). The data collection initiated to evaluate remdesivir has continued.

Remdesivir is a broad-spectrum antiviral drug that has shown activity against Ebola virus in vitro and in non-human primates. 2 It is an adenosine nucleotide prodrug administered via intravenous infusion, and once it is metabolised into its active form 3 , it inhibits the viral RNA-dependent RNA polymerase 4 , a conserved enzyme involved in viral RNA synthesis. Remdesivir has demonstrated in vitro efficacy against other emerging coronaviruses, such as MERS-CoV and SARS-CoV-1 5, 6 , and SARS-CoV-2 [7] [8] [9] [10] [11] . The half-effective concentration (EC50) values against SARS-CoV-2 were below 5M. This promising EC50 combined with high (>100) safety indices in cells, made remdesivir one of the principal compounds of interest for a clinical trial early in the pandemic. While in vivo studies also showed clinical benefits 2,12 , there are limits to what can be extrapolated from the animal models due to important differences in the pharmacokinetics of the drug and disease course, particularly in mice. 13 The pharmacokinetics of remdesivir have been reported in healthy adults, showing a favourable profile, 14 but they are yet to be reported in severely ill patients. At the time of design of this study, findings for two key clinical trials (ACTT-1 15, 16 , and SOLIDARITY 17,18 ) had been publicised (see Research in Context panel).

We used data from a prospective observational cohort study of UK patients hospitalised, using a unique dataset consisting of data collected using the ISARIC WHO CCP-UK protocol. 19 would have been eligible to be initiated on remdesivir and classified at clinical status level 5 of an 8point ordinal scale (hospitalised, requiring oxygen, but not requiring ventilation) (see below for full definitions of baseline and inclusion/exclusion criteria). The study period was patient baseline between 26 th May 2020 and 30 th November 2020, with 28 days follow-up post baseline (see below). Patients that initiated remdesivir treatment within 24 hours of baseline were compared with a propensity-score matched control group that received no remdesivir during hospitalisation. We followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) Statement checklist to guide transparent reporting of this study.

Patients were eligible for inclusion whether their COVID-19 was hospital or community acquired Standard of care evolved over time: corticosteroids, dexamethasone and hydrocortisone, became recommended for some patients part-way through the study period. Interactions between remdesivir and these corticosteroids were not expected 21 such that remdesivir patients will have received these in the same way as non-remdesivir patients. The drugs hydroxychloroquine and chloroquine phosphate were actively not recommended as concomitant medications for remdesivir from 3 rd September 2020 ( Figure 2 ).

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 21, 2021. ; https://doi.org/10.1101/2021.06.18.21259072 doi: medRxiv preprint

The control group was selected from the cohort of eligible non-remdesivir patients using propensity score matching. An optimal propensity score (log-odds of being given remdesivir at baseline) model was derived from the data using the following baseline variables as potential predictors prespecified in the statistical plan: month of baseline; ISARIC4C tier of participating centre (0/1/2); sex; age; broad ethnicity group (White/Asian/Black/Other); clinically extremely vulnerable status (yes [Any of the following: cancer, severe respiratory condition, on immunosuppression therapy, other]/none/unknown); binary indictors of the key comorbidities: diabetes, hypertension, obesity, chronic cardiac disease (CCD), chronic pulmonary disease (CPD), and asthma; where COVID was acquired (community / hospital); admitted to HDU or ICU at baseline (yes/no). A propensity score was then assigned to each patient and used to match remdesivir patients with up to four controls (see Statistical Methods below).

Data were routinely collected on patients at baseline, the first day of admission to an ICU, and then on death, discharge, or day 28 post hospitalisation depending on which was soonest. For some sites, daily CRFs for days 3, 6 and 9 were also collected. Remdesivir patients had daily CRFs completed for each day of remdesivir dosing, and on day 14 after remdesivir initiation. Data collection regarding safety was limited to a tick-box assessment of 30 complications that patients may have experienced during their hospital stay post baseline. Space was provided for free-text entry of other complications, and these were searched for text that indicated any of the 30 listed complications, and to identify commonly occurring others. There was no scope to measure severity or relatedness, and the quality of the data relied on what was recorded in medical notes. An extract of the database was made on 10 th January to assess (a) whether the sample size of eligible remdesivir patients was >500, and (b) the extent of missing data of key baseline variables and the primary outcome. This data extract was judged to be adequate for the purpose of our analysis. No outcome analyses were carried out prior to this decision.

The primary outcome was 14-day mortality. Secondary outcomes were: (1) time-to-recovery; (2) 

The study was planned prior to access to the data being granted and a statistical analysis plan (SAP) was published on the ISARIC website on 16 th December. 22 This was subject to internal clinical and statistical review. The manufacturer was also given opportunity to comment on the SAP. This review was discretionary, and suggested revisions were considered prior to finalisation; however none were considered substantive and the proposed methodology was unchanged.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 21, 2021. ; https://doi.org/10.1101/2021.06.18.21259072 doi: medRxiv preprint

The sample size was limited by the number of patients treated with remdesivir in practice and while a formal sample size calculation was not undertaken a sample size of more than 500 patients treated with remdesivir was considered sufficient for a meaningful analysis. No upper boundary was placed on the numbers included. The number of controls eligible for the study could not be predicted, but the total was expected to be greater than the number that received remdesivir.

Remdesivir patients were matched with to up to four controls using variable-ratio nearest-neighbour propensity-score matching with a calliper width of 0•2 standard deviations. 23 Controls were chosen without replacement, so could not be matched to more than one remdesivir patient. An optimal propensity score model was selected satisfying a pre-defined balance diagnostic threshold (absolute standardised difference in risk score ASDRS ≤ 0•1 24 , where risk score is the estimated underlying risk of 14-day all-cause mortality). Full details and rationale for methods are given in the appendix [p8].

Descriptive statistics were used to describe demographics, other baseline characteristics including preexisting comorbidities, treatments received during hospitalisation, and complications associated with hospitalisation. Outcomes were summarised by treatment group, but inference was obtained through multivariable modelling, adjusting for key confounders chosen a priori (sex, age-group, and number of key comorbidities), and if necessary, additional factors that on seeing the data were considered potential confounders. Models incorporated weights for the control group to account for the variable matching ratio (wj = 1/kj) where kj is the number of controls matched to remdesivir patient j). Stuart 23 suggests that for analysis purposes, the two groups may be treated as independent. Binary outcomes (e.g. 14-day mortality) were modelled using logistic regression; time-to-event outcomes were modelled using Cox Proportional-Hazards models (using a Landmark analysis if proportional hazards assumptions failed); the ordinal outcome -clinical status at 15 dayswas modelled with ordinal logistic regression. The primary outcome analysis was subject to sensitivity analyses regarding specification of the logistic regression model: e.g. adding propensity score as a covariate, adding interactions. Propensity score matching was carried out using the package MatchIt in R v3.6.1; all other analyses were carried out using SAS v9.4.

A total of 39,330 unique patients were identified from a data extract made on 8 th January 2021, with a baseline date between 26 th May and 30 th November 2020. A total of 9,278 patients satisfied inclusion/exclusion criteria, and of these, 6,513 were included in the matched analysis (see Figure 1, and Table S2 ). Patients included in the matched cohort came from all regions of the UK, were 62•1% male, with mean (SD) age 63•1 (15•6) years, and 79•7% of those with ethnicity recorded were . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 21, 2021. ; https://doi.org/10.1101/2021.06.18.21259072 doi: medRxiv preprint categorised as 'White' (see Table 1 ). They had a median (IQR) 4C Mortality Score of 9 (6, 12), meaning that most patients were classified as at intermediate or high risk of in-hospital mortality.

An optimal propensity score model was derived, with sufficient balance between the groups (ASDRS= 0•1). In the eligible cohort, remdesivir was found to have been more likely given: later in the study period, to younger patients, those without an extreme clinical vulnerability, with obesity, who acquired COVID-19 in the community, or who were not admitted to HDU or ICU at baseline (Table S3 ). The fitted primary outcome risk score model results are presented in Table S4 . Figure S1 presents distributions of propensity score split by group. The cohorts are well balanced with respect to baseline characteristics (Table 1 )in particular, 4C Mortality Score 25 statistics were almost identical.

The remdesivir group were generally more medicated post baseline than controls. They were more likely to have been given dexamethasone (93•9% vs. 61•7%) or antibiotics (89•7% vs 79•8%) during hospitalisation. Use of at least one corticosteroid other than dexamethasone was similar in the two groups (9•4% vs 10•7%). In both groups, the use of antiviral agents other than remdesivir was rare -3•0% of the control group were known to have received an antiviral agent (Table S5) . Dexamethasone use was identified as a factor that should be adjusted for in all inferential analyses, given its known efficacy as a therapeutic in treating COVID-19. 

A total of 140/1,507 (9•3%) patients in the remdesivir group and 565/4,734 (11•9%) patients in the control group died within 14 days of baseline. 6,202 patients were included in a logistic regression model ( Table 2 ). The OR of death at 14 days for remdesivir vs controls, adjusted for age, sex, number . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 21, 2021. ; https://doi.org/10.1101/2021.06.18.21259072 doi: medRxiv preprint of comorbidities, dexamethasone use, and viral pneumonia was 0•80 with 95% CI (0•60-1•07), p=0•116.

Time-to-recovery was found to vary with time and treatment ( Figure 3A ). During days 1-5, recovery was more likely in the control group; during days 6-8 there was no treatment effect; and for days 9-28, remdesivir was associated with a faster recovery (Table 3 ). Median (IQR) time to recovery was 9 (9-10) days for remdesivir, and 8 (8-9) days for controls. There is evidence to suggest that reduction in 28day mortality is associated with remdesivir ( Table 2 ): the p-value is 0.03, but uncorrected for multiplicity; this can be interpreted as an estimated 36 (95% CI: 20-290) patients needed to treat to prevent one death. Time-to-death over these 28 days was not significantly associated with treatment group ( Figure 3B , 3,303/4,964 (66•5%) controls. Non-invasive ventilation was more likely in the remdesivir group (Table   2 ). Remdesivir patients required more invasive mechanical ventilation or ECMO; and where data were available, median (IQR) duration of these interventions in days were 10 (5,16) for remdesivir, and 6 (3,14) for controls. Two outcomes could not be derived due to insufficient daily CRF data: length of time requiring supplementary oxygen, and time-to-first-ventilation.

In this analysis of data from a large UK-wide study, we found that remdesivir use was not statistically significantly associated with improved 14-day mortality. There was an absolute difference in incidence of 2.6 percentage points in favour of remdesivir, but this was not statistically significant after adjusting for potential confounders. The confidence interval for the adjusted odds-ratio indicated considerable uncertainty regarding potential effect of remdesivir of a 40% reduction, to having no effect. A small reduction in 28-day mortality was detected, though as a secondary outcome, this result should be considered hypothesis generating. Our study does not exclude the possibility that effectiveness might be present in specific patient subgroups.

Our study population was representative of severe COVID-19 patients of all ages, treated in hospitals across the whole UK, predominantly of white ethnic background, and male. Compared with both the ACTT-1 15 and SOLIDARITY 17 trials, our cohort was older, and less ethnically diverse. As age is a key factor in risk of death from COVID-19 absolute survival outcomes presented here should not be compared directly with the rates published in these trials. Patients in our study received many types of . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 21, 2021. ; https://doi.org/10.1101/2021.06.18.21259072 doi: medRxiv preprint treatment during hospitalisation, but the data do not indicate when these treatments were received, nor the dose, nor duration. Use of other antivirals was rare -the control group can be considered a 'Non anti-viral therapeutics' group. There is an indication that dexamethasone use (and by implication corticosteroid use) was greater in the remdesivir group, and also use of antibiotics. Complications during hospitalisation indicated some imbalance in the groups. The remdesivir group had higher recorded prevalence of viral pneumonia. This measure is difficult to interpret, and arguably our inclusion criteria define patients that were presenting with viral pneumonia at baseline. It may have little meaning recorded in the complication CRF, or may have been a proxy for greater baseline severity, or a secondary effect to the antiviral, though there is less biological plausibility for the latter.

Hyperglycaemia, ARDS, and liver dysfunction were also more observed in the remdesivir group. One in eleven patients in our cohort were recorded as suffering from acute renal injury or failure, but this was similar in the two groups. NIV usage was more likely in the remdesivir group. This could represent a higher level of illness in this group, which was not apparent from the baseline data. Alternatively, it could represent a lower threshold for escalation of care in this group, or a perception that escalation was less likely to be futile in this group. See appendix p17 for further discussion regarding secondary outcomes.

This analysis used data from a prospective observational study, using routine care data collected during a pandemic. There are limitations in that effectiveness estimates are not from randomised patients, and the data collected reflect local practice by the clinical teams at numerous hospital sites. The study was designed pragmatically to be simple enough to be rapidly implemented, using data that were being collected under a generic protocol. The analysis was planned and made open for comment prior to any data being provided for analysis. Data completeness for baseline characteristics and final clinical outcomes were found to be extremely good thanks to the diligence of participating sites throughout the country. Daily follow-up data was less available than expected, and this meant that two outcomes and clinical status at day 15 could not be derived as planned. Inclusion and exclusion criteria were used to define a cohort that represented the type of patient that would have been eligible to have received remdesivir, and with the greatest potential to benefit from it, according to the existing evidence base.

This created a clear analysis cohort with similar baseline level of severity of COVID-19. 73% of patients given remdesivir were excludedchiefly because their treatment did not start within 24 hours of baseline. We justify their exclusion as the data collection tool could not guarantee determination of clinical status at start of treatment at other times. Other remdesivir patients were excluded because they required respiratory support or ECMO at baseline. These excluded patients are of interest, in the overall assessment of effectiveness, but would have to be the subject of a separate study. Propensity matching was used effectively to select a control group that had a similar profile to remdesivir patients and balanced according to the risk-score balance diagnostic. The control group were slightly older, but had similar clinical frailty scores, and numbers of comorbidities. Propensity score matching is a powerful . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 21, 2021. ; https://doi.org/10.1101/2021.06.18.21259072 doi: medRxiv preprint design tool in an observational studyprovided all key factors that increased likelihood of treatment are considered when developing a propensity score model. A priori matching factors were principally demographic and baseline clinical characteristics that might be related to chance of being given remdesivir, or chance of death within 14 days, or both. Several baseline continuous measures were not prespecified in the SAP for inclusion in the development of the propensity score model: oxygen saturation, respiratory rate Glasgow coma score, urea and C-reactive protein (CRP). With hindsight these may have been important to consider, though missing CRP and urea data would have led to at least 20% of patients being excluded from the matching algorithm. Nevertheless, these variables all contribute to the 4C mortality score, for which good balance was observed.

It is possible that further benefit could be gained if remdesivir, or a similar orally available direct acting antiviral, could be given earlier in the disease process, when pharyngeal shedding and by inference viral replication in the lower respiratory tract is at its highest. 12 SARS-CoV-2-infected rhesus macaques were successfully treated when remdesivir dosing was initiated 12 hours after virus inoculum. The authors noted that the efficacy of such direct acting antivirals against acute viral respiratory infections usually drops with time after infection and stressed the importance of dosing humans as quickly as possible.

The ACTT-1 trial 15 confirmed that benefits associated with remdesivir were larger earlier in the disease course (<=10 days vs >10 days). Our study contains too few hospital-acquired patients to explore this hypothesis.

We note that liver dysfunction was increased in the remdesivir group. This is not unexpected, since raised transaminases are an expected adverse event in nucleoside analogues, and indeed alanine aminotransferase (ALT) was elevated in 7% of remdesivir clinical trial participants. 26 The current data does not allow us to distinguish the level of severity of this liver dysfunction or whether it was reversible.

Overall, our study does not provide evidence that remdesivir is of benefit in patients hospitalised with severe COVID-19.

This work uses data provided by patients and collected by the NHS as part of their care and support #DataSavesLives. ISARIC4C welcomes applications for data and material access through our Independent Data and Material Access Committee (https://isaric4c.net).

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 21, 2021. 

All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 21, 2021. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. 

Remdesivir (n=1,549)  Included in primary outcome analysis (n=1,507)  Not included in primary outcome analysis (n=42) Reasons not included: -transferred to another facility before day 14 (n=15) -unable to derive primary outcome due to missing data (n=27)

 Included in primary outcome analysis (n=4,734)  Not included in primary outcome analysis (n=230) Reasons not included: -transferred to another facility before day 14 (n=61) -discharged to palliative care before day 14 (n=17) -unable to derive primary outcome due to missing data (n=152) Analysis adfdfa lkhgljhgl . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 21, 2021. ; https://doi.org/10.1101/2021.06.18.21259072 doi: medRxiv preprint is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 21, 2021. ; https://doi.org/10.1101/2021.06.18.21259072 doi: medRxiv preprint

COVID 19 rapid evidence summary: Remdesivir for treating hospitalised patients with suspected or confirmed COVID-19

Clinical benefit of remdesivir in rhesus macaques infected with SARS-CoV

Therapeutic efficacy of the small molecule GS-5734 against Ebola virus in rhesus monkeys

Mechanism of SARS-CoV-2 polymerase stalling by remdesivir

Broad-spectrum antiviral GS-5734 inhibits both epidemic and zoonotic coronaviruses

Coronavirus Susceptibility to the Antiviral Remdesivir (GS-5734) Is Mediated by the Viral Polymerase and the Proofreading Exoribonuclease

SARS-CoV-2 and SARS-CoV differ in their cell tropism and drug sensitivity profiles

Lack of Antiviral Activity of Darunavir against SARS-CoV-2

Comparative analysis of antiviral efficacy of FDA-approved drugs against SARS-CoV-2 in human lung cells: Nafamostat is the most potent antiviral drug candidate

Identification of inhibitors of SARS-CoV-2 in-vitro cellular toxicity in human (Caco-2) cells using a large scale drug repurposing collection

An orally bioavailable broad-spectrum antiviral inhibits SARS-CoV-2 in human airway epithelial cell cultures and multiple coronaviruses in mice

Remdesivir potently inhibits SARS-CoV-2 in human lung cells and chimeric SARS-CoV expressing the SARS-CoV-2 RNA polymerase in mice

Tolerability, and Pharmacokinetics of Remdesivir, An Antiviral for Treatment of COVID-19, in Healthy Subjects

Remdesivir for the Treatment of Covid-19 -Final Report

Repurposed antiviral drugs for COVID-19 -interim WHO 25•0%) 1,151 (14•9%) 1,539 (16•6%) 388 (25•0%) 1,027 (20•7%) 1,415 (21•7%) 61-70 385 (24•8%) 1,345 (17•4%) 1,730 (18•6%) 385

N=5,712) White 1,046 (78•3%) 5,829 (84•4%) 6,875 (83•4%) 1,046 (78•3%) 3506 (80•1%) 4,552 (79•7%) 24•9%) 1,826 (23•6%) 2,212 (23•8%) 386 (24•9%) 1,208 (24•3%) 1,594 (24•5%) Hypertension

•4%) 218 (14•1%) 770 (15•5%) 988 (15•2%) Asthma 261 (16•8%) 1,091 (14•1%) 1,352 (14•6%) 261

N=6,460) 0 472 (30•5%) 1,974 (25•9%) 2,446 (26•7%) 470 (30•5%) 1,461 (29•7%) 1,931 (29•9%) 1 471 (30•5%) 2,486 (32•6%) 2,957 (32•2%) 470 (30•5%) 1,568 (31•9%) 2,038 (31•5%) 2+ 603

Glasgow Coma Score: n (%) (N=1,511) (N=7,510) (N=9,021) (N=1,508) (N=4,846) (N=6,354) Equal to 15 1,462 (96•8%)

Protein (mg/L) (N=1,365) (N=6,104) (N=7,469) (N=1,362) (N=4,102) (N=5,464) Median (IQR) 106 (61, 172) 72 (30

Defined to be mutually exclusive: Asian = any ticked from Arab, East/South/West Asian; Black = Black ticked, but no Asian categories; White = 'White ticked

Other = not categorised Asian, Black or White, but at least one category ticked in any other box except 'Unknown

No clinically vulnerabilities: all options ticked 'No'; Clinically vulnerable unknown: no boxes are ticked 'Yes', and any level of missingness

Page 17 of 22