key: cord-0102229-wcei23g9 authors: Kirwan, Peter; Charlett, Andre; Birrell, Paul; Elgohari, Suzanne; Hope, Russell; Mandal, Sema; Angelis, Daniela De; Presanis, Anne title: Trends in hospitalised mortality risk and lengths of stay during the first, second and current waves of COVID-19 in England: a cohort study date: 2021-12-20 journal: nan DOI: nan sha: 6ecfd306cd04dc5f403d34c7f2feb637038684ac doc_id: 102229 cord_uid: wcei23g9 Widespread vaccination campaigns have changed the landscape for COVID-19, vastly altering symptoms and reducing morbidity and mortality. We estimate monthly trends in mortality and the impact of vaccination among patients hospitalised with COVID-19 in England, controlling for baseline demographics and hospital load. Among 259,727 hospitalised individuals, 51,948 (20.0%) experienced mortality in hospital. Hospitalised fatality risk ranged from 40.3% (95% confidence interval 39.4-41.3%) in March 2020 to 8.1% (7.2-9.0%) in June 2021. Older patients and those with multiple co-morbidities were more likely to die or else experienced longer stays prior to discharge. Compared to unvaccinated patients, the hazard ratio for mortality following hospital admission was 0.72 (0.67-0.77) with a first vaccine dose, and 0.58 (0.54-0.62) with a second vaccine dose. The prognosis for patients hospitalised with COVID-19 in England has varied substantially throughout the pandemic and is influenced by baseline demographic factors, vaccination status, and hospital load at admission. The extensive vaccination campaign in England during 2021 has dramatically changed the outlook for COVID-19, lessening symptoms and reducing morbidity and mortality [5, 6] . Despite widespread and high levels of vaccination, however, individuals continue to experience COVID-19 infection severe enough to require hospitalisation. Several studies have previously examined hospitalised fatality risk (HFR) in England according to baseline demographic factors [1] [2] [3] [7] [8] [9] , although limitations exist in the coverage of the data and data quality, and there is little information on hospitalised outcomes in the current context of vaccination across the population. Information on the impact of hospital load on outcomes in England is similarly lacking, nevertheless increased hospital load has been linked to poorer outcomes for COVID-19 in Switzerland [10] , and a recent King's Fund study concluded that a shortage of overnight and acute bed availability prior to the pandemic put hospitals under increased strain [11] . In general, studies of hospital cohorts may include several end-points e.g. death or discharge, with a wide variety of survival analysis techniques available to quantify differences in outcomes between groups of patients, a detailed introduction can be found in e.g. [12] . If the competing risks of events are not independent, and the magnitude of the competing risk is large, the assumptions of conventional survival analysis methods such as Kaplan-Meier and Cox proportional hazards regression may provide biased estimates of risk, or of effect sizes on the rate of events, respectively [13] . In the case of hospitalised mortality for COVID-19, the competing risk of discharge may well be large, e.g. for individuals discharged to palliative care. Two survival analysis techniques which account for these competing risks are Aalen-Johansen cumulative incidence [14] and Fine-Grey regression [15] , details of which are included in the supplementary materials. Another form of bias, often ignored in epidemic studies, is the relationship between the time from infection to symptom onset, and an individual's eventual outcome: e.g. those who go on to die may experience more rapid onset of symptoms following infection. Since estimates must be conditioned on an observed quantity (e.g. symptom onset date or hospital admission date) rather than the unobserved infection date, this relationship can introduce bias into results when an epidemic is in a mode of growth or decline [16] . The resulting bias has been termed "epidemic phase bias" and may result in hazards being over or underestimated. We aimed to investigate trends in hospitalised mortality and the impact of vaccination, hospital load, and other factors among patients hospitalised with community-onset COVID-19 in England. We apply statistical methods which account for competing outcomes to estimate absolute and relative risks of hospitalised fatality, and lengths of stay in hospital according to outcome, and assess the potential impact of epidemic phase bias. In this retrospective cohort study we considered data on hospital admissions in England Covariates in the linked dataset included vaccination status (no vaccine, <21 days of first dose, ≥21 days after first dose, ≥14 days after second dose), date of hospital admission (aggregated by month), age group, region of residence (Government office region), Charleson comorbidity index (CCI) [17] , ethnicity, sex, index of multiple deprivation quintile, and a measure of hospital load. The hospital load measure was defined as the number of COVID-19 admissions at an NHS trust within the 7 days around admission (3 before, same day, and 3 after), as a proportion of the busiest 7-day period at that trust. Hospital load was grouped into: 0-20%, 20-40%, 40-60%, 60-80%, 80-90%, and 90-100%. In relative risk analyses the two key exposure variables of interest considered were vaccination status and month of hospital admission. Data comprised all new admissions for COVID-19 reported in England. Numbers of reported admissions were compared with the NHS COVID-19 Situation Reports to ensure data were representative. Hospital-onset COVID-19 (i.e. infection occurring in hospital) cases were excluded: those with hospital-onset infection (n=194,888) tended to be older and have longer lengths of stay than the community-onset cases considered in this study. Data validation was undertaken between the linked datasets, we found no systematic underreporting or mis-reporting of patient characteristics and linked information was used to minimise missing data. Censored outcomes and competing risks were explicitly accounted for by the choice of statistical method. We carried out a sensitivity analysis to assess the potential effect of epidemic phase bias on the estimated hazard ratios in relative risk analyses. Since this bias is caused by conditioning on an observed date later than the date of infection, for this analysis we condition on date of symptom onset, which is nearer to date of infection than date of hospital admission. This should ensure the sensitivity analysis targets bias due to epidemic phase, as opposed to any other factors which may influence time from symptom onset to admission (see Supplementary Information for further details). The SUS dataset receives data daily although has significant reporting delay as records are only submitted once a patient leaves hospital (due to discharge or death). Linkage to ECDS was therefore used to ascertain information for patients initially admitted via emergency care who remained in hospital. Among those with completed hospital episodes, 77% were admitted via emergency care. Two complementary statistical analyses were undertaken to understand both the absolute and relative risks of hospitalised fatality. We used Aalen-Johansen cumulative incidence estimation to obtain estimates of cumulative HFR and median lengths of stay in hospital for specific sub-sets of the population, unadjusted for other factors [14] . We used stratified Fine-Grey competing risk regression with adjustment for confounders to estimate the association of each risk factor with the cumulative incidence of hospitalised fatality; modelling with proportional hazards a "sub-distribution hazard" of hospitalised fatality derived from the cumulative incidence function [15] . Stratification was used for confounders with nonproportional hazards. See Supplementary Information for further details of these methods. To focus our analyses on outcomes following COVID-19 admission, a pragmatic cut-off of 90 days from first positive specimen date was chosen and only those hospital outcomes (death or discharge) occurring within this cut-off were included. All records with outcomes occurring beyond 90 days (n=656) were right-censored at 90 days, meanwhile patients who remained in hospital at the date of data extraction (n=15,460) were right-censored at the shorter of this date or 90 days. To account for palliative discharge, deaths occurring within 14 days of discharge from hospital were classified as deaths rather than discharges and the date of death used as the outcome date. Linkage to UKHSA deaths data enabled these posthospital discharge deaths to be identified. Statistical models were implemented using R version 4.1.1 (R Foundation for Statistical Computing, Vienna, Austria) and packages flexsurv 2.0, survival 3.2-12, and matrixStats 0.60.0. in hospital at the date of data extraction and/or were right-censored at 90 days. Figure 1 presents weekly hospital admissions for COVID-19 over the study period, with an indication of the first, second, and current (third) waves. Compared to all people with PCR-confirmed community-acquired COVID-19 infection, those hospitalised for COVID-19 were older (48.4% aged over 65 vs. 10%), more likely to be male (52.1% vs. 47.6%), and to reside in London (18.0% vs. 16.3%). A greater proportion of those hospitalised were of Black ethnicity (5.9% vs 4.2%) and lived in an area of high deprivation (28.1% vs 23.6%), compared to all those with community-acquired COVID-19. Comparative information on comorbidity was not available, although 7.2% of those hospitalised had a CCI score of 5 or more compared to 2.3% among a sample of 657,264 patients 20 years and older registered at English primary care practices in 2005 [18] . In unadjusted comparisons, older patients experienced poorer outcomes following hospital admission; almost half (46.5%) of those aged 85+ died in hospital compared to just 0.5% of those aged 15-24. Similarly, males (as compared to females), those living outside of London and the South West, and those with an increased CCI score (compared to lower CCI) were more likely to die in hospital (Supp. Table 1 ). Table 2 and Figure 2A 10 .5% (9.7-11.4%) and was maintained at or below 10% throughout subsequent months ( During the initial months of the study those admitted to hospitals which were experiencing higher load had a greater HFR as compared to those admitted to hospitals with lower activity e.g. during March 2020, HFR was 44.9% (39.6-50.9%) for hospitals at 90-100% of their peak load, compared to 38.8% (36.6-41%) for hospitals at 0-20% of their peak load. This disparity appeared to lessen during subsequent months, although estimates in the 90-100% load category are relatively uncertain (Supp. Table 6 ). Figure 2B show estimated median lengths of stay until death or discharge by month of hospital admission. Aside from the first two months of the pandemic (March-April 2020), those with an eventual outcome of death had longer stays in hospital compared to those who were discharged. Trends in length of stay prior to death and discharge followed approximately inverse trends: whilst length of stay prior to discharge decreased throughout the first wave, from 5.9 (5.8-6.0) days in March 2020 to 3.1 (2.9-3.3) days in August 2020, length of stay prior to death increased from 5.6 (5.5-5.6) days to 9.9 (9.2-10.9) days. During the second wave, lengths of stay prior to discharge initially lengthened, peaking at 5 (5-5.1) days in December 2020, before falling to 2.7 (2.6-2.8) days by June 2021. Conversely, length of stay prior to death was shortest in January, at 7.5 (7.5-7.6) days, and lengthened to 10.4 (10.1-10.8) days by June 2021. Examining lengths of stay prior to discharge and death for different subgroups, similar patterns were observed for males and females, although with less pronounced variation among males. Length of stay prior to death estimates were imprecise for younger individuals, due to the small number of events, but for age groups 45-64 and above, the median length of stay prior to death decreased with increasing age. Meanwhile older individuals had longer stays in hospital prior to discharge compared to younger individuals; the median time to discharge among those aged 85+ ranged between 5.1-10.4 days, compared to between 0.8-2.4 days for those aged 0-14 and 15-24 (Supp. Tables 2-5). Figure 4 ). There was an increased hazard of hospitalised mortality for those of Asian (1.19 (1.15-1.23 Supplementary Figure 7 shows the outcome of the shift sensitivity analyses by month of symptom onset, adjusted for the same covariates as above. The greatest effect was observed for the March 2020 hazard ratio estimate, which steadily reduced towards 1 following a shift of c=1, 2, 3 or 4 days. The effect in other months was small, with the previously described monthly trends persisting, although the slight reduction in hazard estimated for the most recent month (September 2021) was no longer apparent. We examined absolute and relative risks of hospitalised fatality and lengths of stay in hospital during the first year and a half of the COVID-19 pandemic in England. In line with epidemiological studies from UKHSA and others [7, [19] [20] [21] , we found that people with community-acquired COVID-19 who became hospitalised were older, more likely to be male, of Black ethnicity, and to live in areas of high deprivation, as compared to everyone diagnosed with the virus. Among those who were hospitalised, we estimated greater absolute fatality risks for men and older individuals, and HFR also varied according to ethnicity, month of admission, hospital load, and region. Lengths of stay in hospital were similarly influenced by demographic factors, with median lengths of stay prior to death typically longer than those prior to discharge. In relative risk analyses controlling for all measured confounders, baseline comorbidity burden was the strongest predictor of death. Our estimates suggest a deterioration in survival as hospital load increases, however, there are several potential biases which make this harder to interpret. It has been suggested that during periods of peak hospital load there is likely a modification of intensive care admission criteria, with only the most severe cases being admitted. Meanwhile, individuals with milder disease may be selected for transfer from overloaded hospitals to those with bed availability, due to the lower inherent risk. Hospital load-dependent changes in outcome are not unique to the English setting, with a recent Swiss study estimating poorer patient outcomes at times of increased hospital load; an ICU occupancy of 70% or greater found to be a tipping point at which outcomes became adversely affected [10] . There is now compelling evidence that vaccination reduces the number of individuals being hospitalised [5] and the risk of mortality, regardless of hospital admission [9, 22] . We found reduced hospitalised mortality among vaccinated patients, with the reduction most clearly seen among older individuals. For those aged 75 and over, vaccination reduced HFR to approximately the risk of an unvaccinated individual aged 10 years younger. In adjusted estimates, each additional vaccine dose reduced the hazard for fatality by a significant margin, with a 42% (38-46%) estimated reduction in the risk of death for double-vaccinated individuals. This is a slightly lower reduction than for all community-acquired PCR-positive COVID-19 cases in England, where a 51% (37-62%) reduced risk of death was estimated for symptomatic patients who had received a single vaccine [5] . This difference may reflect the portion of hospitalised patients who die from other causes, or could be an indication of waning vaccine efficacy among our study population. After controlling for all measured covariates, including hospital load and vaccination, we continued to estimate monthly variation in outcomes, with apparent seasonal variation in hazards. Whilst seasonal patterns in respiratory pathogens such as influenza and respiratory syncytial virus are well-documented [23, 24] , a multitude of interlinked factors including changes in national restrictions and the emergence of new variants may have influenced these trends. The use of high-quality hospital surveillance data linked to several other comprehensive data sources is a strength of this study and enabled a broader understanding of the factors influencing hospitalised mortality. For covariates with varying levels of completeness we undertook sensitivity analyses to confirm minimal effects on our estimates (e.g. indication of injury as a factor for emergency care admission), however, there may have been other unmeasured confounders for which we could not account. Using carefully chosen statistical methods we adjusted for competing risks, and the use of a relatively course monthly timescale likely limited the extent to which our study was affected by epidemic phase bias [16] . Data on hospital pathways following admission were unavailable. As such, we were not able to subdivide the hospitalised population by severity of infection and/or need of respiratory support, whether within or outside of intensive care. Treatment data and changes in patient management were similarly unmeasured in our dataset, although the use of therapeutic agents is likely to have contributed to the reduction in hospital fatality risk, particularly at the start of the pandemic [7, 25] . The measure of hospital burden we used considered acute hospital admissions at and around the time of admission as a proxy for bed occupancy. Whilst no single accepted measure of hospital burden exists, overnight bed occupancy is a widely used metric [11] , and guidance on bed occupancy was issued to ICUs (e.g. alterations in practice upon reaching 150% and 200% above pre-pandemic baseline) [26] . A limitation of the bed occupancy measure is that it only measures demand and not supply (i.e. staffing levels), or the extent of other hospital pressures. Work to access and integrate measures of supply is ongoing. Lastly, this study did not consider the significant proportion of patients (up to 40%) who may have acquired COVID-19 nosocomially (in hospital). Fatality risks and lengths of hospital stay for these patients are complicated by other conditions. Whilst these patients were excluded from our estimates, researchers in Scotland have found similar effects of age, sex, and comorbidity upon patient prognosis following nosocomial COVID-19 acquisition [27] . Case-mix, vaccination, and changes in hospital load continue to impact upon patient outcomes and lengths of stay more than 18 months after the pandemic began in England. One of the primary goals of the lockdown measures implemented in England at various times since start of the pandemic has been to protect against hospitals becoming excessively overburdened. Even with these measures in place, being admitted during a period of high hospital load was correlated with poorer outcomes. Meanwhile, vaccinated individuals admitted to hospital for COVID-19 had a significantly reduced risk of mortality, and third (booster) doses may further reduce this risk. Outcomes following hospitalisation with COVID-19 should continue to be monitored, particularly with the emergence of new variants. The datasets and methods we describe will be vital to estimate future changes in severity. We gratefully acknowledge all the clinicians, data reporters and patients whose data were used in this study, as well as all UK Health Security Agency colleagues involved in the COVID-19 response. We thank Shaun Seaman 10 .5% (10.0 -11.1%) 7.6 (7.5 -7.8) 3.6 (3.5 -3.6) *Median length of stay is a weighted median estimate with weighted ties; when two values satisfy the weighted median requirement, the estimate is the weighted average of the two. Epidemic phase bias may occur if there exists a relationship between the time from infection to symptom onset, and an individual's eventual outcome. E.g. if those who go on to die experience more rapid onset of symptoms following infection. To correct for this bias, a time shift of c days should be added to records with the outcome of interest (e.g. mortality), where c is the mean difference in time from infection to symptom onset date between those experiencing the outcome and those not, as proposed by Seaman et al. [1] . As the value of c is typically unknown, sensitivity analysis with differing values of c can be used to assess the susceptibility of results to this bias. For the sensitivity analysis in this study we shifted the date of symptom onset backwards in time by c = 0,1,2,3,4 days for those who died, to mitigate against the effect of more rapid symptom onset for those with more severe illness, where the shift c represents the average difference in time from infection to symptom onset between those patients who died and those who did not. The effect of this shift is shown in Supplementary Figure 7 . The Aalen-Johansen estimator is the standard non-parametric estimate of the cumulative incidence function for competing risk [2] , also described as the matrix version of the Kaplan-Meier estimator. Let the transition hazard from state ∈ to state ∈ S, ≠ be defined as: The Fine-Grey model estimates the hazard of a competing event (so-termed the subdistribution hazard) among the risk set of those yet to experience an event of the type of interest by time t [3] . The risk set therefore consists of both those who have yet to experience any event; and those who have yet to experience the event of interest (e.g. death) but have experienced a competing event (e.g. discharge). The subdistribution hazard is defined as the instantaneous risk of dying (from a cause ) given that the individual has not already died: Covariate effects on the sub-distribution hazard can then be interpreted as covariate effects on the cumulative incidence, or marginal probability, of a competing event (in this case hospitalised fatality). Stratified survival analyses enable appropriate adjustment for important confounders, by allowing the baseline hazard to vary across strata [4] . Stratification is a similar principle to matched designs, except rather than a 1:n ratio of cases to controls, as many (a:b for a and b both ≥ 1) cases and controls as possible within each strata are used. For the regression on month of admission, stratification was by age group, region of residence and vaccination status, with regression adjustment (main effects) on sex, ethnicity, IMD quintile and CCI. For the regression on vaccination status, stratification was by age group, region of residence and month of hospital admission, with regression adjustment (main effects) similarly on sex, ethnicity, IMD quintile and CCI. Supplementary figure 8 demonstrates the high degree of agreement between the Aalen-Johansen and Fine-Grey model estimates. Prognostic Factors for 30-Day Mortality in Critically Ill Patients With Coronavirus Disease 2019: An Observational Cohort Study Trends in risks of severe events and lengths of stay for COVID-19 hospitalisations in England over the pre-vaccination era: results from the Public Health England SARI-Watch surveillance scheme Patient factors and temporal trends associated with COVID-19 in-hospital mortality in England: an observational study using administrative data Association between multimorbidity and mortality in a cohort of patients admitted to hospital with COVID-19 in Scotland Effectiveness of the Pfizer-BioNTech and Oxford-AstraZeneca vaccines on covid-19 related symptoms, hospital admissions, and mortality in older adults in England: test negative case-control study Direct and Indirect Impact of the Vaccination Programme on COVID-19 Infections and Mortality Changes in UK hospital mortality in the first wave of COVID-19: the ISARIC WHO Clinical Characterisation Protocol prospective multicentre observational cohort study. bioRxiv. medRxiv ICNARC report on COVID-19 in critical care: England, Wales and Northern Ireland Understanding COVID-19 trajectories from a nationwide linked electronic health record cohort of 56 million people: phenotypes, severity, waves & vaccination Mortality among people hospitalised with covid-19 in Switzerland: a nationwide population-based analysis NHS hospital bed numbers past, present, future Survival Analysis: A Practical Approach Statistical Analysis of Clinical COVID-19 Data: A Concise Overview of Lessons Learned, Common Errors and How to Avoid Them An Empirical Transition Matrix for Non-Homogeneous Markov Chains Based on Censored Observations A proportional hazards model for the subdistribution of a competing risk Adjusting for time of infection or positive test when estimating the risk of a post-infection outcome in an epidemic A new method of classifying prognostic comorbidity in longitudinal studies: development and validation A comparison of the recording of comorbidity in primary and secondary care by using the Charlson Index to predict short-term and long-term survival in a routine linked data cohort COVID-19: review of disparities in risks and outcomes Ethnic differences in SARS-CoV-2 infection and COVID-19-related hospitalisation, intensive care unit admission, and death in 17 million adults in England: an observational cohort study using the OpenSAFELY platform Changes in COVID-19 inhospital mortality in hospitalised adults in England over the first seven months of the pandemic: An observational study using administrative data COVID-19 vaccine weekly surveillance reports (weeks 39 to 4, 2021 to 2022) Mortality caused by influenza and respiratory syncytial virus by age group in England and Wales 1999-2010. Influenza Other Respi Viruses The burden of influenza in England by age and clinical risk group: a statistical analysis to inform vaccine policy Remdesivir for the Treatment of Covid-19 -Preliminary Report Advice on acute sector workforce models during COVID-19 Does nosocomial COVID-19 result in increased 30-day mortality? A multi-centre observational study to identify risk factors for worse outcomes in patients with COVID-19 5%) Region of residence London 5%) 733,250 (11.1%) Yorkshire and Humber 25,892 (10%) 723,861 (10.9%) Index of multiple deprivation 1st quintile Month of hospital admission .9%) .9%) Oct-20 15,235 (5.9%) 474,083 (7.2%) 4%) Adjusting for time of infection or positive test when estimating the risk of a post-infection outcome in an epidemic An Empirical Transition Matrix for Non-Homogeneous Markov Chains Based on Censored Observations A proportional hazards model for the subdistribution of a competing All tables in Supplementary Excel workbook.