Comparison of COVID-19 outcomes among shielded and non-shielded populations

Many western countries used shielding (extended self-isolation) of people presumed to be at high-risk from COVID-19 to protect them and reduce healthcare demand. To investigate the effectiveness of this strategy, we linked family practitioner, prescribing, laboratory, hospital and death records and compared COVID-19 outcomes among shielded and non-shielded individuals in the West of Scotland. Of the 1.3 million population, 27,747 (2.03%) were advised to shield, and 353,085 (26.85%) were classified a priori as moderate risk. COVID-19 testing was more common in the shielded (7.01%) and moderate risk (2.03%) groups, than low risk (0.73%). Referent to low-risk, the shielded group had higher confirmed infections (RR 8.45, 95% 7.44–9.59), case-fatality (RR 5.62, 95% CI 4.47–7.07) and population mortality (RR 57.56, 95% 44.06–75.19). The moderate-risk had intermediate confirmed infections (RR 4.11, 95% CI 3.82–4.42) and population mortality (RR 25.41, 95% CI 20.36–31.71) but, due to their higher prevalence, made the largest contribution to deaths (PAF 75.30%). Age ≥ 70 years accounted for 49.55% of deaths. In conclusion, in spite of the shielding strategy, high risk individuals were at increased risk of death. Furthermore, to be effective as a population strategy, shielding criteria would have needed to be widely expanded to include other criteria, such as the elderly.

The likelihood of being tested increased with age, was higher in women and the moderate-risk category and highest in the shielded group (Table 1). Overall, 3348 (0.25%) people had confirmed COVID-19 infection. The likelihood of laboratory-confirmed COVID-19 infection followed similar patterns as testing. It increased with age, was higher in women, was highest in the shielded group and lowest in the low-risk category ( Table 2). After adjustment for sex and deprivation quintile, the risk of laboratory-confirmed infection remained higher in the moderate-risk category and highest in the shielded group ( Fig. 1 and Supplementary Table S2).
Overall, 1661 people were hospitalised for COVID-19. Within the general population, hospitalisations increased with age but were comparable between men and women ( Table 2). Hospitalisations were more common in the moderate-risk category and most common in the shielded group (Table 2), remaining so after adjustment for sex and deprivation ( Fig. 1 and Supplementary Table S2). Overall, 122 people were admitted to ICU wards for COVID-19. ICU admissions were significantly more common among people aged 45-64 years of age than among older people ( Table 2). Compared with the low-risk category, the shielded group were 18 times more likely to be hospitalised but only 4 times more likely to be admitted to ICU ( Fig. 1 and Supplementary Table S2). Overall, 1027 (0.08%) people died from COVID-19. Within the general population, mortality increased with age but was similar in men and women (Table 2). Population mortality was higher in the moderate-risk category and highest in the shielded group (Table 2) and remained so after adjustment for sex and deprivation ( Fig. 1 and Supplementary Table S2). Mortality in the community before/without admission to hospital was higher in the moderate-risk category (RR moderate/low [95% CI] 79.7 [39.9-159.0], p < 0.0001) and highest in the shielded group (RR shielded/low [95% CI] 182. 1 [86.5-383.3], p < 0.0001). A similar pattern was observed for mortality before/ without ICU admission (RR moderate/low [95% CI] 40.8 [29.0-57.4], p < 0.0001; RR shielded/low [95% CI] 120. 8 [83.0, 175.9], p < 0.0001).
Among the sub-group with laboratory-confirmed (test-positive) COVID-19 infection, 1661 (49.6%) were hospitalised. Hospitalisations increased with age but were comparable between men and women ( Table 3). The moderate-risk category was more likely to be hospitalised and the shielded group most likely (Table 3), remaining so after adjustment for age and deprivation ( Fig. 2 and Supplementary Table S3). Among those with laboratoryconfirmed infection, ICU admissions were more common in men and more common in people aged 45-64 years than those older (Table 3). Low-risk cases were more likely to be admitted to ICU than the moderate-risk and shielded groups (Fig. 2). Among the sub-group with clinically-confirmed (test-positive or COVID-19 related death) COVID-19 infection, 1027 (26.70%) died (Table 3). Case-fatality increased by age and was higher in men than women. It was lowest in the low-risk category but not significantly different between the moderate-risk and shielded groups (RR shielded/moderate [95% CI] 1.12 [0.96-1.31], p = 0.14) ( Fig. 2 Figure S1).
Individual risk criteria. Due to insufficient numbers, the individual risk criteria models could not be run for pregnant women with severe heart disease or for COVID-19 related ICU admission in the shielded category. All the remaining individual risk criteria were associated with higher likelihood of being tested for COVID-19 (Table 1), laboratory-confirmed infection (Table 2), hospitalisation, population mortality ( Fig. 1 and Supplementary Table S2) and case-fatality ( Fig. 2

Discussion
The 2.03% of people advised to shield were, nonetheless, eight times more likely to have confirmed infections than the low-risk category, five times more likely to die following confirmed infection and 49 times more likely to die from COVID-19 overall. Whilst selective testing might explain the first outcome, it does not explain higher overall mortality which suggests that the shielding strategy was not as effective as was hoped. www.nature.com/scientificreports/ One quarter of the population were classified as moderate-risk and not advised to shield. Nonetheless, they were four times more likely to have confirmed infections than the low-risk category, five times more likely to die following confirmed infection and 25 times more likely to die overall, suggesting that, where shielding is employed, the shielding criteria should be expanded. In particular, older age needs to be considered since the elderly are both at high individual risk and contribute significantly to population burden due to their relatively high numbers. A survey of 1695 rheumatology patients reported that the shielded sub-group has slightly lower risk of COVID-19 infection than the non-shielded sub-group (3.0% vs 4.1%) but the difference did not reach statistical significance 21 . In spite of people in the shielded and moderate-risk categories having poorer prognosis, they were less likely to be admitted to ICU following hospitalisation for COVID-19, especially patients ≥ 70 years. This finding reinforces the importance of protection in those with the worst prognosis.
Our finding that 26.85% of people satisfied the moderate-risk criteria is consistent with limited existing evidence. A study linking English primary and secondary care records on 3.9 million people reported that 20% of population satisfied similar criteria 22 . Similarly, analysis of the Global Burden of Diseases Study estimated that 22% of the global population are at increased risk of severe COVID-19 disease 19 . A USA study using data from the Behavioral Risk Factor Surveillance System reported that 45.4% of 444,649 adults had one or more of a longer list of morbidities that may be associated with higher risk from COVID-19 23 . Another USA study estimated that 14.2% of participants in the National Health Interview Survey had more than two-fold risk and 1.6% had more than tenfold risk 24 .
The evidence on COVID-19 related complications among those classified as high risk, and therefore advised to shield, has mainly come from case series and expert opinion. Case series found higher COVID-19 related complications among organ transplant recipients 25,26 , patients receiving chemotherapy, radiotherapy or immunotherapy for cancer 27,28 , and patients with haematological cancers 29 . Systematic review suggested higher COVID-19 complication risk among COPD patients, but the effect of COPD severity was not investigated 30 . Patients with cystic fibrosis and sickle cell disease were classified as high risk based on expert opinion 31,32 . While pregnant women with COVID-19 were found to have higher risk of poor maternal and perinatal outcomes 33,34 , outcomes were not investigated specifically for pregnant women with heart disease. There was no evidence of worse COVID-19 related complications among patients on immunosuppressants 35 . A large community study in England found strong association between severe asthma (hazard ratio 1.25) and COVID-19 related mortality but did not investigate the risk of COVID-19 infection or hospitalisation 36 .
In common with previous studies 37 , we demonstrated that age was a major individual-level risk factor for death. Additionally, we showed it is important at the population level with 49.55% of deaths attributable to age ≥ 70 years. The higher mortality in the elderly was mediated in part by higher case-fatality but they also had a higher incidence of infection, possibly due to transmission within care homes. Lower ICU admissions following hospitalisation for COVID-19 may have contributed to their higher case-fatality. Previous studies have reported that men are at higher risk of COVID-19 7 . Our study demonstrated they are less likely to be tested for COVID-19, have confirmed infection, and be hospitalised. They have comparable overall mortality from COVID-19, due to their lower incidence, but their case-fatality is higher. www.nature.com/scientificreports/ This study adds to the existing evidence of the possible effectiveness of a shielding strategy which is largely limited to mathematical modelling of population effects based on assumptions [12][13][14][15][16][17][18][19][20] . The modelling papers concluded that shielding could potentially reduce deaths and demand for ICU beds, subject to a range of parameters including the transmission rate, proportion shielding, and baseline case-fatality rate, but could not reduce to reproduction rate to below 1 in the absence of other non-pharmaceutical interventions. In Scotland, prior to the introduction of shielding, those individuals subsequently advised to shield accounted for 40% of COVID-19 deaths 8 . However, an unpublished study reported that, following the introduction of shielding, the rate of decline in COVID-19 incidence in the shielding group was comparable to the general population 38 , suggesting www.nature.com/scientificreports/ that the decline was most likely due to the suite of non-pharmaceutical interventions introduced at that time, rather than shielding specifically 8,38 .
Ours was a large-scale, unselected general population study. The data cover a period when shielding was in place. Linkage of family practitioner, laboratory, hospital and death data enabled us to examine a range of COVID-19 outcomes and study a range of exposure variables including the overall risk categories and their individual criteria. The datasets were linked using exact, rather than probabilistic, matching. We were able to adjust for potential sociodemographic confounders. This is important since, in spite of eligibility for Statutory Sick Pay, the financial impact of shielding was greater among people on short, fixed term or zero hours contracts, in self-employment, or unable to work from home, and compliance with shielding was worse in more deprived households 39 . The exposure data were collected prior to the outcomes occurring avoiding potential reverse causation and recall or recording bias. Our analysis of potential risk factors was restricted to those used as criteria for shielding and moderate-risk at the time of the study. We did not have data on ethnicity and rurality and there may be other unmeasured confounders such as exposure to health and social care workers and compliance with restrictions 20,40 . The list of people meeting the shielding and moderate-risk criteria was extracted at the study mid-point; therefore, some people may have been misclassified prior to and after this date. This is a pragmatic study that evaluates the effectiveness of shielding as delivered and supported in Scotland over the first wave of COVID-19. Shielding was recommended but not monitored or enforced. Forty one percent of people advised to shield in Scotland reported stringently following shielding guidance and 21% reported they were unable to comply for a variety of reasons such as supporting other household members (e.g. only driver), caring for pets, avoiding domestic abuse, or undertaking essential chores 39 . Many also reported difficulties social distancing within their home due to shared facilities or carer roles. Our findings are representative of Glasgow and Greater Clyde area but may be less so for other areas or countries. In particular, the effectiveness of shielding may be different in countries with different levels of support, monitoring or compliance.
Our findings suggest that attempts to shield those at highest risk have not been as successful as hoped, with those advised to shield experiencing higher rates of infection and death. Since this group was also less likely to be admitted to ICU, protecting them from infection is essential. For shielding to be effective as a population level strategy, the current criteria would need to be expanded, since three-quarters of deaths were associated with moderate-risk criteria for which shielding had not hitherto been recommended. In our study, more than one-quarter of the general population would have needed to be effectively shielded to prevent over 80% of deaths. Since this is unlikely to be acceptable at a time when governments are under pressure to avoid further lock-downs, shielding is probably best viewed as an intervention to protect individuals, to be used alongside other population-wide interventions such as physical distancing, face coverings and hand hygiene.

Methods
This was a general population cohort study of all 1.3 million residents of NHS GGC in the West of Scotland between March and May 2020. In Scotland, shielding was recommended from the start of pandemic, throughout the whole study period, until end of July 2020. The Community Health Index (CHI), a unique identifier attached to all Scottish health records, enabled individual-level record linkage of nine databases: Community Health Index (CHI) register, NHS GGC Shielding List, Egton Medical Information Systems (EMIS) and Vision, Electronic Communication of Surveillance in Scotland (ECOSS), Prescribing Information System (PIS), Strathclyde Electronic Renal Patient Record (SERPR), Rapid Preliminary Inpatient Data (RAPID), and death certificates.
The CHI register provided sociodemographic information (age, sex, area socioeconomic deprivation). Deprivation was measured using the Scottish Index of Multiple Deprivation (SIMD), derived from seven domainsincome, education, health, employment, crime, housing, and access to services-and categorised into general population quintiles. ECOSS collects laboratory data on infectious diseases, including test date and result. Albasoft software extract data from the family practitioner electronic health record systems EMIS and Vision, and PIS collects data on medications prescribed by family practitioners. SERPR records data on renal replacement therapy and transplantation. RAPID collects real-time data on hospitalisation, including dates of admission and discharge, and type of ward, and the Scottish Morbidity Record 01 (SMR01) subsequently records the relevant disease codes. Death certificates provide the date and cause of all deaths, whether in-hospital or in the www.nature.com/scientificreports/ community. Follow-up data were available until the end of May 2020, before the shielding recommendation was lifted. Supplementary Table S1 lists the criteria for the shielded and medium risk categories applied at the time of data extraction. All remaining patients were categorised as low-risk. The Scottish list of high-risk individuals is compiled centrally, and regularly updated, using family practitioner, hospital admission, disease registry and medication data. Family practitioners check the completeness and accuracy of the list before letters, recommending shielding, are sent to patients. The NHS GGC Shielding List we used contains the validated data including the criterion satisfied. We ascertained moderate risk individuals using Albasoft extraction of EMIS and Vision data, and PIS data.
Separate models were conducted by overall risk category (low-risk, moderate-risk or shielded) and by the individual criteria for the moderate-risk and shielded categories. The four general population outcomes investigated were: confirmed COVID-19 infection; COVID-19 related hospitalisation; COVID-19 related ICU admission; COVID-19 related mortality. The three outcomes investigated among those with confirmed infection were: COVID-19 related hospitalisation; COVID-19 related ICU admission; and COVID-19 related case fatality.
Laboratory-confirmed cases were defined as positive PCR test. Clinically-confirmed cases were defined as either positive PCR test or death from COVID-19 without testing. COVID-19 related deaths were defined as International Classification of Diseases 10th revision (ICD-10) code U07.1 or U07.2 recorded on the death certificate. COVID-related hospitalisation was defined as an SMR01 hospitalisation record with an ICD code U07.1 or U07.2 or, for more recent admissions, a RAPID hospitalisation record plus positive PCR test taken between two weeks before and two days after hospitalisation. ICU admission during such hospitalisations was assumed to be COVID-related.
Sociodemographic characteristics were compared by risk category using chi-square tests. Poisson regression models with robust standard errors were used to compare risk ratios (RR) for the shielded and moderate-risk categories referent to the low-risk category. Applying Poisson models to binary outcomes can be subject to an under-dispersion problem. This was overcome by deriving robust 'sandwich' standard errors 41 . The models were run univariately; then adjusted for sex, SIMD quintile, and other risk categories as potential confounders. Age was not included as a covariate because it was a moderate-risk criterion. The models were re-run using the individual criteria for the shielded and moderate-risk categories as the exposure variables, referent to the low-risk category.
Population attributable fractions (PAFs) 42 were calculated, from prevalence and adjusted RR, to determine the proportion of each outcome that could be attributed to being shielded and moderate-risk, as well as the proportion due to each individual criterion. The PAFs of individual criteria were proportionally calibrated so that their sum equated to the overall PAF of the relevant risk category. PAF confidence intervals were estimated using bootstrapping (× 1000) taking account of the variance for prevalence and RRs.
Ethical approvals. The study was approved by the NHS GGC Primary Care Information Sharing Group and the NHS GGC Local Privacy Advisory Committee (Reference GSH/20RM005) and was covered by the generic Safe Haven Research Ethics Committee approval (GSH20RM005_COVID_Community). Explicit consent was not feasible to obtain as the data custodians provided us with an anonymised extract of secondary data. Requirement for consent was waived by the Safe Haven Research Ethics Committee, as part of the study approval, on the basis of adequate data protection safeguards including: anonymisation, analysis with a safe haven environment, information governance training, disclosure control and signed agreements. The study has been performed in accordance with the Declaration of Helsinki.