Living risk prediction algorithm (QCOVID) for risk of hospital admission and mortality from coronavirus 19 in adults: national derivation and validation cohort study

1The first cases of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection were reported in the UK on 24 January 2020, with the first death from coronavirus disease 2019 (covid-19) on 28 February 2020. As of 18 August 2020, more than 41 000 deaths from covid-19 had occurred in the UK and more than 773 000 deaths globally. In the initial absence of any vaccination or prophylactic or curative treatments, the UK government implemented social distancing and shielding measures to suppress the rate of infection and protect vulnerable people, thereby trying to minimise the risk of serious adverse outcomes.

2Emerging evidence throughout the course of the pandemic, initially from case series and then from cohorts of patients with confirmed SARS-CoV-2 infection, has shown associations of age, sex, certain comorbidities, ethnicity, and obesity with adverse covid-19 outcomes such as hospital admission or death. The knowledge base regarding risk factors for severe covid-19 is growing. As many countries are cautiously attempting to ease “lockdown” measures or reintroduce measures if rates are rising, an opportunity exists to develop more nuanced guidance based on predictive algorithms to inform risk management decisions. Better knowledge of individuals’ risks could also help to guide decisions on mitigating occupational exposure and in targeting of vaccines to those most at risk. Although some prediction models have been developed, a recent systematic review found that they all have a high risk of bias and that their reported performance is optimistic.

3The use of primary care datasets with linkage to registries such as death records, hospital admissions data, and covid-19 testing results represents a novel approach to clinical risk prediction modelling for covid-19. It provides accurately coded, individual level data for very large numbers of people representative of the national population. This approach draws on the rich phenotyping of individuals with demographic, medical, and pharmacological predictors to allow robust statistical modelling and evaluation. Such linked datasets have an established track record for the development and evaluation of established clinical risk models, including those for cardiovascular disease, diabetes, and mortality. We aimed to develop and validate population based prediction models to estimate the risks of becoming infected with and subsequently dying from covid-19 and of becoming infected and subsequently admitted to hospital with covid-19. The model we have developed is designed to be applied across the adult population so that it can be used to enable risk stratification for public health purposes in the event of a “second wave” of the pandemic, to support shared management of risk and occupational exposure, and in early targeting of vaccines to people most at risk. An ongoing companion study will externally validate the models, using datasets across all four nations of the UK, and will be reported separately.

4This study was commissioned by the Chief Medical Officer for England on behalf of the UK Government, who asked the New and Emerging Respiratory Virus Threats Advisory Group (NERVTAG) to establish whether a clinical risk prediction model for covid-19 could be developed in line with the emerging evidence. The protocol has been published. The study was conducted in adherence with TRIPOD and RECORD guidelines and with input from our patient advisory group.

5We did a cohort study of primary care patients using the QResearch database (version 44). QResearch was established in 2002 and has been extensively used for the development of risk prediction algorithms across the National Health Service (NHS) and for epidemiological research. By April 2020, 1205 practices in England were contributing to QResearch, covering a population of 10.5 million patients. The database is linked at individual patient level, using a project specific pseudonymised NHS number, to hospital admissions data (including intensive care unit data), positive results from covid-19 real time reverse transcriptase polymerase chain reaction tests held by Public Health England, cancer registrations (including detailed radiotherapy and systemic chemotherapy records), the national covid-19 shielded patient list in England, and mortality records held by NHS Digital.

6We identified a cohort of people aged 19-100 years registered with participating general practices in England on 24 January 2020. We excluded patients (approximately 0.1%) who did not have a valid NHS number. Patients entered the cohort on 24 January 2020 (date of first confirmed case of covid-19 in the UK) and were followed up until they had the outcome of interest or the end of the first study period (30 April 2020), which was the date up to which linked data were available at the time of the derivation of the model, or the second time period (1 May 2020 until 30 June 2020) for the temporal cohort validation.

7The primary outcome was time to death from covid-19 (either in hospital or outside hospital), defined as confirmed or suspected death from covid-19 as per the death certification or death occurring in an individual with confirmed SARS-CoV-2 infection at any time in the period 24 January to 30 April 2020. The secondary outcome was time to hospital admission with covid-19, defined as an ICD-10 (International Classification of Diseases, 10th revision) code for either confirmed or suspected covid-19 or new hospital admission associated with a confirmed SARS-CoV-2 infection in the study period.

8We selected candidate predictor variables on the basis of the presence of existing clinical vulnerability group criteria (table 1), associations with outcomes in other respiratory diseases, or hypothesised to be linked to adverse outcomes on clinical/biological plausibility and likely to be available for implementation. They are summarised in box 1 and supplementary box A. We defined variables according to information recorded using Read Codes in general practices’ electronic health records at the start of the study period. The exception to this was information on chemotherapy, radiotherapy, and transplants, which was based on linked hospital records.

9Age in years (continuous)

10Townsend deprivation score (continuous)—This is an area level continuous score based on the patient’s postcode. Originally developed by Townsend, it includes unemployment (as a percentage of those aged ≥16 who are economically active), non-car ownership (as a percentage of all households), non-home ownership (as a percentage of all households), and household overcrowding. These variables are measured for a given area of approximately 120 households, via the 2011 census, and combined to give a Townsend score for that area. A greater Townsend score implies a greater level of deprivation

11Ethnicity in nine categories (White, Indian, Pakistani, Bangladeshi, Other Asian, Caribbean, Black African, Chinese, other ethnic group)

12Domicile in three categories: homeless, care home residence (nursing or residential), other

13Smoking status in five categories (non-smoker, ex-smoker, 1-10 per day, 11-19 per day, ≥20 per day)

14Body mass index in kg/m2 (continuous)

15Crack cocaine and injected drug use

16Solid organ transplant recipient on long term immune suppression treatment

17Cancers:

18Active chemotherapy

19Radical radiotherapy for lung cancer

20Blood/bone marrow cancer at any treatment stage

21Immunotherapy or continuing antibody treatment

22Targeted cancer treatments that affect immune system (PARP inhibitor or PKI)

23Bone marrow or stem cell transplant in previous 6 months or remain on immunosuppression

24Immunosuppression sufficiently increasing infection risk

25Severe respiratory disease:

26Severe asthma (≥3 prescribed courses of steroids in preceding 12 months)

27Severe COPD (≥3 prescribed courses of steroids in preceding 12 months)

28Cystic fibrosis, interstitial lung disease, sarcoidosis, non-cystic fibrosis bronchiectasis, or pulmonary hypertension

29Rare diseases or inborn errors of metabolism:

30Severe combined immunodeficiency

31Homozygous sickle cell disease

32Pregnant with significant heart disease:

33Acquired or congenital

34Chronic, non-severe respiratory disease:

35Asthma

36COPD (emphysema and chronic bronchitis)

37Extrinsic allergic alveolitis

38Chronic kidney disease (CKD):

39CKD stage 3 or 4

40End stage renal failure requiring dialysis

41End stage renal failure ever undergoing a transplant

42Cardiac disease:

43Congestive cardiac failure

44Valvular heart disease

45Chronic liver disease:

46Chronic infective hepatitis (hepatitis B or C)

47Alcohol related liver disease

48Primary biliary cirrhosis

49Primary sclerosing cholangitis

50Haemochromatosis

51Chronic neurological conditions:

52Epilepsy

53Parkinson’s disease

54Motor neurone disease

55Cerebral palsy

56Dementia (Alzheimer’s, vascular, frontotemporal)

57Down’s syndrome

58Diabetes mellitus:

59Type 1

60Type 2

61Conditions or treatments that predispose to infection (eg, steroid treatment):

62Rheumatoid arthritis

63Systemic lupus erythematosus

64Ankylosing spondylitis or other inflammatory arthropathy (eg, psoriatic arthritis)

65Connective tissue disease (eg, Ehlers-Danlos syndrome, scleroderma, Sjögren’s syndrome)

66Polymyositis or dermatomyositis

67Vasculitis (eg, giant cell arteritis, polyarteritis nodosa, Behçet’s syndrome)

68Cardiovascular disease:

69Atrial fibrillation

70Cardiovascular events (myocardial infarction, stroke, angina, transient ischaemic attack)

71Peripheral vascular disease

72Treated hypertension

73Hyperthyroidism

74Chronic pancreatitis

75Cirrhosis (if not above; eg, non-alcoholic fatty liver disease)

76Malabsorption:

77Coeliac disease

78Steatorrhoea

79Blind loop syndrome

80Peptic ulcer (gastric or duodenal)

81Learning disability

82Osteoporosis

83Fragility fracture (hip, spine, shoulder, or wrist fracture)

84Severe mental illness:

85Bipolar affective disorder

86Psychosis

87Schizophrenia or schizoaffective disorder

88Severe depression

89HIV infection

90Hyposplenism

91Sickle cell disease

92Sphingolipidoses (eg, Tay-Sachs disease)

93History of venous thromboembolism

94Tuberculosis

95Drugs affecting the immune response, including systemic chemotherapy based on hospital data

96Drugs affecting the immune system prescribed in primary care (focus on BNF chapter 8.2)

97Long acting β agonists

98Long acting muscarinic antagonists

99Inhaled corticosteroids

100We randomly allocated 75% of practices to the derivation dataset, which we used to develop the models. We evaluated the models’ performance in the remaining 25% of practices (the validation set). All models were fitted separately in men and women. The outcomes of interest are subject to competing risks. For the primary outcome of death from covid-19, the competing risk is death due to other causes. For the secondary outcome of hospital admission, the competing risk is death from any cause before admission. We fitted a sub-distribution hazard (Fine and Gray) model for each outcome to account for competing risks. Individuals who did not have the outcome of interest were censored at the study end date, including those who had a competing event.

101For all predictor variables, we used the most recently available value at the entry date (24 January 2020). We used second degree fractional polynomials to model non-linear relations for continuous variables (age, body mass index, and Townsend material deprivation score, an area level score based on postcode). Initially, we fitted a complete case analysis by using a model within the derivation data to derive the fractional polynomial terms. For indicators of comorbidities and medication use, we assumed the absence of recorded information to mean absence of the factor in question. Data were missing in four variables: ethnicity, Townsend score, body mass index, and smoking status. We used multiple imputation with chained equations under the missing at random assumption to replace missing values for these variables. For computational efficiency, we used a combined imputation model for both outcomes. The imputation model was fitted in the derivation data and included predictor variables, the Nelson-Aalen estimators of the baseline cumulative sub-distribution hazard, and the outcome indicators (death from covid-19 and hospital admission with covid-19). We carried out five imputations. Each analysis model was fitted in each of the five imputed datasets. We used Rubin’s rules to combine the model parameter estimates and the baseline cumulative incidence estimates across the imputed datasets.

102We initially sought to fit models using all predictor variables. Owing to sparse cells, some conditions were combined if clinically similar in nature (such as rare neurological disorders). We examined interactions between body mass index and ethnicity and interactions between predictor variables and age, focusing on predictor variables that apply across the age range (asthma, epilepsy, diabetes, severe mental illness). We explored the use of penalised models (LASSO) to screen variables for inclusion, but this retained all the predictor variables and most interaction terms. In line with the protocol, we subsequently removed a small number of variables with low numbers of events and adjusted (sub-distribution) hazard ratios close to 1 (as these will have minimal effect on predicted risks) or with uncertain clinical credibility, defined as counterintuitive results in light of the emerging literature. Lastly, we combined regression coefficients from the final models with estimates of the baseline cumulative incidence function evaluated at 97 days to derive risk equations for each outcome. We used all the available data in the database.

103We did all model evaluation using the validation data with two separate periods of follow-up. The first validation study period was the same as for the derivation cohort: 24 January to 30 April 2020. The second temporal validation covered the subsequent period of 1 May 2020 to 30 June 2020. This was carried out with the same validation cohort except for exclusion of patients who died during 24 January to 30 April 2020. In the validation cohort, we fitted an imputation model to replace missing values for ethnicity, body mass index, Townsend score, and smoking status. This excluded the outcome indicators and Nelson-Aalen terms, as the aim was to use covariate data to obtain a prediction as if the outcome had not been observed to reflect intended use.

104We applied the final risk equations developed from the derivation dataset to men and women in the validation dataset and evaluated R2 values, Brier scores, and measures of discrimination and calibration for the two time periods. R2 values refer to the proportion of variation in survival time explained by the model. Brier scores measure predictive accuracy, where lower values indicate better accuracy. D statistics (a discrimination measure that quantifies the separation in survival between patients with different levels of predicted risks) and Harrell’s C statistics (a discrimination metric that quantifies the extent to which people with higher risk scores have earlier events) were evaluated at 97 days (the maximum follow-up period available at the time of the derivation of the model) and 60 days for the second temporal validation, with corresponding 95% confidence intervals. We assessed model calibration by comparing mean predicted risks with observed risks by twentieths of predicted risk for each of the validation cohorts. Observed risks were derived in each of the 20 groups by using non-parametric estimates of the cumulative incidences. Additionally, we did a recalibration for the mortality outcome, using the method proposed by Booth et al by updating the baseline survivor function based on the temporal validation cohort with the prognostic index as an offset term. We also applied the algorithms to the validation cohort for the first time period to define the centile thresholds based on absolute risk. We also defined centiles of relative risk (defined as the ratio of the individual’s predicted absolute risk to the predicted absolute risk for a person of the same age and sex with a white ethnicity, body mass index of 25, and mean deprivation score with no other risk factors).

105We calculated the performance metrics in the whole validation cohort and in the following pre-specified subgroups: within age groups (19-39, 40-49, 50-59, 60-69, 70-79, ≥80 years), within nine ethnic groups, and within each of the 10 English regions (where numbers allowed). In response to open peer review of the published protocol, we evaluated performance by calculating Harrell’s C statistics in individual general practices and combining the results using a random effects meta-analysis.

106Patients were involved in setting the research question and in developing plans for design and implementation of the study. Patients were asked to aid in interpreting and disseminating the results.

107Overall, 1205 practices in England met our inclusion criteria. Of these, 910 practices were randomly assigned to the derivation dataset and 295 to the validation cohort. The practices had 8 256 158 registered patients aged 19-100 years on 24 January 2020. We included 6 083 102 of these in the derivation cohort, and the validation dataset comprised 2 173 056 people.

108Table 2 shows the baseline characteristics of patients in the derivation cohort. Of these patients, 3 035 409 (49.9%) were men and 990 799 (16.3%) were of black, Asian, or other minority ethnic (BAME) background.

109In the derivation cohort, 10 776 (0.18%) patients had a covid-19 related hospital admission and 4384 (0.07%) had a covid-19 related death during the 97 days’ follow-up, of which 4265 (97.3%) were recorded on the death certificate and 119 (2.71%) were based only on a positive test (and of these <15 were based on a test more than 28 days before death). Admissions and deaths due to covid-19 occurred across all regions, with the greatest numbers in London, which accounted for 3799 (35.3%) of admissions and 1287 (29.4%) of deaths. Of those who died, 2517 (57.4%) were male, 732 (16.7%) were BAME, 3616 (82.5%) were aged 70 and over, 1417 (32.3%) had type 2 diabetes, 1311 (29.9%) had dementia, and 1033 (23.6%) were identified as living in a care home.

110The characteristics of the validation cohort were similar to those of the derivation cohort, as shown in supplementary tables A and B. In the first validation period (24 January to 30 April 2020), 1722 deaths and 3703 hospital admissions due to covid-19 occurred. In the second validation period (1 May to 30 June 2020), 621 deaths and 1002 admissions due to covid-19 occurred.

111The variables included in the final models were fractional polynomial terms for age and body mass index, Townsend score (linear), ethnic group, domicile (residential care, homeless, neither), and a range of conditions and treatments as shown in figure 1, figure 2, figure 3, and figure 4. These conditions and treatments were cardiovascular conditions (atrial fibrillation, heart failure, stroke, peripheral vascular disease, coronary heart disease, congenital heart disease), diabetes (type 1 and type 2 and interaction terms for type 2 diabetes with age), respiratory conditions (asthma, rare respiratory conditions (cystic fibrosis, bronchiectasis, or alveolitis), chronic obstructive pulmonary disease, pulmonary hypertension or pulmonary fibrosis), cancer (blood cancer, chemotherapy, lung or oral cancer, marrow transplant, radiotherapy), neurological conditions (cerebral palsy, Parkinson’s disease, rare neurological conditions (motor neurone disease, multiple sclerosis, myasthenia, Huntington’s chorea), epilepsy, dementia, learning disability, severe mental illness), other conditions (liver cirrhosis, osteoporotic fracture, rheumatoid arthritis or systemic lupus erythematosus, sickle cell disease, venous thromboembolism, solid organ transplant, renal failure (CKD3, CKD4, CKD5, with or without dialysis or transplant)), and medications (≥4 prescriptions from general practitioner in previous six months for oral steroids, long acting β agonists or leukotrienes, immunosuppressants).

112Figure 1 and figure 2 show the adjusted hazard ratios in the final models for covid-19 related death in the derivation cohort in women and men. Figure 3 and figure 4 show the adjusted hazard ratios for the final models for covid-19 related hospital admission in the derivation cohort.

113Supplementary figures A and B show graphs of the adjusted hazard ratios for body mass index, age, and the interaction between age and type 2 diabetes for deaths and hospital admissions due to covid-19 (which showed higher risks associated with younger ages). Supplementary figures C and D show fully adjusted hazard ratios for variables for the full model, including variables that were not retained in the final model (for example, adjusted hazard ratios close to one or those which lacked clinical credibility). Other variables with too few events for inclusion were HIV, sphingolipidoises, short bowel syndrome, polymyositis, dermatomyositis, Ehlers-Danlos syndrome, biliary cirrhosis, hepatitis B and C, haemochromatosis, non-alcoholic fatty liver disease, chronic pancreatitis, drug misuse, asplenia, cholangitis, scleroderma, Sjogren’s syndrome, and pregnancy. Supplementary figures E and F show fully adjusted hazard ratios for a combined outcome of either covid-19 related death or hospital admission. This gave very similar absolute risks to the hospital admission outcome.

114Table 3 shows the performance of the risk equations in the validation cohort for women and men over 97 days for the main study period and for the temporal validation cohort evaluated from 1 May 2020 to 30 June 2020. Overall, the values for the R2, D, and C statistics were similar in women and men. Values for the mortality outcome tended to be higher than those for the hospital admission outcome. For example, in the first validation period, the equation explained 74% of the variation in time to death from covid-19 in women; the D statistic was 3.46, and Harrell’s C statistic was 0.933. The corresponding values in men were 73.1%, 3.37, and 0.928. The results for the second validation period were similar except for covid-19 related admissions in women, for which the explained variation and discrimination were lower than for the first period (explained variation 45.4%, D statistic 1.87, and Harrell’s C statistic 0.776).

115Supplementary tables C-F show the corresponding results by region, age band, and fifth of deprivation and within each ethnic group in men and women in both validation periods. Performance was generally similar to the overall results except for age, for which the values were lower within individual age bands.

116Figure 5 shows funnel plots of Harrell’s C statistic for each general practice in the validation cohort versus the number of deaths in each practice in men and women in the first validation period. The summary (average) C statistic for women was 0.916 (95% confidence interval 0.908 to 0.924) from a random effects meta-analysis. The corresponding summary C statistic for men was 0.919 (0.912 to 0.926).

117Figure 6 (top two rows) shows the calibration plots for the covid-19 related hospital admission equation and for the covid-19 related death equation in the first validation period. These show that both sets of equations were well calibrated in the first time period except for a small degree of under-prediction in the highest risk category for mortality. In the second validation period, calibration was good for the hospital admission outcome but not for the mortality outcome at the high levels of risk (fig 6, third and fourth rows). The calibration was improved with recalibration, with observed risks more closely matching the predicted risks (fig 6, bottom row).

118Table 4 shows the sensitivity values for the mortality equation over 97 days evaluated at different thresholds based on the centiles of the predicted absolute risk in the validation cohort. For example, it shows that 75.73% of deaths occurred in people in the top 5% for predicted absolute risk of death from covid-19 (predicted absolute risks above 0.237%). People in the top 20% for predicted absolute risk of death accounted for 94% of deaths, and the top 25% accounted for 95.99% of deaths. Supplementary table G shows a similar table based on centiles of relative risk. This example shows that 50.93% of deaths occurred in people in the top 5% for predicted relative risk of covid-19 related death (predicted relative risk above 5.3). People in the top 20% for predicted relative risk of death accounted for 77.53% of deaths, and the top 25% accounted for 81.59% of deaths. As an example, table 5 shows characteristics of the 5% of patients with the highest predicted absolute risk of death for all individuals aged 19-100 years.

119Supplementary figures G and H show two clinical examples from the web calculator (https://qcovid.org/BMJ/), showing the absolute and relative risk of catching and dying from covid-19 and the risk of hospital admission due to covid-19. It also shows a ranking of mortality risk based on centiles across the validation cohort. Supplementary figure G shows a 55 year old black African man with type 2 diabetes, a body mass index of 27.7, and no other risk factors. His absolute risk of catching and dying from covid-19 over the 90 day period is 0.1095% (or 1 in 913). His relative risk compared with a white man aged 55 years and a body mass index of 25 is 10.84. The graph shows that he is in the top 10% of the population at highest risk. Supplementary figure H shows a 30 year old white woman with Down’s syndrome with a body mass index of 40. Her absolute risk of catching and dying from covid-19 is 0.024% (or 1 in 4184). Her relative risk compared with a white woman aged 30 years with a body mass index of 25 and no other risk factors is 59.75, and the rank is 75. Her absolute risk of admission to hospital with covid-19 over 90 days is 1 in 272.

120We have developed and evaluated a novel clinical risk prediction model (QCOVID) to estimate risks of hospital admission and mortality due to covid-19. We have used national linked datasets from general practice and national SARS-CoV-2 testing, death registry, and hospital episode data for a sample of more than 8 million adults representative of the population of England. The risk models have excellent discrimination (Harrell’s C statistics >0.9 for the primary outcome). Although the calibration for the hospital admission outcome was good in both time periods, some under-prediction existed for the mortality outcome in the second validation cohort, which improved after recalibration. The recalibration method could be used to transport the risk models to other settings or time periods with different absolute risks of covid-19. QCOVID represents a new approach for risk stratification in the population. It could also be deployed in several health and care applications, either during the current phase of the pandemic or in subsequent “waves” of infection (with recalibration as needed). These could include supporting targeted recruitment for clinical trials, prioritisation for vaccination, and discussions between patients and clinicians on workplace or health risk mitigation—for example, through weight reduction as obesity may be an important modifiable risk factor for serious complications of covid-19 if a causal association is established. Although QCOVID has been specifically designed to inform UK health policy and interventions to manage covid-19 related risks, it also has international potential, subject to local validation. One of the variables in our model (the Townsend measure of deprivation) may need to be replaced with locally available equivalent measures, or some recalibration may be needed. Previous risk prediction models based on QResearch have been validated internationally and found to have good performance outside of the UK.

121Although similarities exist between our study and the recently reported analysis of risk factors from another English general practice database using a different clinical computer system, our project had a different aim—namely, to develop and evaluate a risk prediction model. We used a more comprehensive outcome (including deaths in patients with positive tests for SARS-CoV-2), a much wider range of predictors, and a more granular assessment of ethnicity and body mass index. Our C statistic for mortality (>0.92) is substantially higher than the previous study’s reported value of 0.77. Other prediction models have been reported, although these focus on other outcomes of covid-19, including risk of admission to intensive care or death following a positive test, or clinical decision tools that integrate biochemical and imaging parameters to aid diagnostis. However, most such studies are at high risk of bias, as they have been developed in highly selected cohorts, have limited transparency, are likely to have optimistic reported performance, or did not use covid-19 specific data. This study represents a substantial improvement on previously developed risk algorithms in terms of the size and representativeness of the study population, the richness of data linkages enabling accurate ascertainment of cases (including both in-hospital and out of hospital deaths) across the health network, and the breadth of candidate predictor variables considered. Importantly, it analyses risks at the population level, rather than risks in people with confirmed or suspected infection, and may have relevance for shielding or other policies that seek to mitigate risk of viral exposure.

122Several complexities of modelling adverse risks from covid-19 in the general population warrant discussion. We used a general population approach which, although not able to incorporate all determinants of being infected, offers an overall estimate of risk of adverse outcomes from covid-19 that could be used in discussions between clinicians and patients about adjustment of lifestyle or occupational and behavioural factors that could limit viral exposure. Our model predicts risks of “catching covid-19 and then having a severe outcome,” on the basis of data collected during the first peak of the pandemic. The endpoint in this study examines a risk trajectory that comprises two elements: becoming infected, which is predominantly a function of behavioural/environmental factors including occupation, local infection rate, and numbers of social interactions; and risk of hospital admission and death due to the infection, which is arguably primarily driven by “vulnerability” (that is, biological/physiological factors including age, sex, body mass index, comorbidities, and medications). Although producing a prediction model for risk of “death if infected” is feasible in principle, this approach is not yet possible owing to the approach to testing in the UK and the context of an as yet incompletely quantified degree of asymptomatic background transmission. Limited covid-19 testing data are available, but the difficulty is that no systematic community testing was done in the UK during the study period, so only patients unwell enough to attend hospital were tested. This means that a risk score developed in those who tested positive would overestimate risks of severe outcomes. As more widespread testing is done and those data become available, we will be able to update the model to take background infection rates into account and also model regional differences. Although the absolute risk levels will of course change over time, depending on the incidence of the disease, our analysis over two validation time periods indicates that the relative risk measures and discrimination are likely to remain stable.

123Secondly, the model estimates the absolute risk for a non-infected individual in the general population of becoming infected and then dying (or needing to be admitted to hospital) from the virus over a 97 day period. Although many more than 40 000 people have died from covid-19 in the UK to date, when the denominator is a population of multi-millions, the absolute risk for most people may be low. Therefore, when conveying this type of risk score to an individual, due emphasis is needed on the different meanings of absolute and relative risk.

124Thirdly, the absolute risk of catching covid-19 depends not only on the incidence of the infection but also on the number of people one gets close to. For this reason, non-pharmacological interventions such as social distancing and shielding were introduced in the UK during the study period. We have included some measures of multi-occupancy, as we have factored care homes into the analysis. The data generated during the study period will therefore be affected by the uptake of interventions such as social distancing and shielding, intended to mitigate the risks of SARS-CoV-2 infection. This could result in underestimation of some model coefficients and hence underestimation of absolute risk in people who were shielded. Also, as this is a prediction model derived from an observational study, the associations estimated for individual predictor variables should not be interpreted as causal effects.

125However, ethical questions must be considered regarding how the tools may be used. We have presented two ways of stratifying risk based on either absolute or relative risk measures with associated centile values, but the choice of whether to have a threshold (given that risk is a continuous measure), and if so what threshold, will depend on the purpose for which the risk assessment tool is to be used, the available resources, and the ethical framework for decision making. We have analysed this within the “four ethical principles” framework that is widely used in medical decision making. The four principles are autonomy, beneficence, justice, and non-maleficence. The new risk equations, when implemented in clinical software, are designed to provide more accurate information for patients and clinicians on which to base decisions, thereby promoting shared decision making and patient autonomy. They are intended to result in clinical benefit by identifying where changes in management are likely to benefit patients, thereby promoting the principle of beneficence. Justice can be achieved by ensuring that the use of the risk equations results in fair and equitable access to health services that is commensurate with patients’ level of risk. Lastly, the risk assessment must not be used in a way that causes harm either to the individual patient or to others (for example, by introducing or withdrawing treatments where this is not in the patient’s best interest), thereby supporting the non-maleficence principle. How this applies in clinical practice will naturally depend on many factors, especially the patient’s wishes, the evidence base for any interventions, the clinician’s experience, national priorities, and the available resources. The risk assessment equations therefore supplement clinical decision making and do not replace it. With these caveats, the predicted risk estimates can be used to identify people at higher risk, to inform shared decision making between healthcare professionals and service users, or for population level stratification.

126Our study has some major strengths, but some important limitations, which include the specific factors related to covid-19 along with others that are similar to those for a range of other widely used clinical risk prediction algorithms developed using the QResearch database. Key strengths include the use of a very large validated data source that has been used to develop other risk prediction tools; the wealth of candidate risk predictors; the prospective recording of outcomes and their ascertainment using multiple national level database linkage; lack of selection, recall and respondent biases; and robust statistical analysis. We have used non-linear terms for body mass index and age. We examined interaction terms, which show increased risks at younger ages for adults with type 2 diabetes. We also established a new linkage to the systemic anti-cancer therapy (SACT) database for chemotherapy prescribed and administered in secondary care (which may not be recorded well in general practice software) to circumvent possible missing data for this important variable.

127Specific limitations include the occurrence of shielding during the study period and that the study was conducted during the first phase of the UK epidemic. We have accounted for many risk factors for covid-19 mortality, but risks may be conferred by some rare medical conditions or other factors such as occupation that have not yet been observed or are poorly recorded in general practice or hospital data. In particular, the model does not include two important predictors—namely, prevailing infection rate and personal social distancing measures. A lack of comprehensive testing has led to some missing data on covid-19 admissions and/or deaths, which means that development of a valid model for predicting death in people infected with SARS-CoV-2 is not yet possible. We acknowledge that absolute risks are changing during the course of the pandemic, so these should be interpreted with caution. However, we would expect predictors of risk, relative risk measures, and discrimination to be more stable over time, which is consistent with the results from our temporal validation. Although this tool was modelled on the best available data from the first wave of the pandemic, it will be updated as further testing and outcome data accrue, immunity levels change, and (potentially) a vaccine becomes available. Nevertheless, having a risk score available at this stage of the pandemic may be useful to identify people at high risk before a vaccine or treatment is available.

128We have reported a validation in each of two time periods using practices from QResearch, but these practices were completely separate from those used to develop the model. We have used this approach previously to develop and validate other widely used prediction models. When these have been further externally validated on completely different clinical databases, by ourselves and others, the results have been very similar. Work is already under way to evaluate the models in external datasets across all four nations of the UK and to integrate the algorithms within NHS clinical software systems.

129This study presents robust risk prediction models that could be used to stratify risk in populations for public health purposes in the event of a “second wave” of the pandemic and support shared management of risk. We anticipate that the algorithms will be updated regularly as understanding of covid-19 increases, as more data become available, as behaviour in the population changes, or in response to new policy interventions. It is important for patients/carers and clinicians that a common, appropriately developed, evidence based model exists that is consistently implemented and is supported by the academic, clinical, and patient communities. This will then help to ensure consistent policy and clear national communication between policy makers, professionals, employers, and the public.

130Public policy measures and clinical risk assessment relevant to covid-19 can be aided by rigorously developed and validated risk prediction models

131Published risk prediction models for covid-19 are subject to a high risk of bias with optimistic reported performance, raising concern that these models may be unreliable when applied in practice

132Novel clinical risk prediction models (QCOVID) have been developed and evaluated to identify risks of short term severe outcomes due to covid-19

133The risk models have excellent discrimination and are well calibrated; they will be regularly updated as the absolute risks change over time

134QCOVID has the potential to support public health policy by enabling shared decision making between clinicians and patients, targeted recruitment for clinical trials, and prioritisation for vaccination