REACT-1 study of over 1 million people

REACT-1 is a series of community prevalence surveys of SARS-CoV-2 virus swab positivity in England, conducted at approximately monthly intervals since May 2020. Using the National Health Service patient register across the 315 lower-tier local authorities (LTLAs) in England, recruitment letters were sent to a random nationally representative sample of individuals aged 5 years and over, with a separate (non-overlapping) sample selected at each round. We therefore approached a different base sample at each round and did not have repeat samples to account for. Participant sampling aimed to achieve approximately equal numbers of participants in each LTLA. Random samples were drawn stratified by LTLA, with larger numbers selected in some LTLAs to address variable response rates at the LTLA level. Up to 160,000 valid responses and viable swabs were obtained at each round [9]. Participation involved a self-administered throat and nasal swab, and the completion of a short online or telephone questionnaire including information on demographic variables, household composition, behaviors, and recent symptoms. The questionnaires used are available on the study website (https://www.imperial.ac.uk/medicine/research-and-impact/groups/react-study/react-1-study-materials/). At the time of survey completion, participants were unaware of the result of their swab test. In the survey, participants were asked about new symptoms occurring in the week preceding the swab. These included a set of 26 clinically relevant symptoms potentially related to COVID-19: (i) loss or change of sense of smell and loss or change of sense of taste, (ii) coryzal symptoms (runny nose, sneezing, blocked nose, sore eyes, sore throat, and hoarse voice), (iii) gastrointestinal symptoms (appetite loss, nausea/vomiting, diarrhea, and abdominal pain/belly ache), (iv) fatigue-related symptoms (tiredness, severe fatigue, heavy arms/legs, and difficulty sleeping), (v) respiratory and cardiac symptoms (new persistent cough, shortness of breath, chest pain, and tight chest), and (vi) other flu-like and miscellaneous symptoms (fever, muscle aches, chills, headache, dizziness, and numbness/tingling). Data from round 1 were excluded as the symptom questions asked in that round were not consistent with those in the subsequent rounds [9].

PCR-positive swab samples from round 8 with N-gene cycle threshold value < 34 and sufficient sample volume underwent genome sequencing. Viral RNA was amplified using the ARTIC protocol [10], and sequencing libraries were prepared with CoronaHiT [11]. Samples were sequenced using the Illumina NextSeq 500 platform. Each run included 1 positive and 1 negative control per 96 samples. Sequencing data were analyzed using the ARTIC bioinformatic pipeline [12]. Lineages were assigned using PangoLEARN [13].

Twenty-five participants (19 in rounds 2–7 and 9 in round 8) were excluded due to missing information on sex. We used univariate logistic regression to model the risk of testing positive for SARS-CoV-2 as a function of symptoms reported in the week prior to testing. For multivariable models, round 2–7 data (restricted to people reporting any of the 26 surveyed symptoms) were split into a 70% training set (76,187 observations, of which 1,078 were positive) and a 30% test set (32,651 observations, of which 461 were positive). We adopted a variable selection approach with stability selection using least absolute shrinkage and selection operator (LASSO) penalized logistic regression with all 26 surveyed symptoms as predictors and PCR positivity as the outcome [14]. LASSO models were fit on 1,000 independent random 50% subsamples of the 70% training set. The stability selection penalty parameter was calibrated to give a per family error rate of fewer than 5 falsely selected symptoms [14,15]. The proportion of penalized models where each symptom was included (its selection proportion) was used as a measure of its importance; those with selection proportion above 50% were considered stably selected and were included in a predictive model of PCR positivity, where mean penalized odds ratios across models (where selected) were used as weightings. The resultant model was then applied to the 30% holdout test set (rounds 2–7) and, separately, to round 8 data. We then estimated the proportion of symptomatic COVID-19 cases that would be detected using this predictive model across all levels of symptomatic community testing.

All calculations were done with the R computational environment, version 4.0.2, using the logistic LASSO algorithm as implemented in the glmnet package and in-house scripts for the stability selection (available at https://github.com/mrc-ide/reactidd/tree/master/R/symptom_prediction_paper_scripts).

Participants in the REACT-1 study were not involved in the definition of our research question or in the outcome measurements. They were not involved in developing the analytical plan or the implementation of the study. There are no plans to disseminate the results of the research directly to the study participants or the relevant patient community. However, a public advisory panel provides regular review of the study processes and results, and links to published reports from the study are available on the study website (https://www.imperial.ac.uk/medicine/research-and-impact/groups/react-study/real-time-assessment-of-community-transmission-findings/).

The key characteristics of the study population in rounds 2–7 and 8 for the full sample and for those reporting symptoms are summarized in S1 Table. Of the 979,709 respondents (after 19 exclusions) with a valid swab test in REACT-1 rounds 2–7, 870,872 (88.9%) reported no symptoms in the week prior to testing while 108,837 (11.1%) reported at least 1 of the 26 surveyed symptoms. We detected 4,168 PCR-positive cases, of whom 1,538 (36.9%) reported 1 or more symptoms in the past week. In round 8, there were 167,636 participants (after 6 exclusions), of whom 146,701 (87.5%) reported no symptoms in the past week and 20,935 (12.5%) reported at least 1 of the 26 surveyed symptoms. We detected 2,282 PCR-positive cases, of whom 1,031 (45.1%) reported 1 or more symptoms in the past week (Fig 1; S2 Table). In univariate analyses, each of the 26 surveyed symptoms was associated with PCR positivity in rounds 2–7 and round 8 (Fig 2; S2 Table). In rounds 2–7, with a mean prevalence of PCR positivity of 0.46%, the positive predictive value for any symptom was 1.4%, with an odds ratio for PCR positivity of 4.7 compared with non-symptomatic individuals. In round 8, with a PCR positivity prevalence of 1.36%, the positive predictive value for any symptom was 4.9%, with an odds ratio for PCR positivity of 6.0 compared with non-symptomatic individuals (Fig 1).

Fig 2. Results from univariate logistic regression models of PCR positivity for 26 surveyed symptoms.

Effect size estimates are expressed as odds ratios (95% confidence intervals) in rounds 2–7 (left) and round 8 (right).

https://doi.org/10.1371/journal.pmed.1003777.g002

Seven symptoms were selected as jointly positively predictive of PCR positivity in a LASSO stability selection model trained on round 2–7 data: loss or change of sense of smell, loss or change of sense of taste, fever, new persistent cough, chills, appetite loss, and muscle aches. In addition, numbness/tingling was selected as jointly but negatively predictive of PCR positivity. The resultant model gave an area under the curve (AUC) of 0.75 for holdout round 2–7 test data and 0.77 for round 8 data, with no improvement in AUC upon inclusion of other symptoms (Fig 3). Testing people in the community with at least 1 of the 7 selected positively predictive symptoms gave sensitivity, specificity, and positive predictive value of 71%, 64%, and 1.9% in holdout round 2–7 test data and 74%, 64%, and 9.7% in round 8 data, respectively. Performance of the stability selection model at varying levels of community testing is shown in S1 Fig; for example, testing 10% of symptomatic individuals would detect around half of all symptomatic cases.

Fig 3. Selected symptoms predictive of COVID-19.

Results of LASSO stability selection using 1,000 models (with 50% subsamples of training data from rounds 2–7). Mean (penalized) log odds ratios (log ORs) across all models are shown in the top panel. Positive regression coefficients are presented in teal, and negative in red. Only symptoms selected at least once are displayed. The selection proportions (selection prop.; proportion of 1,000 models that included each symptom) are shown in the middle panel; the horizontal dashed line shows the selection threshold of 50%. Symptoms are ordered according to their selection proportions, and selected symptoms are in black. The bottom panel shows the area under the curve (AUC) of models adding each variable in order of selection proportion (from left to right) in both holdout data from rounds 2–7 (grey) and data from round 8 (red).

https://doi.org/10.1371/journal.pmed.1003777.g003

Of the 2,282 positive cases in round 8, 1,088 had cycle threshold value < 34 and underwent viral genome sequencing; 898 (82.5%) were B.1.1.7, 181 (16.6%) were wild type, and 8 were other lineages. Characteristics of the positive cases with B.1.1.7 and wild type in round 8 were similar except for the region of the case, with a marked excess of B.1.1.7 cases in East of England, South East, and London (S3 Table). LASSO stability selection models fit on round 8 data with an outcome of B.1.1.7 versus PCR negative selected the same 7 positively predictive symptoms as in rounds 2–7 (Fig 4). In univariate analysis within round 8, there was slightly higher reporting of any symptom for B.1.1.7 infection versus wild type. Specifically, we found a higher prevalence of sore throat, new persistent cough, fever, difficulty sleeping, dizziness, and nausea/vomiting at a nominal 0.05 significance level (Fig 5A). Stability selection models identified sore throat and new persistent cough as jointly and positively associated with B.1.1.7 infection compared with wild-type infection, while loss or change of sense of smell was selected as negatively associated, i.e., more predictive of wild type infection (Fig 5B).

Fig 4. Symptoms predictive of B.1.1.7 infection.

LASSO stability selection for symptoms predictive of B.1.1.7 (Alpha) lineage infection versus symptomatic people (aged 5+ years) testing PCR negative in round 8. Mean log odds ratio (Log OR) and selection proportion (selection prop.) are represented for each symptom in the top and bottom panels, respectively. Positive regression coefficients are presented in teal, and negative in red.

https://doi.org/10.1371/journal.pmed.1003777.g004

Fig 5. B.1.1.7 versus wild-type symptoms.

Comparison of symptom profile in B.1.1.7 (Alpha) lineage versus wild-type infection among 1,124 people testing positive in round 8 (other lineages excluded, N = 8). (A) Proportion of people reporting each symptom by lineage (left panel), and the differences in proportions with 95% confidence intervals (right panel). (B) Results of LASSO stability selection (using 1,000 models with 50% subsampling) with B.1.1.7 infection, versus wild-type infection as the outcome, summarized by the mean log odds ratio (Log OR) and selection proportion (selection prop.) for each of the symptoms selected at least once. Positive regression coefficients are presented in teal, and negative in red. The horizontal dashed line represents the selection threshold of 50%.

https://doi.org/10.1371/journal.pmed.1003777.g005

In age-stratified analyses of round 2–7 data, 6 (5–17 years) or 7 (18–54 and 55+ years) symptoms were jointly selected. New persistent cough was not selected for those aged 5–17 years but was for adults, while chills, fever, loss or change of sense of smell, and loss or change of sense of taste were selected in all age groups. Additional age-specific selected symptoms (positive coefficients) were headache (5–17 years), appetite loss (18–54 and 55+ years), and muscle aches (18–54 years), and age-specific negatively associated symptoms (negative coefficients) were runny nose (5–17 years) and numbness/tingling (55+ years) (S2 Fig).

Considering first reported symptom, instead of any symptom, as the predictor, in univariate analysis, all first reported symptoms among the 26 surveyed were associated with PCR positivity in both round 2–7 and round 8 data (S3 Fig). In multivariable analysis, LASSO stability selection on round 2–7 data included headache as positively predictive of PCR positivity instead of appetite loss (S4 Fig); otherwise, the selected (positively associated) symptom set was the same as in Fig 3, although with lower predictive performance (AUC of 0.68 and 0.66 for holdout round 2–7 test data and round 8 data, respectively).