key: cord-0697501-8yb1yosk authors: Houston, Hamish; Hakki, Seran; Pillay, Timesh D; Madon, Kieran; Derqui-Fernandez, Nieves; Koycheva, Aleksandra; Singanayagam, Anika; Fenn, Joe; Kundu, Rhia; Conibear, Emily; Varro, Robert; Cutajar, Jessica; Quinn, Valerie; Wang, Lulu; Narean, Janakan S; Tolosa-Wright, Mica R; Barnett, Jack; Kon, Onn Min; Tedder, Richard; Taylor, Graham; Zambon, Maria; Ferguson, Neil; Dunning, Jake; Deeks, Jonathan J; Lalvani, Ajit title: Broadening symptom criteria improves early case identification in SARS-CoV-2 contacts date: 2021-11-25 journal: Eur Respir J DOI: 10.1183/13993003.02308-2021 sha: 78514a4b30648e2c7c9de2564651e061c8e728c6 doc_id: 697501 cord_uid: 8yb1yosk INTRODUCTION: The success of case isolation and contact tracing for the control of SARS-CoV-2 transmission depends on the accuracy and speed of case identification. We assessed whether inclusion of additional symptoms alongside three canonical symptoms (CS) - fever; cough; loss or change in smell or taste – could improve case definitions and accelerate case identification in SARS-CoV-2 contacts. METHODS: Two prospective longitudinal London-based cohorts of community SARS-CoV-2 contacts, recruited within 5 days of exposure, provided independent training and test datasets. Infected and uninfected contacts completed daily symptom diaries from the earliest possible time-points. Diagnostic information gained by adding symptoms to the CS was quantified using likelihood ratios and AUC-ROC. Improvements in sensitivity and time-to-detection were compared to penalties in terms of specificity and number-needed-to-test. RESULTS: Of 529 contacts within two cohorts, 164 (31%) developed PCR-confirmed infection and 365 (69%) remained uninfected. In the training dataset (n=168), 29% of infected contacts did not report the CS. Four symptoms (sore throat, muscle aches, headache and appetite loss) were identified as early-predictors (EP) which added diagnostic value to the CS. The broadened symptom criterion “≥1 of the CS, or ≥2 of the EP” identified PCR-positive contacts in the test dataset on average 2 days earlier after exposure (p=0.07) than “≥1 of the CS”, with only modest reduction in specificity (5.7%). CONCLUSIONS: Broadening symptom criteria to include individuals with at least 2 of muscle aches, headache, appetite loss and sore throat identifies more infections and reduces time-to-detection, providing greater opportunities to prevent SARS-CoV-2 transmission. INSTINCT (Integrated Network for Surveillance, Trials and Investigations into COVID-19 Transmission) and ATACCC (Assessment of Transmission And Contagiousness of COVID-19 in Contacts) were two community-based cohort studies in which contacts of COVID- 19 In INSTINCT, household-contacts living with their index cases were enrolled at home by research nurses (day 0) and visited again on days 7, 14 and 27. Date of ISO was recorded at enrolment and served as a proxy for exposure. Combined nose and throat swabs (CNTS) for RT-PCR testing and blood samples for serology were taken by research nurses at each visit and an additional CNTS by participants on day 4. Samples were processed at the Molecular Diagnostics Unit, Imperial College London. Antibody (IgM and IgG) to SARS-CoV-2 receptor binding domain (anti-RBD) was measured using a two-step double antigen binding assay (DABA) with recombinant S1 antigen on the solidphase and labelled recombinant RBD as detector in the fluid-phase. [16] In ATACCC, household and non-household-contacts (i.e. not residing with their index) were enrolled. Dates of ISO (householdcontacts) or exposure event (non-household-contacts) were provided by NTAT. After nurse-delivered training, participants self-sampled CNTS daily for 14 consecutive days. SARS-CoV-2 RT-PCR testing was performed at the Virus Reference Department, PHE Colindale. At enrolment, demographic information was collected and participants recorded the onset date of prior symptoms. After enrolment, participants completed a daily symptom diary which assessed 20 symptoms (S2). Loss or change in smell or taste was recorded as one item (hereafter referred to as anosmia). INSTINCT data were used as the training dataset. 'Current infection' was set as the target condition and a rigorous composite reference standard was constructed to establish its presence or absence with maximum accuracy. [17] Contacts were assigned to the 'infected' group if they were PCRpositive at day 0, 4 or 7. Contacts were assigned to the 'uninfected' group if they were PCR-negative and had undetectable SARS-CoV-2 antibodies at all time-points. Participants were excluded if they had no serology results or were PCR-negative at all time-points but had detectable SARS-CoV-2 antibodies at study day 0, 7 or 27. ATACCC data were used as the test dataset. In this cohort, daily PCR results were available but serological testing was not performed routinely. Contacts were assigned to the 'PCR-positive' group if they had a positive PCR result by 7 days after enrolment and PCR-negative if all results were negative. Participants who became PCR-positive after study day 7 or had no PCR results were excluded from the analysis. Participants with only one positive PCR result with a high Ct value (>28) were excluded to minimise false-positives caused by recent rather than current infection. In both cohorts, participants were made aware of their PCR results as they became available. Participants with missing ISO or exposure dates were excluded from analyses requiring these data. The study flowchart ( Figure 1 ) depicts participant numbers included in each analysis. [ Figure 1 ] We used time-to-event analysis to describe the onset of COVID-19-related symptoms relative to ISO in INSTINCT (Figure 1 cohort A) . Briefly, we used symptoms reported by 'uninfected' contacts to define baseline time-dependent hazards, and the difference between 'infected' and 'uninfected' contacts to define COVID-19-related hazards for each symptom (S3 for detailed methodology). Symptoms with a probability of occurring due to COVID-19 of >15% by 10 days post-ISO were selected as candidate symptoms for further evaluation. We aimed to quantify any additional diagnostic value gained by adding each of the candidate symptoms to the CS using likelihood ratios (LRs) estimated for individual symptoms within combinations of symptoms. The Spiegelhalter Knill-Jones (SKJ) method was used rather than the independence Bayes approach in order to adjust for dependency caused by symptom cooccurrence. [18] This method is summarised in S4, having been described in detail previously. [18] [19] [20] [21] [22] Symptoms were considered as a series of binary tests based on their occurrence by each study day (e.g. fever by day 3 would be regarded as positive if fever had been reported on study day 2). Persistent cough and productive cough were combined into a single cough variable. We compared models using the CS to those with an additional symptom. AUC-ROC allows evaluation of model discrimination in training and test datasets. Candidate symptoms with useful LRs after adjustment for dependency with the CS and whose addition improved AUC-ROC across multiple early timepoints were considered 'early-predictors' (EP). To assess real-world impact through readily applicable case definitions, each of the EP were added to the CS individually and together as a list requiring more than 1 to be positive by using the words "at least". Diagnostic performance was assessed against the serial PCR reference standard in the test dataset (cohort D, Figure 1 ) at each day post-exposure. We used time-to-event analysis to measure how quickly broadened case definitions would identify PCR-positive individuals and log-rank tests to make comparisons with the CS. Finally, we quantified the prevalence-dependent trade-off between true-positives and false-positives by calculating the NNT: the number of false-positives for every true-positive plus 1. Statistical analyses were performed in Stata version 17.0 (StataCorp, LLC, College Station TX) and R (R Core Team, 2020). (21/73, 28 .8%) did not report fever, cough or anosmia by day 7. Time-to-event analysis of symptom onset following exposure (S7-8) showed that fever preceded anosmia and persistent cough preceded productive cough. Sore throat and rhinitis occurred early, and breathlessness later. Fatigue was commonly reported by 'uninfected' contacts. Thirteen symptoms had a probability of occurring due to COVID-19 of >15% by 10 days post-ISO (fever, persistent cough, productive cough, anosmia, headache, muscle aches, sore throat, rhinitis, appetite loss, breathlessness, diarrhoea, nausea and abdominal pain). 9 of these 13 symptoms are not included in the CS and were denoted candidate symptoms in further analyses. Other than the CS: rhinitis, sore throat, headache, muscle aches, and appetite loss had the largest cumulative COVID-19-related hazards. Raw counts of participants who had reported each symptom by each study day in the training cohort are presented in S9. Used alone, cough, rhinitis, headache and muscle aches were the most sensitive symptoms whilst nausea and abdominal pain were insensitive (S10). Anosmia, fever, and appetite loss were highly specific symptoms. The crude LRs (S11) show that any of the symptoms will affect post-test odds when they are used alone. However, when used in combination with other symptoms, their LRs after adjustment using the SKJ approach ( Figure 2 , Table 1 , S12) were all less extreme than their crude LRs, indicating considerable dependency between symptoms. [ Figure 2 ] [ Table 1] When cough was used in combination with fever and anosmia it's adjusted LRs were closer to 1 than those of anosmia or fever (Figure 2A , S12). This was most likely due to the higher specificity of anosmia and fever (S10). When combined with the CS the presence or absence of nausea did not independently affect post-test odds, its adjusted LRs lying close to 1 (Figure 2 ). Breathlessness was more common in the 'infected' group. However, whilst breathlessness was reported without fever, anosmia or cough by 'uninfected' contacts, this was rare in the 'infected', explaining why its adjusted positive LRs are below 1 and negative LRs above 1. In training and test datasets, AUC-ROCs increased with study day reflecting improved discrimination afforded by greater accumulation of symptoms by later study days in 'infected' contacts ( Figure 3 ). AUC-ROC was often greater in test data than training data, likely reflecting the longer median time to recruitment in ATACCC. [ Figure 3 ] Between study days 0 to 3, the addition of headache, sore throat, muscle aches and appetite loss to the CS yielded the greatest improvements in AUC-ROC in the test dataset. When combined with the CS, appetite loss, headache, sore throat and muscle aches all consistently had positive adjusted LRs above 1 and negative adjusted LRs below 1 showing that both their presence and their absence added to the CS's ability to discriminate between the infected and uninfected. These symptoms were therefore considered 'early-predictors' (EP). Each of the four EP were combined individually with the CS using an "or" operator, as well as together using "or" and "at least" operators (Box 1). The addition of any symptom to the CS using an "or" operator increased sensitivity ( Figure 4A ) whilst reducing specificity ( Figure 4B ). The addition of appetite loss produced the smallest changes compared to the CS. [ Figure 4 ] The CS identified 50% of PCR-positives by 6 days post exposure ( Figure 5 , Table 2 , S14). Adding headache yields the greatest increase in sensitivity ( Figure 4A , S13A) and would identify PCRpositives on average 2 days earlier (p=0.02), but causes the largest reduction in specificity (15.2% at 5 days post-exposure; Figure 4B , S13A). In contrast, CS or sore throat only reduced specificity by 5.7% at 5 days and identified PCR-positive cases earlier than the CS, by 1 day on average. This change was not statistically significant given the small number of PCR-positive participants in the cohort (p=0.1, n=91). [ Figure 5 ] [ Table 2 ] When all 4 EP are added to the CS, if all 4 are required, there is very little difference to the CS. In contrast, the case definition 1 of the CS, or 1 of the EP would increase sensitivity and identify PCRpositive cases a median 2 days earlier than the CS (p=0.002). However, the corresponding reduction in specificity by 5 days post-exposure (19.7%, Figure 4B , S13B) would lead 25% of PCR-negative individuals to be inappropriately identified (Table 2, Figure 5B ). 1 of the CS, or 2 of the EP identified PCR-positive cases a median 2 days earlier than the CS (p=0.07) with a reduction in specificity of only 5.7% at 5 days post-exposure. This reduction is smaller than that caused by moving from the CS to various other international case definitions (S16). None of the EP were dispensable from this proposed criterion (S17). The number of individuals identified in order to yield a single PCR-positive case, the NNT, increases rapidly immediately after exposure, reflecting an initial accumulation of false-positives because no one has yet developed symptoms actually caused by infection ( Figure 4C ). NNT plateaus around 4-5 days following exposure, reflecting the incubation period. At 25.6% prevalence, 1 of the CS, or 2 of the EP had a NNT at 5 days post-exposure of 1.78 compared to 1.61 for the CS, indicating 17 additional individuals identified for every 100 infected individuals identified. This, to our knowledge, is the first study to use daily symptom data prospectively collected from recently exposed infected and uninfected SARS-CoV-2 contacts to evaluate the diagnostic performance of symptom combinations for detecting infection. Using this definitive study design, we found that 29% of individuals with PCR-confirmed COVID-19 did not report any of the CS, but 93% reported at least one symptom from a broader list of 20. We identified 4 EP symptoms (sore throat, headache, muscle aches, and appetite loss) providing additional early predictive power for identifying SARS-CoV-2-infected contacts. The case definition 1 of the CS, or 2 of the EP identified PCR-positive contacts 2 days earlier after exposure than the CS alone (p=0.07). This time-saving is critical given that shortening the delay from infectiousness to self-isolation from 2.6 to 1.2 days has been estimated to reduce transmission by 47%. [3] Moreover, the proportion of "symptomatic" infections and time-to-symptom onset are critical parameters in studies modelling effectiveness of testing and isolation strategies for contacts. [24] Consistent with previous studies, headache and sore throat were sensitive symptoms, [7] which occurred early in the course of infection, [15] and were prevalent in our relatively young participants. [24] Importance of these symptoms will increase as vaccination of older age groups increases the proportion of infections occurring in the young. In agreement with the Real-time Assessment of Community Transmission-1 (REACT-1) study, we found that headache, muscle aches and appetite loss improved discrimination within statistical prediction models. [14] We add a crucial evaluation of readily applicable case definitions. We observed that both the structure of symptom criteria (e.g. use of the Boolean operator "at least") and time-from-exposure had a considerable effect on diagnostic performance. The SKJ approach enabled another important new observation. Although an important indicator of disease severity, [26] breathlessness was not a useful additional symptom for identifying early and mild infections because a hierarchy of symptoms exists. Breathlessness is unlikely to occur due to COVID-19 without prior fever, cough or anosmia and its inclusion reduces specificity. Further strengths include day-by-day measurement of diagnostic performance following exposure and prospective data collection which mitigates recall bias. The rigorous reference standard employed in our training cohort maximised accuracy for the target condition and ensured only the most useful symptoms were taken forward to the test data. Neither serology nor PCR have 100% sensitivity for SARS-CoV-2 infection. [27] Click or tap here to enter text. Using both serology and PCR at multiple time-points to define the absence of infection we minimised false-negatives. Falsepositive PCR results caused by recent rather than current infection were likely less common in our longitudinal study of recently exposed contacts than in studies involving random communitysampling. [14] Limitations include modest sample size, largely White British population, minor differences between training and test cohorts and the potential for tick-box and behavioural biases. Study participants were usually highly motivated and attentiveness to mild symptoms (e.g. rhinitis) may have been increased by awareness of exposure, frequent study visits and co-residence with other participants. Contacts could not be blinded to their PCR results or those of their index. Since we studied community-based COVID-19 contacts identified through NTAT, our findings are very likely generalizable. As large-scale cross-sectional data replicate our findings in smaller-scale daily-resolution longitudinal data, the combined evidence-base is now sufficient to influence policy. Broadening symptom criteria for use in the general population would likely identify more infections and reduce time-to-detection, reducing transmission. We propose that symptom criteria within case definitions to prompt symptomatic isolation and testing of SARS-CoV-2 contacts should include headache, sore throat, muscle aches and appetite loss as well as the canonical symptoms to optimise sensitivity. Two of these additional symptoms should be required to maximise specificity. As highly vaccinated regions transition to lower COVID-19 incidence, investment in RT-PCR testing capacity will make such broader case definitions feasible. As societies develop alternatives to testtrace-isolate, application of evidence-based symptom criteria alongside judicious testing will be critical for early discrimination of infected and uninfected contacts. Accordingly, our findings should inform development of evidence-based national testing policies in many parts of the world now and in subsequent phases of the pandemic. Results are presented in accordance with STARD (Standards for Reporting of Diagnostic Accuracy Studies) and STROBE (Strengthening Reporting of Observational Studies in Epidemiology) guidelines. No competing interests were declared by any of the study authors. Table 1: Adjusted Likelihood Ratios for individual symptoms when canonical symptoms are used in combination with an additional candidate symptom. LRs for individual symptoms in combination with other symptoms were calculated using the Spiegelhalter Knill-Jones method which adjusts for the dependency caused by symptom co-occurrence. Adjusted LRs for study days 0-7 are represented graphically in Figure 2 and values for study days 0, 2 & 4 are presented here. For any combination of symptoms by a particular time-point, the adjusted LRs can be multiplied together to estimate effect on posttest odds. For example, on study day 0, a contact with fever and headache but without cough or anosmia has times the odds of being infected compared to a contact where symptoms are unknown (see S4 for worked example The 'early-predictors' were each combined individually with the CS using an "OR" operator. fever OR cough OR anosmia OR headache CS or headache fever OR cough OR anosmia OR sore throat CS or sore throat fever OR cough OR anosmia OR muscle aches CS or muscle aches fever OR cough OR anosmia OR appetite loss CS or appetite loss The 'early-predictors were all combined together with the CS using "at least" and "OR" operators. Figure 1 ). AUC Test = Area under the receiver operating characteristic curve in test dataset (cohort C, Figure 1 ). The hard-line marks AUC Test for the CS model. The dotted-line marks AUC Train for the CS. Models where the addition of a candidate symptom yielded better predictions in the training dataset lie to the right of the dotted-line and models where better predictions were yielded in the test dataset lie above the hard-line. Figure 4 : Sensitivity, specificity and number-needed-to-test for the canonical symptoms and broadened symptom criteria. The 'early-predictors' (EP): sore throat, headache, muscle aches, appetite loss, were each combined individually with the canonical symptoms (CS): fever, cough and anosmia, using an "OR" operator and all were added together using "at least" and "OR" operators (as described in Box 1). Sensitivity [A], specificity [B] and number-needed-to-test [C] were calculated for each symptom criterion by day post exposure (ISO for household-contacts) against a serial PCR reference standard. Full results are given in S13. Number-needed-to-test is calculated by dividing the number of false-positives by the number of truepositives and adding 1. Rarely, symptoms were reported at enrolment without an onset date. We imputed onset dates for these symptoms by assuming the median number of days pre-enrolment (maximum 2 participants [0.55%] for rhinitis). The 'early-predictors' (sore throat, headache, muscle aches, appetite loss) were each combined individually with the canonical symptoms (CS) using an "OR" operator and all were added together using "at least" and "OR" operators (as described in Box 1). The proportion of PCR-positive (left) and PCR-negative (right) participants who were positively identified by each case definition by each day following exposure (ISO for householdcontacts) is shown using a Kaplan-Meier plot. The plot for the CS is shown in black. Life-tables are presented in S14 and median time-to-diagnosis in Table 2 . Rarely, symptoms were reported at enrolment without an onset date. We imputed onset dates for these symptoms by assuming the median number of days pre-enrolment (maximum 2 participants [0.55%] for rhinitis). Exact wordings of criteria can be found at: World Health Organization, [ We aimed to identify symptoms with a high probability of occurring due to COVID-19 by 10 days post index symptom onset. Index symptom onset (ISO) was defined as the onset date of the first of any of the 20 symptoms. We used time-to-event analysis to describe the onset of COVID-19 related symptoms by comparing symptom onset in 'infected' and 'uninfected' contacts in cohort A (Figure 1 ). Participants were excluded from this analysis if the date of index symptom onset was unknown, or if the contact reported symptoms which occurred more than 3 days prior to index symptom onset. Rarely, symptoms were reported at enrolment without an onset date. We imputed onset dates for these symptoms using the median number of days preenrolment (maximum 9 participants [5.4%] for muscle aches). Participants with incomplete symptom diaries were right-censored on the day their symptom diary ended. We selected symptoms where this probability was more than 15% at days as candidate symptoms for further evaluation. Bootstrap 90% confidence intervals were calculated for h(t) and H(t). Likelihood ratios (LRs) have clinical value as they quantify how much a diagnostic test result (or symptom) will raise or lower the pre-test probability of the target condition. Bayes's theorem (by which LRs are calculated) assumes that diagnostic tests act independently. However, symptoms often co-occur and this dependency must be accounted for to avoid overoptimistic evaluations of diagnostic performance. Spiegelhalter and Knill-Jones combined Bayes's theorem with logistic regression to produce LRs which are adjusted for dependency. We used Spiegelhalter Knill-Jones models to study the predictive power of symptom combinations and measure the diagnostic usefulness of each symptom within the combinations. The analysis was performed using Cohort B as the training dataset and Cohort C as the test dataset (Figure 1 ). Symptoms were considered as diagnostic tests based on their occurrence by each study day for the day of recruitment (day 0) and the first 7 days thereafter. Diagnostic performance is assessed compared to a composite serial PCR and serology reference standard in the training dataset and a serial PCR reference standard in the test dataset (described in the main text). A series of 8 Spieglhalter Knill-Jones models were created, one for each study day, using 3 predictors (fever, cough, anosmia) to evaluate the canonical symptoms (CS). Nine further sets of 8 models were created using 4 predictors (fever, cough, anosmia, and the candidate symptom) to evaluate the effect of adding one of the 9 candidate variables to the 3 CS at each study day. Adjusted LRs were calculated for each of the canonical symptoms (fever, cough and anosmia) when used together, and for each candidate symptom when added to the canonical symptoms. In brief, rather than including sets of (0,1) indicator variables for each symptom in a logistic regression, the indicator variables take the value of the crude positive and negative likelihood ratios in a logistic regression equation of the form: The formula for calculating adjusted LRs is given by We calculated the area under the receiver operating characteristic curve (AUC-ROC) for each predictive model in the training dataset and test dataset (cohort C, Figure 1 ) and compared to the AUC-ROCs for predictive models containing only fever, cough and anosmia. Candidate symptoms whose addition yielded consistent improvements in AUC-ROC across multiple early time-points were considered 'early predictors' and were evaluated within simple diagnostic decision tools. A Spiegelhalter Knill-Jones model using fever, cough, anosmia and headache as predictor The log of the odds of infection at study day 1 is used as the offset term: Instead of 1 and 0, the indicator variables take the value of the crude positive and negative likelihood ratios for each of the CS with missing data entered as having log likelihood ratios of zero. The model fitted for the example is: The shrinkage factor is used to calculate the adjusted LRs ( Figure 2 ) for symptom . For example, the positive adjusted LR for fever when used in combination with cough, anosmia & sore throat on day 0 is calculated as follows: Note how the magnitude of the LR has been shrunk to account for the symptom's dependency with other symptoms in the model. Figure 1 ) and compare models using the CS to those with an additional symptom (Figure3). S5: Recruitment cascade. Early in the study, contacts were referred to our study through the Royal College of General Practitioners Research and Surveillance Centre. Later in the study, contacts were referred either directly through National Test And Trace or, in the case of most household-exposed contacts, through their index case, which PHE had identified. In total, 53011 referrals were received via these three pathways. If contacts were identified via PHE, PHE would initially ascertain whether participants consented to being contacted to take part in a research study. Consenting Contacts would then be contacted by Imperial, of which a subset would agree to take part. PHE and Imperial strove to contact all eligible contacts, but were limited by staff resources. The potentially eligible contacts PHE and Imperial attempted to call were non-biased in that the staff systematically worked down their list of participants, which was organised at random. Where referrals concerned index cases, multiple contacts may have been recruited following a single referral. This analysis was performed using Cohort A (Figure 1 ). For each 2 day time interval, the number of participants at risk of onset of each symptom at the start of each time period is given. The numbers of participants reporting symptom onset, and participants whose study diary ended within the time interval are given in brackets. At enrolment, some contacts reported symptom onset prior to any symptom onset in the index. These events are shown in time periods -4 to -2 and -2 to 0 (days post index symptom onset). As prospective study diaries always started after index symptom onset, the number of participants whose symptom diary ended within these time intervals is not applicable (NA). Rarely, symptoms were reported at enrolment without an onset date. We imputed onset dates for these symptoms using the median number of days preenrolment (maximum 9 participants Symptoms were considered as a series of binary tests based on their occurrence by each study day. Contacts were defined as infected or uninfected based on a composite PCR and serology reference standard (see main text). These raw counts were used to calculate diagnostic sensitivity and specificity of individual symptoms (S10) as well as crude likelihood ratios (S11) which were used in the Spiegelhalter Knill-Jones analysis. TP = True-positive, FN = False-negative, TN = True-negative, FP = False-positive. TP TP TP TP TP TP TP TP TP TP TP TP Infected 0 73 13 25 15 25 29 14 26 5 15 9 7 9 1 72 20 31 15 34 39 20 37 11 20 12 10 12 2 72 24 33 21 39 45 25 40 15 23 14 14 15 3 71 28 34 26 40 47 27 43 18 26 14 14 15 4 71 29 39 28 42 49 29 46 20 28 14 16 18 5 69 28 38 28 41 48 30 48 20 28 14 16 20 6 65 29 40 28 38 46 30 46 20 27 14 16 22 7 57 25 33 25 35 39 26 40 16 23 12 15 18 Day TN + FP FP FP FP FP FP FP FP FP FP FP FP FP Uninfected 0 95 2 10 1 14 16 5 7 6 4 5 4 3 1 62 3 9 2 11 16 7 10 5 4 4 4 3 2 62 4 10 2 13 18 10 11 5 5 4 4 3 3 61 4 9 2 15 19 10 11 7 5 5 4 5 4 61 4 10 2 15 20 11 11 8 5 5 4 6 5 58 4 10 2 14 19 11 11 8 5 5 4 6 6 57 3 10 2 13 18 11 11 7 4 4 3 6 7 55 2 9 2 12 17 12 11 6 5 3 1 ) against a serial PCR reference standard. We compared the canonical symptoms (CS), our proposed symptom criteria (1 of the CS, or 2 of the EP) and international case definitions (S1, [3] [4] [5] [6] [7] ). Where the precise wording of items in our symptom diaries (S2) did not correspond exactly with wording in international case definitions the closest approximation was chosen. In the WHO case definition, anorexia/nausea/vomiting is considered one symptom item whereas in our analysis loss of appetite and nausea/vomiting were considered two separate items, possibly overestimating sensitivity and underestimating specificity. In the US Centers for Disease Control case definition, chills and rigors are included as two separate symptom items as well as fever. In our analysis we considered fever as one variable, possibly underestimating sensitivity and overestimating specificity. The Clopper-Pearson interval was used to calculate 95% confidence intervals. [2] Kaplan-Meier plots were constructed showing time from exposure (index symptom onset for household-contacts) until positive identification by these symptom criteria in PCRpositive (C) and PCR-negative (D) contacts. The plot for the CS is shown in black. Rarely, symptoms were reported at enrolment without an onset date. We imputed onset dates for these symptoms by assuming the median number of days pre-enrolment (maximum 2 participants [0.55%] for rhinitis). Table 7 . We performed a leave-one-out analysis to determine whether each of the 'early predictor' symptoms (headache, muscle aches, sore throat, appetite loss) were necessary within the Community transmission and viral load kinetics of the SARS-CoV-2 delta (B.1.617.2) variant in vaccinated and unvaccinated individuals in the UK: a prospective, longitudinal, cohort study Contact tracing in the context of COVID-19. World Health Organization Effectiveness of isolation, testing, contact tracing, and physical distancing on reducing transmission of SARS-CoV-2 in different settings: a mathematical modelling study. The Lancet Infectious Diseases Lancet Publishing Group Predicting infectious severe acute respiratory syndrome coronavirus 2 from diagnostic samples Epidemiological Correlates of Polymerase Chain Reaction Cycle Threshold Values in the Detection of Severe Acute Respiratory Syndrome Coronavirus Clinical Infectious Diseases Symptom profiles and accuracy of clinical definitions for COVID-19 in the community Optimal symptom combinations to aid COVID-19 case identification: Analysis from a community-based, prospective, observational cohort World Health Organisation. WHO COVID-19 Case definition Coronavirus disease 2019 (covid-19) 2020 interim case definition Estrategia de detecci n precoz, vigilancia y control de covid-19 Get a free PCR test to check if you have coronavirus Could expanding the covid-19 case definition improve the UK's pandemic response? Predictive symptoms for COVID-19 in the community: REACT-1 study of over 1 million people Symptom diary-based analysis of disease course among patients with mild coronavirus disease Symptoms of COVID-19 in a population-based cohort study Testing for responses to the wrong SARS-CoV-2 antigen? Target product profile: In Vitro Diagnostic (IVD) self-tests for the detection of SARS-CoV-2 in people without symptoms Three methods to construct predictive models using logistic regression and likelihood ratios to facilitate adjustment for pretest probability give similar results Statistical and Knowledge-Based Approaches to Clinical Decision-Support Systems, with an Application in Gastroenterology Probabilistic prediction in patient management and clinical trials Diagnostic systems as an aid to clinical decision making Making better decisions: Construction of clinical scoring systems by the Spiegelhalter-Knill-Jones approach Quarantine and testing strategies in contact tracing for SARS-CoV-2: a modelling study Clinical characteristics of children and young people admitted to hospital with covid-19 in United Kingdom: Prospective multicentre observational cohort study Clinical, laboratory and radiological characteristics and outcomes of novel coronavirus (SARS-CoV-2) infection in humans: A systematic review and series of meta-analyses Clinical and immunological assessment of asymptomatic SARS-CoV-2 infections Confidence Intervals for the Ratio of Two Binomial Proportions The use of confidence or fiducial limits illustrated in the case of the binomial Get a free PCR test to check if you have coronavirus World Health Organisation. WHO COVID-19 Case definition Coronavirus disease 2019 (covid-19) 2020 interim case definition Ministerio de Sanidad vigilancia y control de covid-19 We would like to thank all the participants who were involved in the study and the support of our