Skip to main content

Real-time tracking of self-reported symptoms to predict potential COVID-19

Our take —

Among 18,401 users of a symptom-tracking app in the UK and US, the loss of smell and taste (as well as cough, fatigue, and skipped meals) was strongly associated with a positive SARS-CoV-2 test result. Although a symptom-based predictive model of infection performed reasonably well, caution is warranted because of possible bias arising from the sample of individuals studied, the self-reported nature of the data, and the lack of information on the timing of symptoms relative to testing.

Study design

Cross-Sectional, Other

Study population and setting

The study included 2,618,862 individuals in the UK and the US (65% female, 95% from the UK) who used a COVID-19 symptom tracking app (COVID Symptom Study, developed by Zoe Global) between March 24 and April 21, 2020. Analyses were conducted separately in the UK and US cohorts. Authors were focused on loss of smell and taste as specific symptoms of COVID-19. Patients used the app to report symptoms, whether they were tested for SARS-CoV-2 infection via PCR, and the outcome of the tests.

Summary of Main Findings

18,401 (0.7%) individuals reported being tested for SARS-CoV-2 infection, and of these, 7,178 (39%) reported a positive test. Among those tested, loss of smell and taste was strongly associated with a positive result in multivariable regression (UK:OR 6.40, 95% CI: 5.96-6.87; US: OR 10.01, 95% CI: 8.23-12.16). With data from the UK cohort, the authors used step-wise regression to develop a prognostic model for SARS-CoV-2 infection. The final model included age, sex, loss of smell and taste, cough, fatigue, and skipped meals; this model had an area under the curve (AUC) of 0.76 in 10-fold cross-validation, a sensitivity of 0.65, a specificity of 0.78, a positive predictive value of 0.69, and a negative predictive value of 0.75 in the UK cohort. The model performed similarly in the US cohort, with an AUC of 0.76. When applied to all 805,753 individuals who reported symptoms but were not tested, the model predicted that 17.42% (14.45–20.39%) were likely to be infected, representing 5.36% of all app users.

Study Strengths

The sample size was large, even when limited to those receiving tests. The US cohort allowed validation of the predictive model.


Olfactory dysfunction (anosmia) and taste dysfunction (dysgeusia) were combined in the app questionnaire. Self-reporting of symptoms and of SARS-CoV-2 test results are subject to misclassification; for example, respondents with positive tests may be more likely to report symptoms that have received media attention, such as loss of smell. Furthermore, defining cases by a single PCR test may result in misclassification. The timing of symptoms relative to test results was not reported, limiting the interpretability of the prognostic tool. The population of all respondents may not be representative of the general population (there was a considerably higher proportion of females, for example). Moreover, the subpopulation of respondents receiving SARS-CoV-2 tests are not a random sample of all respondents; they may have been more likely to have known exposures or to have experienced severe symptoms. Associations between reported symptoms and infection status may differ in this selected population.

Value added

This study is the largest published to date of self-reported symptoms as they relate to the likelihood of SARS-CoV-2 infection.