18.05.2020
Real-time tracking of self-reported symptoms to predict potential COVID-19
Diagnostic Transversal
Menni C et al
Nature Med
Main result
- Of the 2,618,864 users of the symptom monitoring application, 32.2% reported at least 1 symptom potentially indicative of covid-19 involvement and 18,401 (0.7%) were tested and were able to report results. Of these individuals tested, 35% tested positive and 65% tested negative.
- 64.76% of the participants who tested positive had reported a loss of sense of smell compared to only 22.68% in the negative group (corresponding to an odds ratio of 6.40 after adjustment for age, sex and BMI).
- A model is proposed to predict, based on the presence or absence of a combination of symptoms (loss of smell and taste, fatigue, persistent cough and loss of appetite) whether a person is COVID-19 positive or not. The model thus constructed has a sensitivity of 0.65, a specificity of 0.78 and an area under the curve of 0.76. The implementation of the predictive model on all users of the application leads to an estimate of 17.42% of infected people among the users of the application.
- The authors also related the predictive character of the loss of taste for a diagnosis of COVID-19 to the media relay of the specificity of this symptom. It seems that patients' awareness of their loss of smell and taste, as well as their predictive character, increases with the media exposure of this symptom in the UK, but not in the USA.
Takeaways
Loss of taste and smell appear to be symptoms that are particularly correlated with the contraction of COVID-19.
There is a formula for assessing the likelihood of a patient being contaminated with COVID-19, based on self-diagnosis of a number of symptoms.
There is a formula for assessing the likelihood of a patient being contaminated with COVID-19, based on self-diagnosis of a number of symptoms.
Strength of evidence Weak
All results are based on self-diagnosed symptoms. In addition, training data for the model are relatively sparse (only 0.7% of participants were actually tested) and the participants (especially those tested by PCR) are not representative of the general population (non-randomized trial).
Even more: the tests were naturally performed on patients with the most severe symptoms.
Objectives
Produce a model to estimate the number of individuals with COVID-19 among the 2,618,862 participants who reported their symptoms on an application. This model is based on data from the 18,401 subjects who were also tested for SARS-CoV-2.
Method
A set of data for monitoring symptoms - or the absence of symptoms - self-reported by the 2,618,862 patients, US and British, via a mobile application is studied. The monitoring is updated daily and includes symptoms, hospitalizations, PCR tests and medical history.
A logistic regression is performed to identify the prevalence of patients with anosmia among those who tested positive versus those who tested negative. Logistic regression is reapplied to determine which other set of symptoms might be associated and indicative of an SARS-CoV-2 infection. Age, gender, loss of smell and taste, fatigue, persistent cough and loss of appetite are therefore retained.
The model to assess the likelihood of being infected based on symptoms and age is calibrated and tested on two subsets of the cohort of patients who actually underwent PCR testing.
A logistic regression is performed to identify the prevalence of patients with anosmia among those who tested positive versus those who tested negative. Logistic regression is reapplied to determine which other set of symptoms might be associated and indicative of an SARS-CoV-2 infection. Age, gender, loss of smell and taste, fatigue, persistent cough and loss of appetite are therefore retained.
The model to assess the likelihood of being infected based on symptoms and age is calibrated and tested on two subsets of the cohort of patients who actually underwent PCR testing.
bibliovid.org and its content are bibliovid property.