All Times ET
Keywords: COVID-19, Bayesian Modelling, Social Media, Conceptualization
A method will be presented, which allows us to track diseases by monitoring the mentions of symptoms on social media. We demonstrate this method by tracking COVID-19 in the Netherlands. Messages mentioning three symptoms were collected from social media: fever, cough, and sore throat. Each message was evaluated if it concerned a real symptom (e.g.: "I have a terrible cough") or not (e.g.: "I will cough on purpose when somebody gets too close to mee") by using a binary classifier. The idea behind our approach is that, when a disease with certain symptoms spreads, the number of messages for each symptom will rise proportionally. A Bayesian model was used to estimate the number of messages mentioning symptoms, probably caused by COVID-19. In this model, the hazard rate of COVID-19 was modeled as a log-Gaussian random walk and was split into three underlying hazard rates: one for cough mentions, one for sore throat mentions, and one for fever mentions. The messages, mentioning the symptoms, were modeled as three individual Poisson processes. Results of this COVID-19 indicator were compared to the Dutch national data and suggested that this social media indicator could be a fast indicator COVID-19 in the Netherlands. The model will be extended to more symptoms to see if we can distinguish different diseases.