An artificial intelligence-based first-line defence against COVID-19: digitally screening citizens for risks via a chatbot

Alistair Martin; Jama Nateqi; Stefanie Gruarin; Nicolas Munsch; Isselmou Abdarahmane; Bernhard Knapp

Below we compare Symptoma to alternative approaches used to predict the likelihood of bearing COVID-19. Using a restricted set of symptoms as input, the probability that a user is suffering from COVID-19 in comparison to either influenza, the common cold or hay fever is calculated based on their respective literature-derived symptom frequencies (see Table S1 ). We implemented four different methods: the distance in symptom-space between case presentation and symptom frequency (SF-DIST), the distance normalised by the standard deviation (SF-SD), the distance normalised by the first principal component (SF-PCA) and the cosine similarity (SF-COS). These methods are described in detail within the Methods.