Using Deep Learning Transformer Networks to Identify Symptoms Associated with COVID-19 on Twitter

Resumo

This study aims to present a methodology to identify, through Twitter posts, predefined symptoms
of COVID-19 aided by Deep Learning techniques, namely Transformers Networks. The proposed approach
was evaluated on a public Twitter database in Brazilian Portuguese, using user reports of COVID-19 symptoms.
We mine the Twitter database, extract phrases with symptoms, compare distributions, and build a database to
construct high accuracy Deep Learning networks, which can be used to identify symptoms. We use a crossvalidation
procedure to evaluate the result’s performance. Additionally, we interpret the results using a Local
Interpretable Model-Agnostic Explanations (LIME) algorithm. We identified 907 tweets containing one or more
of the 14 previously chosen COVID-19 symptoms. The most frequently reported symptoms were a cough (392),
headache (154), runny nose (143), fever (124), nausea (106), and diarrhea (105) amongst users who reported at
least one symptom. The BERT architecture identified all 14 symptoms reported in Twitter phrases in Portuguese,
resulting in identifying each symptom with over 97% accuracy and over 0.95 of AUC-ROC at the test dataset.
This project is a step towards a complementary tool to identify symptoms in future automated clinical settings,
e.g., medical chatbots, to support faster clinical assessment in Portuguese.