Analysis of longitudinal social media for monitoring symptoms during a pandemic

Abstract

Objective

Current studies leveraging social media data for disease monitoring face challenges like noisy colloquial language and insufficient tracking of user disease progression in longitudinal data settings. This study aims to develop a pipeline for collecting, cleaning, and analyzing large-scale longitudinal social media data for disease monitoring, with a focus on COVID-19 pandemic.

Materials and methods

This pipeline initiates by screening COVID-19 cases from tweets spanning February 1, 2020, to April 30, 2022. Longitudinal data is collected for each patient, two months before and three months after self-reporting. Symptoms are extracted using Name Entity Recognition (NER), followed by denoising with a combination of Graph Convolutional Network (GCN) and Bidirectional Encoder Representations from Transformers (BERT) model to retain only User-experienced Symptom Mentions (USM). Subsequently, symptoms are mapped to standardized medical concepts using the Unified Medical Language System (UMLS). Finally, this study conducts symptom pattern analysis and visualization to illustrate temporal changes in symptom prevalence and co-occurrence.

Results

This study identified 191,096 self-reported COVID-19-positive cases from COVID-19-related tweets and retrospectively collected 811,398,280 historical tweets, of which 2,120,964 contained symptoms information. After denoising, 39 % (832,287) of symptom-sharing tweets reflected user-experienced mentions. The trained USM model achieved an average F1 score of 0.927. Further analysis revealed a higher prevalence of upper respiratory tract symptoms during the Omicron period compared to the Delta and Wild-type periods. Additionally, there was a pronounced co-occurrence of lower respiratory tract and nervous system symptoms in the Wild-type strain and Delta variant.

Conclusion

This study established a robust framework for analyzing longitudinal social media data to monitor symptoms during a pandemic. By integrating denoising of user-experienced symptom mentions, our findings reveal the duration of different symptoms over time and by variant within a cohort of nearly 200,000 patients, providing critical insights into symptom trends that are often difficult to capture through traditional data source.