Web and Social Media Listening Using AI/NLP for COVID-19 Vaccine Safety Monitoring

Background: While traditional Adverse Event (AE) reporting systems are slow to adapt on online AE reporting from patients, the number of patients using social media to share their experiences with vaccines is increasing. As global mass COVID-19 vaccination campaigns are ongoing, social media using artificial intelligence (AI)/natural language processing (NLP) is an important way to monitor AEs.

Objectives: To monitor AE shared in social media by those who received COVID-19 vaccines.

Methods: An AI powered NLP web app was developed to analyze web content (social media, and patient blogs and forums excluding press-release) from 190 countries in 61 languages for mentions of COVID-19 vaccine related keywords. Following machine translation to English, lay language safety terms are detected using PubmedBERT based named-entity recognition model (precision = 0.76, recall = 0.82) which are mapped to Medical Dictionary for Regulatory Activities (MedDRA) terms using knowledge graphs. Insights were then generated using weekly and cumulative aggregated counts, proportions and ratios. Visual analytics as word clouds, tabulations and line graphs were used to analyze across periods any trending by vaccine brand, mechanism, country and MedDRA terms level (PT, HLT, SOC).

Results: Selected AE examples found using the app were consistent with those communicated shortly before or concomitantly by different Health Authorities. Anaphylactic reactions started to trend mid-December 2020 (under HLT "Anaphylactic and anaphylactoid responses") shortly after the introduction of mRNA COVID-19 vaccines. Thrombo-embolic events started to trend on 13 March 2021 for adenovirus-vectored COVID-19 vaccine with popping up of different HLTs including Pulmonary thrombotic and embolic conditions, Non-site specific embolism and Thrombosis. Myocarditis/pericarditis started to trend for mRNA COVID-19 vaccines on 1 May 2021 with an increased ratio (proportion of Non-infectious myocarditis HLT of the current period as compared to previous one) on 29 May 2021. The proportion of Myocarditis (count of Myocarditis HLT reported to the count of all HLTs within the period) remains the highest since September 2021.

Conclusions: Web and social media listening and monitoring using AI/NLP provide opportunities to detect both safety insights as well as consumer concerns around COVID-19 vaccines safety efficiently. This tool demonstrates AI/NLP capability to detect early safety insights from consumers at scale. It could help to adjust signal detection strategies and communication to external stakeholders, contributing to increase confidence in vaccines' monitoring and safety.