Fig. 1.

Data processing results. Public social media data were obtained for Facebook and Twitter, with 935,246 posts identified where each of the 10 products under review were mentioned. An automated classifier was used to remove spam, followed by exclusion of verbatim duplicate posts. Posts with an indicator score of ≥0.65 were considered a ‘post with resemblance to an adverse event’ and manually curated to remove false positives. The resulting 13 posts that mentioned product–event associations from alerts or negative controls were assessed for causality by an independent third party. Six posts contained certain, probable, and possible cases. AE adverse event, Proto-AE post with resemblance to an adverse event