Table 2.

Recommendations relating to the role of social media data in pharmacovigilance

Recommendation	Rationale
Social media should not be used as a source of ICSRs	With the exception of posts made by patients, carers and healthcare professionals on pharmaceutical company websites that make explicit mention of adverse events, the use of social media data for pharmacovigilance is secondary to the original intended use of these data [10, 14]. Although some posts may give detailed descriptions of an adverse event, the vast majority of posts lack the detail required for meaningful evaluation. Furthermore, large volumes of generally poor quality, non-informative data from social media should not be used to generate ICSRs since this has the potential to negatively impact signal detection systems [15]
Facebook and Twitter are currently not worthwhile to employ for the purpose of broad-ranging statistical signal detection at the expense of other pharmacovigilance activities	Applying disproportionality-based signal detection algorithms to automatically annotated Twitter/Facebook posts did not result in any predictive ability against two reference sets of signals and non-signals, in contrast to applying disproportionality analysis to VigiBasea cases. In addition, neither the first detected Twitter or Facebook posts nor the first occurrence of disproportionality in these sources would precede the actual time point of signalling, whereas in VigiBase this was more frequent, thereby negating any timing advantage of social media. This same lack of predictive ability was encountered with a relatively small sample of patient forum posts [15]
Future research should explore the value of social media as a source of information for additional cases in signal refinement/evaluation of ADRs that may significantly affect a patient’s quality of life	Approximately 12% of posts inspected in WEB-RADR contained information relevant to quality-of-life issues, e.g. lack of sleep, anxiety etc. [15]. This was an average across 38 medical products; however, further analyses (unpublished) indicate that drugs in the neuropsychiatric area have much higher proportions of mentions with quality-of-life issues
If social media is considered for use in pharmacovigilance, it is recommended that a prior assessment of the absolute and relative number of available posts related to the drug and/or event of interest in different online sources is made in relation to its intended use	There is substantial variation across drugs and adverse event terms in the amount of information in social media as well as substantial variation across different social media sources. Of the 38 medicinal products included in the WEB-RADR signal detection reference set, the range of substance mentions was from five (ranibizumab) to approximately 24,000 (methylphenidate) over a 3-year period (1 March 2012 to 31 March 2015)—see Fig. 2 Within the data collected prospectively for WEB-RADR (acquired from September 2014 through September 2017b), products with orphan or oncology-related indications were more likely to have higher volumes of posts describing potential AEs in patient fora than in Twitter (ruxolitinib had 3 × more posts describing potential AEs in fora, nilotinib had 8 ×, tobramycin 70 × and anastrozole 85 ×). Products with psychiatric indications were more likely to have a higher volume of posts in general, as well as a higher volume of posts describing potential AEs and mentions in Twitter than in patient fora (methylphenidate – 1.5 × more posts describing AEs in Twitter, zolpidem 7 ×) [16]
Further research should be carried out to determine whether there is value in social media data for niche areas of pharmacovigilance	WEB-RADR has demonstrated that there are niche areas of pharmacovigilance where social media data are more plentiful [17, 18] and can complement more traditional sources. For example, there is significant discussion about drug use in pregnancy [34] on social media to suggest that a combination of spontaneous reports and social media is likely to result in improved signal detection. However, the performance of this combined spontaneous/social media approach in specific areas is yet to be demonstrated as value-added relative to spontaneous reporting alone. In order to investigate this relative performance, additional work in algorithms and representative reference sets is needed
Consider using a predictive algorithm to identify and eliminate from the search query any medicinal product names with high levels of ambiguity to optimise time efficiency and, where applicable, cost effectiveness	The study by Hedfors et al. [13] showed that this could decrease the number of search terms by 67% and the number of extracted social media posts by 78%, with an associated increase in precision from 21.4% to 98.6% at a loss of only 0.9% of all relevant social media posts

ADRs adverse drug reactions, AEs adverse events, ICSRs individual case safety reports

^aVigiBase is the World Health Organisation’s (WHO) global database of ICSRs maintained by the Uppsala Monitoring Centre, Uppsala, Sweden. It is the largest database of its kind in the world, with over 19 million reports of suspected adverse effects of medicines submitted since 1968 by member countries of the WHO Programme for International Drug Monitoring

^bIn fact, data collection continued until December 2017; however, only data through September 2017 were included in the final report