Background/Introduction: The large use of social media leads to their data mining not only for commercial and political purposes but also for pharmacovigilance (PV) ones [1, 2]. Data mining in social media included: listening (safety data reporting), engaging (followup), and broadcasting (risk communication). Aim: To review evidences on data mining strategies in social media, to assess the quality of information, and to evaluate the forecasting power of social media compared with the safety warnings issued by Health Authorities. Methods: We reviewed English studies published up to December 31st, 2017 in MEDLINE, EMBASE, and Google Scholar in accordance with PRISMA and MOOSE statements. We included studies: investigating the frequency of proto-adverse drug events (ADEs) or proto-signals in social media, and reporting at least one identifiable data source. We excluded: reviews, prospective studies, sentiment analyses, text mining, and studies using search engines only and focusing on events following immunization. We evaluated seriousness, notoriety, and causality of proto-ADEs for assessing the quality. Results: Out of included 38 studies, 9 reported users’ age, 8 the geographical areas, 16 provided notoriety, and 18 used a single social media as data source. The methodological approaches can be classified into three sequential steps. The first was the selection of posts, performed as a drug-based approach (n. = 37), or an event-based one (n. = 1). The second step dealt with the identification of proto-ADEs in social media, classified as: non-medical social media (n. = 11), medical social media (n. = 26), and both of them (n. = 1). Studies performed in non-medical social media investigated specific drug classes (e.g. antiretroviral, antidiabetic, anti-rheumatic, antipsychotics, antidepressants, glucocorticoids) (n. = 8) or drugs most frequently mentioned in the posts (n. = 3). Studies carried out in medical social media focused on general medical contents (n. = 19), specific diseases (n. = 2) or both subjects (n. = 5). Serious and unexpected proto-ADEs were detected in 21 and 15 studies, respectively. The third step included the identification of proto-signals through a signal detection performed as descriptive analysis or disproportiona analysis or semantic association. Out of these, 8 studies assessed the forecasting power of social media and only 6 showed it. Conclusions: Listening social media has the potential of identifying serious and unexpected ADEs and forecasting signals. However, the poor quality of information allows rarely assessing the causality as compared with spontaneous reporting databases. The patients’ perception reported in social media could be useful for risk communication strategies. On the contrary, signal detection in social media has logistic and technical limits for suggesting the usage in routine PV.
Listening Social Media for Pharmacovigilance Objectives: Which Usefulness?
Convertino, I;Leonardi, L;Tuccori, M;Ferraro, S;Blandizzi, C
2019-01-01
Abstract
Background/Introduction: The large use of social media leads to their data mining not only for commercial and political purposes but also for pharmacovigilance (PV) ones [1, 2]. Data mining in social media included: listening (safety data reporting), engaging (followup), and broadcasting (risk communication). Aim: To review evidences on data mining strategies in social media, to assess the quality of information, and to evaluate the forecasting power of social media compared with the safety warnings issued by Health Authorities. Methods: We reviewed English studies published up to December 31st, 2017 in MEDLINE, EMBASE, and Google Scholar in accordance with PRISMA and MOOSE statements. We included studies: investigating the frequency of proto-adverse drug events (ADEs) or proto-signals in social media, and reporting at least one identifiable data source. We excluded: reviews, prospective studies, sentiment analyses, text mining, and studies using search engines only and focusing on events following immunization. We evaluated seriousness, notoriety, and causality of proto-ADEs for assessing the quality. Results: Out of included 38 studies, 9 reported users’ age, 8 the geographical areas, 16 provided notoriety, and 18 used a single social media as data source. The methodological approaches can be classified into three sequential steps. The first was the selection of posts, performed as a drug-based approach (n. = 37), or an event-based one (n. = 1). The second step dealt with the identification of proto-ADEs in social media, classified as: non-medical social media (n. = 11), medical social media (n. = 26), and both of them (n. = 1). Studies performed in non-medical social media investigated specific drug classes (e.g. antiretroviral, antidiabetic, anti-rheumatic, antipsychotics, antidepressants, glucocorticoids) (n. = 8) or drugs most frequently mentioned in the posts (n. = 3). Studies carried out in medical social media focused on general medical contents (n. = 19), specific diseases (n. = 2) or both subjects (n. = 5). Serious and unexpected proto-ADEs were detected in 21 and 15 studies, respectively. The third step included the identification of proto-signals through a signal detection performed as descriptive analysis or disproportiona analysis or semantic association. Out of these, 8 studies assessed the forecasting power of social media and only 6 showed it. Conclusions: Listening social media has the potential of identifying serious and unexpected ADEs and forecasting signals. However, the poor quality of information allows rarely assessing the causality as compared with spontaneous reporting databases. The patients’ perception reported in social media could be useful for risk communication strategies. On the contrary, signal detection in social media has logistic and technical limits for suggesting the usage in routine PV.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
