Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=ieds20
Expert Opinion on Drug Safety
ISSN: 1474-0338 (Print) 1744-764X (Online) Journal homepage: http://www.tandfonline.com/loi/ieds20
The usefulness of listening social media for
pharmacovigilance purposes: a systematic review
Irma Convertino, Sara Ferraro, Corrando Blandizzi & Marco Tuccori
To cite this article: Irma Convertino, Sara Ferraro, Corrando Blandizzi & Marco Tuccori (2018):
The usefulness of listening social media for pharmacovigilance purposes: a systematic review,
Expert Opinion on Drug Safety, DOI: 10.1080/14740338.2018.1531847
To link to this article: https://doi.org/10.1080/14740338.2018.1531847
Accepted author version posted online: 04
Oct 2018.
Submit your article to this journal
View Crossmark data
Accepted Manuscript
Publisher: Taylor & Francis
Journal: Expert Opinion on Drug Safety
DOI: 10.1080/14740338.2018.1531847
Article type: review
The usefulness of listening social media for pharmacovigilance purposes: a systematic review
Irma Convertino
1
, Sara Ferraro
1
, Corrando Blandizzi
1,2
,
Marco Tuccori
2
1
Unit of Pharmacology and Pharmacovigilance, University of Pisa, Pisa, Italy
2
Unit of Adverse Drug Reaction Monitoring, University Hospital of Pisa, Pisa, Italy
Corresponding author
Marco Tuccori
Unit of Adverse Drug Reaction Monitoring, University Hospital of Pisa, Medical School, Via Roma 55, 56126,
Pisa, Italy
Email: m.tuccori@ao-pisa.toscana.it
Accepted Manuscript
Abstract
Introduction: Social media mining could be a possible strategy to retrieve drug safety information. The
mining of social media is a complex process under progressive evolution, falling into three broad
categories: listening (safety data reporting), engaging (follow-up) and broadcasting (risk communication).
This systematic review is aimed at evaluating the usefulness and quality of proto-signals by social media
listening.
Areas covered: In this systematic search, performed according to MOOSE and PRISMA statements, we
selected studies, published in MEDLINE, EMBASE and Google Scholar until December 31
st
, 2017, that
listened at least one social media to identify proto-adverse drug events and proto-signals.
Expert opinion: The selected thirty-eight studies identified serious and unexpected proto-adverse drug
events characterized by poorer information quality as compared with spontaneous reporting databases.
This feature allows rarely the evaluation of causal relationships. Proto-signals identified by social media
listening had the potential of anticipating pre-specified known signals in only six studies. Moreover, the
personal perception of patients reported in social media could be used to implement effective risk
communication strategies. However, signal detection in social media cannot be currently recommended for
routine pharmacovigilance, due to logistic and technical issues.
Keywords: social media, pharmacovigilance, proto-signal, signal detection
Article highlights
• Social media listening has the potential of identifying serious and unexpected adverse drug events
• Limited evidence suggest that social media can identify some signals earlier than the traditional
pharmacovigilance approaches
• The poor quality of information provided in social media comments, rarely allows the evaluation of causal
relationships as compared with spontaneous reporting databases. Therefore, drug-event pairs will often
need to be corroborated by alternative data sources
• Signal detection in social media cannot be currently recommended for routine pharmacovigilance, due to
logistic and technical issues
• In the future, the standardization of methodologies and the development of technologies could make the
social media promising sources of information suitable for supporting pharmacovigilance
Accepted Manuscript
1. Introduction
The widespread use of social media is likely one of the most revolutionary global phenomenon
occurred after 2000. Over those years, social media proliferated and differentiated, creating large people
communities sharing passions, needs and interests. The possibility of giving access via web to any kind of
personal information, not only to the circle of friends and relatives, but even to unknown peoples, became
a sort of appealing highway to popularity. The toll to be payed to take this highway is often the renounce,
at least in part, to privacy. As a consequence, the information shared on social media has become a fruitful
goldmine for data owners, since they may represent an easy way to achieve personalized advertising. Social
media started to be mined initially for commercial purposes and later for other purposes, like political
ones
1
. Mining strategies and methodologies were developed and evolved in parallel with the differentiation
of these scopes
2
.
The Council for International Organizations of Medical Sciences (CIOMS) defined a “signal” as an
information that arises from one or multiple sources (including observations and experiments), which
suggest a putative causal association, or a new aspect of a known association, between a medical
intervention and an event or set of related events, either adverse or beneficial, that is judged to be of
sufficient likelihood to justify verification actions
3
. Theoretically, social media could be considered a source
of data that might be exploited to retrieve drug safety information
2,4,5
. Several investigations, such as the
“WEB-Recognising Adverse Drug Reactions” (Web-RADR), have been then implemented with the aim of
providing some directives on 'what and how' to use social media to further proactive pharmacovigilance
and protection of public health
6
. The Web-RADR team coined the term “proto-adverse event” (proto-AE) in
order to signify “posts with resemblance to AEs’, designating posts containing discussion of AEs identified in
social media sources” and to distinguish them from the traditional pharmacovigilance regulatory
definition
7
.
The mining of social media is a complex process under progressive evolution that falls into three
broad categories: listening (safety data reporting), engaging (follow-up) and broadcasting (risk
communication). Therefore, reviews highlighting deeply the state of the art on the progress of related
methodologies
8
, ethical aspects
9
and reliability
10
are frequently required. Of note, the usefulness and
quality of proto-signals generated by social media listening remain questionable
11
, and a systematic review
is lacking. In this paper, we reviewed systematically current medical literature, searching for articles that
attempted a detection of proto-signals by social media mining, with the aim of evaluating the usefulness of
listening social media for pharmacovigilance purposes. This review is aimed at: 1) attempting to summarize
different data mining strategies; 2) assessing whether the information quality (i.e. seriousness, notoriety,
causality, clinical features) of detected drug-event pairs may be affected by the different scopes of social
media; 3) evaluating whether the different methodological approaches used in the studies were able to
identify proto-signals that could anticipate the safety warnings issued by Health Authorities.
2. Methods
2.1 Data sources
This systematic review was performed in accordance with PRISMA
12
and MOOSE
13
statements.
Studies were selected by using MEDLINE, EMBASE, and Google Scholar.
2.2 Definitions
In this review, social media are intended as computer-based technologies that allow the creation
and sharing of information, ideas, interests, and other forms of expression via virtual communities as social
networks and forums. Users create service-specific profiles for the website or app that are designed and
maintained by the social media organization. The network is populated spontaneously by user-generated
contents, and may vary from texts to photos and videos. We further classified social media into two large
groups, based on their purposes: medical (e.g. askapatient.com) and non-medical (e.g. Facebook, Twitter,
and Google Plus). Each post containing a drug-event pair was analysed as a spontaneous report. In this
article, a proto-ADE is defined as a drug-event that emerges in a post published on social media. As well, a
proto-signal is identified when an abnormal frequency distribution of a proto-ADE is identified in social
media (either by considering a defined threshold or not).
Accepted Manuscript
2.3 Search strategy
We conducted a systematic search of studies using the above-mentioned data sources by a
combination of the keywords “Pharmacovigilance” and “Social Media”. The research was performed in
English language without limits of time up to December 31
st
, 2017. The reference list of selected studies
was also checked for additional relevant articles. Duplicates were removed firstly by Mendeley auto-
deduplication tool
14
, and lastly by manual assessment. Two different reviewers (I.C., S.F.) examined the
retrieved papers. The relevance of studies was evaluated by the title and the abstract. If the study eligibility
remained unclear, the full text was checked. Any disagreement was resolved by discussion with a senior
reviewer (M.T.).
2.4 Study inclusion & exclusion criteria
We included studies performed retrospectively on dataset obtained by at least one identifiable
social media (i.e. by name or URL), in accordance with the definition of social media given above. Review
articles, sentiment analysis only
15
, studies that provided only methodological assessments of text mining
strategies, and studies performed on search engines only as data sources were excluded. Notably, when a
study used indifferently search engines and social media as data sources, we included it only if data
resulting from social media listening were provided separately. Studies in which at least one proto-ADE was
not identifiable or for which at least a simple numerical frequency for proto-ADEs was not provided were
excluded, as well. Studies performed on proto-ADEs following immunization (vaccines) were not included.
Finally, we excluded studies in which the reporting of an ADE was prospectively stimulated by means of
questionnaire or by means of ad hoc designed forms (active surveys) promoted by the social media
managers.
2.5 Data extraction
From each selected study, the following information were collected: authors, year of publication,
social media data sources, study population (all users or sub-groups of users and their features), outcomes
(identified proto-ADE, clinical features and causality assessment), method of data extraction, aims (survey
or signal detection) and conclusion key-points.
2.6 Proto-ADEs classification
Proto-ADEs seriousness, notoriety (expectedness) and causality assessment were recorded when
provided by the selected articles. When such information were not reported, seriousness and notoriety
were classified according to the European Medicines Agency (EMA) Important Medical Event list of the
Medical Dictionary for Regulatory Activities (MedDRA)
16
and the related summary of product
characteristics, respectively. Proto-ADEs were clustered according to clinical features identified by means at
least of the system organ class (SOC) that was extracted by the original article or attributed by the MedDRA
classification. Drugs and drug classes were classified, when necessary, according to their Anatomical
Therapeutic Chemical (ATC) classification groups.
2.7 Methodology classification
Study methodologies were analysed and classified by strategy for the selection of the primary data
source, and by design (drug-based approach when the identification of proto-ADE begins with the
identification of posts containing the name of the drug/drugs of interest, or event-based approach, when
the posts are initially selected by the presence of at least one event). A description of their approach to the
identification of proto-ADEs is also provided based on the dictionaries, lexicon creation, and strategies for
identification of a semantic association. Finally, we classified the studies, based on their approach to signal
assessment, in descriptive (the authors provided a simple frequency of the event in the posts of the social
media) or analytic (the authors identified proto-signals by using a disproportionality approach such as
proportional reporting ratio, or measures of semantic association such as “lift” or “leverage”).
3. Results
The selection flowchart is displayed in figure 1. Based on the inclusion and exclusion criteria, thirty-
eight studies were identified. Nine studies provided information about the age of users/patients included in