Documents

17 pages
97 views

Researching Mental Health Disorders in the Era of Social Media_ Systematic Review

Please download to get full document.

View again

of 17
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Share
Description
Social Media
Transcript
  Review Researching Mental Health Disorders in the Era of Social Media:Systematic Review Akkapon Wongkoblap 1 , BSc, MSc; Miguel A Vadillo 2,3 , PhD; Vasa Curcin 1,2 , PhD 1 Department of Informatics, King's College London, London, United Kingdom 2 Primary Care and Public Health Sciences, King’s College London, London, United Kingdom 3 Departamento de Psicología Básica, Universidad Autónoma de Madrid, Madrid, Spain Corresponding Author: Akkapon Wongkoblap, BSc, MScDepartment of InformaticsKing's College LondonStrandLondon, WC2R 2LSUnited KingdomPhone: 44 20 7848 2588Fax: 44 20 7848 2017Email: akkapon.wongkoblap@kcl.ac.uk  Abstract  Background: Mental illness is quickly becoming one of the most prevalent public health problems worldwide. Social network  platforms, where users can express their emotions, feelings, and thoughts, are a valuable source of data for researching mentalhealth, and techniques based on machine learning are increasingly used for this purpose. Objective: The objective of this review was to explore the scope and limits of cutting-edge techniques that researchers are usingfor predictive analytics in mental health and to review associated issues, such as ethical concerns, in this area of research. Methods: We performed a systematic literature review in March 2017, using keywords to search articles on data mining of social network data in the context of common mental health disorders, published between 2010 and March 8, 2017 in medicaland computer science journals. Results: The initial search returned a total of 5386 articles. Following a careful analysis of the titles, abstracts, and main texts,we selected 48 articles for review. We coded the articles according to key characteristics, techniques used for data collection,data preprocessing, feature extraction, feature selection, model construction, and model verification. The most common analyticalmethod was text analysis, with several studies using different flavors of image analysis and social interaction graph analysis. Conclusions: Despite an increasing number of studies investigating mental health issues using social network data, somecommon problems persist. Assembling large, high-quality datasets of social media users with mental disorder is problematic, notonly due to biases associated with the collection methods, but also with regard to managing consent and selecting appropriateanalytics techniques. (J Med Internet Res 2017;19(6):e228) doi:10.2196/jmir.7215 KEYWORDS mental health; mental disorders; social networking; artificial intelligence; machine learning; public health informatics; depression;anxiety; infodemiology Introduction  Mental illness is quickly becoming one of the most serious and prevalent public health problems worldwide [1]. Around 25%of the population of the United Kingdom have mental disordersevery year [2]. According to statistics published by the WorldHealth Organization, more than 350 million people havedepression. In terms of economic impact, the global costs of mental health problems were approximately US $2.5 trillion in2010. By 2030, it is estimated that the costs will increase further to US $6.0 trillion [3]. Mental disorders include many differentillnesses, with depression being the most prominent.Additionally, depression and anxiety disorders can lead tosuicidal ideation and suicide attempts [1]. These figures show J Med Internet Res 2017 | vol. 19 | iss. 6 | e228 | p.1http://www.jmir.org/2017/6/e228/ (page number not for citation purposes) Wongkoblap et alJOURNAL OF MEDICAL INTERNET RESEARCH XSL ã FO RenderX  that mental health problems have effects across society, anddemand new prevention and intervention strategies. Earlydetection of mental illness is an essential step in applying thesestrategies, with the mental illnesses typically being diagnosedusing validated questionnaires designed to detect specific patterns of feelings or social interaction [4-6]. Online social media have become increasingly popular over thelast few years as a means of sharing different types of user-generated or user-curated content, such as publishing personal status updates, uploading pictures, and sharing currentgeographical locations. Users can also interact with other users by commenting on their posts and establishing conversations.Through these interactions, users can express their feelings andthoughts, and report on their daily activities [7], creating awealth of useful information about their social behaviors [8].To name just 2 particularly popular social networks, Facebook is accessed regularly by more than 1.7 billion monthly activeusers [9] and Twitter has over 310 million active accounts [10],  producing large volumes of data that could be mined, subjectto ethical constraints, to find meaningful patterns in users’ behaviors.The field of data science has emerged as a way of addressingthe growing scale of data, and the analytics and computational power it requires. Machine learning techniques that allowresearchers to extract information from complex datasets have been repurposed to this new environment and used to interpretdata and create predictive models in various domains, such asfinance [11], economics [12], politics [13], and crime [14]. In medical research, data science approaches have allowedresearchers to mine large health care datasets to detect patternsand accrue meaningful knowledge [15-18]. A specific segment of this work has focused on analyzing and detecting symptomsof mental disorders through status updates in social networkingwebsites [19].Based on the symptoms and indicators of mental disorders, itis possible to use data mining and machine learning techniquesto develop automatic detection systems for mental health problems. Unusual actions and uncommon patterns of interactionexpressed in social network platforms [19] can be detectedthrough existing tools, based on text mining, social network analysis, and image analysis.Even though the current performance of predictive models issuboptimal, reliable predictive models will eventually allowearly detection and pave the way for health interventions in theforms of promoting relevant health services or delivering usefulhealth information links. By harnessing the capabilities offeredto commercial entities on social networks, there is a potentialto deliver real health benefits to users.This systematic review aimed to explore the scope and limitsof cutting-edge techniques for predictive analytics in mentalhealth. Specifically, in this review we tried to answer thefollowing questions: (1) What methods are researchers usingto collect data from online social network sites such as Facebook and Twitter? (2) What are the state-of-the-art techniques in predictive analytics of social network data in mental health? (3)What are the main ethical concerns in this area of research? Methods  We conducted a systematic review to examine how social mediadata have been used to classify and predict the mental healthstate of users. The procedure followed the guidelines of thePreferred Reporting Items for Systematic Reviews andMeta-Analyses (PRISMA) to outline and assess relevant articles[20]. Literature Search Strategy We searched the literature in March 2017, collecting articles published between 2010 and March 8, 2017 in medical andcomputer science databases. We searched PubMed, Institute of Electrical and Electronics Engineers (IEEE Xplore), Associationfor Computing Machinery (ACM Digital Library), Web of Science, and Scopus using sets of keywords focused on the prediction of mental health problems based on data from socialmedia. We restricted our searches to common mental healthdisorders, as defined by the UK National Institute for Healthand Care Excellence [21]: depression, generalized anxietydisorder, panic disorder, phobias, social anxiety disorder,obsessive-compulsive disorder (OCD), and posttraumatic stressdisorder (PTSD). To ensure that our literature search strategywas as inclusive as possible, we explored Medical SubjectHeadings (MeSH) for relevant key terms. MeSH terms wereused in all databases that made this option available. Searchterms are outlined in Textbox 1.In addition, we manually searched the proceedings of theComputational Linguistics and Clinical Psychology Workshops(CLPsych) and the outputs of the World Well-Being Project[22] to find additional articles that our search terms might haveexcluded. Furthermore, we examined the reference lists of included articles for additional sources. Textbox 1. Search strategy to identify articles on the prediction of mental health problems based social media data.Medical Subject Headings (MeSH)1.Depression/ or Mental Health/ or Mental Disorders/ or Suicide or Life Satisfaction/ or Well Being/ or Anxiety/ or Panic/ or Phobia/ or OCD/ or PTSD2.Social Media/ or Social Networks/ or Facebook/ or Twitter/ or Tweet3.Machine Learning/ or Data Mining/ or Big Data/ or Text Analysis/ or Text Mining/ or Predictive Analytics/ or Prediction/ or Detection/ or DeepLearning4.(1) and (2)5.(1) and (3) J Med Internet Res 2017 | vol. 19 | iss. 6 | e228 | p.2http://www.jmir.org/2017/6/e228/ (page number not for citation purposes) Wongkoblap et alJOURNAL OF MEDICAL INTERNET RESEARCH XSL ã FO RenderX  Inclusion and Exclusion Criteria We further filtered the titles and abstracts of articles retrievedusing the search terms outlined in Textbox 1. Only articles published in peer-reviewed journals and written in English wereincluded. Further inclusion criteria were that studies had to (1)focus on predicting mental health problems through social mediadata, and (2) investigate prediction or classification models based on users’ text posts, network interactions, or other featuresof social network platforms. Within this review, we focused onsocial network platforms—that is, those allowing users to create personal profiles, post content, and establish new or maintainexisting relationships.Studies were excluded if they (1) only analyzed the correlation between social network data and symptoms of mental illness,(2) analyzed textual contents only by human coding or manualannotation, (3) examined data from online communities (eg,LiveJournal), (4) focused on the relationship between socialmedia use and mental health disorders (eg, so-called Internetaddiction), (5) examined the influence of cyberbullying onmental health, or (6) did not explain where the datasets camefrom. Data Extraction After screening articles and obtaining a set of studies that metour inclusion criteria, we extracted the most relevant data fromthe main texts. These are title, author, aims, findings, methods,data collection on machine learning techniques, sampling,questionnaire, platform, and language. Results  Overview Figure 1 presents a PRISMA flow diagram of the results of searching and screening articles following the above searchmethodology. The initial search resulted in a total of 5371articles plus 11 additional articles obtained through CLPsych,1 from the World Well-Being Project, and 3 from the referencelists of included articles. We removed 1864 of these articles because of duplication. Each of the remaining articles (n=3522)was screened by reviewing its title and abstract. If an articleanalyzed data from other sources (such as brain signals, mentalhealth detection from face detection, or mobile sensing), wediscarded it. This resulted in a set of 106 articles. By matchingthese with our inclusion and exclusion criteria, we removed afurther 58 articles. To sum up, we excluded 5338 articles andincluded 48 in the review (see Figure 1).We extracted data from each of the 48 articles. Table 1andMultimedia Appendix 1(whose format is adapted from previouswork [11,23]) show the key characteristics of the selected studies [24-71], ordered by year published. Of the studies reviewed, 46 were published from 2013 onward, while only 2 peer-reviewedarticles were published between 2011 and 2012. None of theselected articles was published in 2010. Figure 1. Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram. CLPsych: Computational Linguistics andClinical Psychology Workshops. J Med Internet Res 2017 | vol. 19 | iss. 6 | e228 | p.3http://www.jmir.org/2017/6/e228/ (page number not for citation purposes) Wongkoblap et alJOURNAL OF MEDICAL INTERNET RESEARCH XSL ã FO RenderX  Table 1. Summaries of articles reviewed.FindingsAimsFirst author,date, referenceThere was assortativity among users with eating disorder. Theclassifier distinguished 2 groups of people.To explore and characterize the structure of the community of  people with eating disorders using Twitter data and then classifyusers into those with and without the disorder.Wang, 2017[24]Tweets from students across 44 universities were related to studentsurveys on satisfaction and happiness.To explore academic discourse from tweets and build predictivemodels to analyze the data.Volkava, 2016[25]The proposed method and a classifier were built as an online sys-tem, which distinguished 2 groups of individuals and providedmental illness information.To present a new data collection method and classify individualswith mental illness and nonmental illness.Saravia, 2016[26]The models detected users with depression.To propose classification models to detect tweets of users withdepression for a long period of time. Classifiers were based onthe texts, emoticons, and images they posted.Kang, 2016[27]A combination of message- and user-level aggregation of posts performed well.To present predictive models to estimate individual well-beingthrough textual content on social networks.Schwartz, 2016[28]Future mental illness severity could be predicted from user-gener-ated messages.To explore posts from Instagram to forecast levels of mentalillness severity of pro-eating disorder.Chancellor,2016 [29]Machine learning algorithms successfully classified users withsuicidal ideation.To explore machine learning algorithms to measure suicide risk in the United States.Braithwaite,2016 [30]There were quantifiable signals of suicide attempt in tweets.To explore linguistics and emotional patterns in Twitter userswith and without suicide attempt.Coppersmith,2016 [31]The Chinese suicide dictionary detected individuals and tweets atsuicide risk.To build a Chinese suicide dictionary, based on Weibo posts,to detect suicide risk.Lv, 2015 [32]Machine learning classifiers estimated the level of concern fromsuicide-related tweets.To explore machine learning models to automatically detect thelevel of concern for each suicide-related tweet.O’Dea, 2015[33]Users’ subjective well-being could be predicted from posts andtheir time frame.To investigate and predict users’ subjective well-being basedon Facebook posts.Liu, 2015 [34]Classification models classified tweets into relevant suicide cate-gories.To explore suicide-related tweets to understand users’ commu-nications on social media.Burnap, 2015[35]Participants with depression had fewer interactions, such as receiv-ing likes and comments. Depressed users posted at a higher rate.To analyze the relationships between Facebook activities andthe depression state of users.Park, 2015 [36]Behavioral and linguistic features predicted depression. A 2-month period of observation enabled prediction cues of depression half a month in advance.To present classifiers with different lengths of observation timeto detect depressed users.Hu, 2015 [37]Activities extracted from Twitter were useful to detect depression;2 months of observation data enabled detection of symptoms of depression. The topics estimated by LDA a were useful.To develop a model to recognize individuals with depressionfrom non-English social media posts and activities.Tsugawa, 2015[38]LDA automatically detected suicide probability from textual con-tents on social media.To explore 2 natural language processing algorithms to identify posts predicting the probability of suicide.Zhang, 2015[39]There were quantifiable signals of 10 mental health conditions insocial network messages and relations between them.To explore tweet content with self-reported health sentencesand language differences in 10 mental health conditions.Coppersmith,2015 [40]The combination of linear classifiers performed better than averageclassifiers. All unigram features performed well.To implement linear classifiers to detect users with PTSD c anddepression based on user metadata, and several textual andtopic features.Preotiuc-Pietro,2015 [41]Character ngram features were used to train models to classifyusers with and without schizophrenia. LDA outperformed linguisticinquiry and word count.To use several natural language processing techniques to explorethe language of schizophrenic users on Twitter.Mitchell, 2015[42]Personality and demographic data extracted from tweets detectedusers with depression or PTSD.To study differences in language use in tweets about mentalhealth depending on the role of personality, age, and sex of users.Preotiuc-Pietro,2015 [43]Bigram features underperformed ngram 1-6 features.To explore and study the accuracy of decision lists of ngramsto classify users with depression and PTSD.Pedersen, 2015[44] J Med Internet Res 2017 | vol. 19 | iss. 6 | e228 | p.4http://www.jmir.org/2017/6/e228/ (page number not for citation purposes) Wongkoblap et alJOURNAL OF MEDICAL INTERNET RESEARCH XSL ã FO RenderX
Related Documents
View more...
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks
SAVE OUR EARTH

We need your sign to support Project to invent "SMART AND CONTROLLABLE REFLECTIVE BALLOONS" to cover the Sun and Save Our Earth.

More details...

Sign Now!

We are very appreciated for your Prompt Action!

x