We searched the literature in March 2017, collecting articles published between 2010 and March 8, 2017 in medical and computer science databases. We searched PubMed, Institute of Electrical and Electronics Engineers (IEEE Xplore), Association for 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 social media. We restricted our searches to common mental health disorders, as defined by the UK National Institute for Health and Care Excellence [21]: depression, generalized anxiety disorder, panic disorder, phobias, social anxiety disorder, obsessive-compulsive disorder (OCD), and posttraumatic stress disorder (PTSD). To ensure that our literature search strategy was as inclusive as possible, we explored Medical Subject Headings (MeSH) for relevant key terms. MeSH terms were used in all databases that made this option available. Search terms are outlined in Textbox 1.

In addition, we manually searched the proceedings of the Computational 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 have excluded. Furthermore, we examined the reference lists of included articles for additional sources.

We further filtered the titles and abstracts of articles retrieved using the search terms outlined in Textbox 1. Only articles published in peer-reviewed journals and written in English were included. Further inclusion criteria were that studies had to (1) focus on predicting mental health problems through social media data, and (2) investigate prediction or classification models based on users’ text posts, network interactions, or other features of social network platforms. Within this review, we focused on social network platforms—that is, those allowing users to create personal profiles, post content, and establish new or maintain existing 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 manual annotation, (3) examined data from online communities (eg, LiveJournal), (4) focused on the relationship between social media use and mental health disorders (eg, so-called Internet addiction), (5) examined the influence of cyberbullying on mental health, or (6) did not explain where the datasets came from.

After screening articles and obtaining a set of studies that met our inclusion criteria, we extracted the most relevant data from the main texts. These are title, author, aims, findings, methods, data collection on machine learning techniques, sampling, questionnaire, platform, and language.

Figure 1 presents a PRISMA flow diagram of the results of searching and screening articles following the above search methodology. The initial search resulted in a total of 5371 articles plus 11 additional articles obtained through CLPsych, 1 from the World Well-Being Project, and 3 from the reference lists 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 article analyzed data from other sources (such as brain signals, mental health detection from face detection, or mobile sensing), we discarded it. This resulted in a set of 106 articles. By matching these with our inclusion and exclusion criteria, we removed a further 58 articles. To sum up, we excluded 5338 articles and included 48 in the review (see Figure 1).

We extracted data from each of the 48 articles. Table 1 and Multimedia Appendix 1 (whose format is adapted from previous work [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-reviewed articles were published between 2011 and 2012. None of the selected articles was published in 2010.

The selected studies can be divided into several distinct categories. Several studies [27,36-38,40-46,52,56-59,61-64, 66,67] used datasets from social networks to examine depression. Postpartum depression disorder was explored by De Choudhury et al [50,65], PTSD was investigated by 8 studies [40,41,43-46,53,59]. Anxiety and OCD were investigated by 2 studies [40,69]. Borderline disorder and bipolar disorder were investigated by 3 studies [26,40,59]. Seasonal affective disorder was studied by Coppersmith et al [40,59]. Eating disorder was explored by Chancellor et al [29], Coppersmith et al [40], and Prieto et al [56]. Attention-deficit/ hyperactivity disorder, anxiety, and schizophrenia were examined by Coppersmith et al [40], and sleep disorder was studied by Jamison-Powell et al [70]. None of the included studies explored phobias or panic disorders. Users with suicidal ideation were investigated by 8 studies [30-33,35,39,51,60]. Happiness, satisfaction with life, and well-being were investigated by 7 studies [28,34,47,54, 55,68,71].

Of the studies included in this review, 31 analyzed social network contents written in English [24-31,33-35,40-46,49, 50,52,53,58-60,65-68,70,71]; 11 studies investigated Chinese text [32,37,39,48,51,54,55,57,62,63,69]; 2 focused on Korean [36,61] and 2 on Japanese text [38,64], 1 looked at Turkish content [47], and 1 jointly at Spanish and Portuguese [56].

Each of the selected articles was based on a dataset directly or indirectly obtained from social networks. We identified 2 broad approaches to data collection: (1) collecting data directly from the participants with their consent using surveys and electronic data collection instruments (eg, Facebook apps), and (2) aggregating data extracted from public posts.

The methods for collecting data directly from participants varied with the purpose of the studies and the target platform. These methods included posting project information on relevant websites inviting participants to take part in the project [32,38,50] and posting tasks on crowdsourcing platforms asking for project volunteers [28,30,66,67]. For crowdsourcing, researchers posted detailed information about their studies on platforms such as Amazon Mechanical Turk [74] to attract participants. As part of a questionnaire, the participants would typically be asked to provide informed consent allowing collection of their social network data.

A range of questionnaires were used to measure participants’ levels of depression and life satisfaction, including the Center for Epidemiologic Studies Depression Scale [36,38,61,66,67], Patient Health Questionnaire-9 [50], Beck Depression Inventory [36,38,61,67], Zung Self-Rating Depression Scale [64], Depressive Symptom Inventory-Suicidality Subscale [30], and Symptom Checklist-90-Revised [69]. The instruments used to detect suicidal ideation and the possibility of an individual committing suicide were the Suicide Probability Scale [32,39,48], the Acquired Capability for Suicide Scale [30], and the Interpersonal Needs Questionnaire [30]. Satisfaction with life and well-being were measured with the Satisfaction with Life Scale [28,34], the Positive and Negative Affect Schedule [55], and the Psychological Well-Being Scale [55]. One study used the Revised NEO Personality Inventory-Revised to assess personality [58].

The second approach was to pool only public posts from social network platforms, by using regular expressions to search for relevant posts, such as “I was diagnosed with [condition name]” [40,42,43,59].

To collect social network data, each data source required a custom capture mechanism, due to a lack of standards for data collection. Facebook-based experiments gathered user datasets by developing custom tools or Web apps connecting to the Facebook application programming interfaces (APIs) [36,50,61]. Another group of studies used Twitter APIs to explore cues for mental disorders [24-27,30,31,33,35,38,47,52,53,56-60,64-68, 70,71]. A similar approach was used for Instagram APIs [29] and Sina Weibo APIs [32,37,39,51,54,55,57,62,63].

Another way of obtaining data was promoted by the myPersonality project, which provides both social network data and a variety of psychometric test scores for academic researchers [75], and was used by 3 studies [28,34,58]. Some studies [41,44-46] originated from workshops where the organizers shared data already approved by an institutional review board (IRB) for analysis.

In all of the selected studies, several standard steps had to be taken before machine learning algorithms could be applied to data. First, data were cleaned and preprocessed to ensure that they were in the form required by the analytical algorithms. Second, the key features (the term “feature” in machine learning denotes a set of observations that is relevant to the modelling problem, typically represented numerically [76]) related to the research domain were prepared for model construction. Overall, this involves feature extraction and feature selection, producing sets of features to be used in learning and validating predictive models.

The ethical aspects of using social network data for research are still not clearly defined, particularly when working with information that is publicly available. Thus, the studies that we surveyed adopted a wide range of approaches to handle ethical constraints.

Among the articles included in this review, 9 [30,32,33,36,38,40,42,48,61] were approved by their authors’ IRBs, and 8 [34,36-38,48,50,55,61] reported receiving informed consent from participants prior to data analysis. For public data collected from crowdsourcing platforms, participants who opted in provided their consent to data sharing [67]. For myPersonality data, Liu et al [34] stated that the dataset itself had IRB approval, so the authors did not report obtaining any further approval from their institution. Youyou et al [83] also concluded that no IRB approval was needed for using myPersonality data. Chancellor et al [29] did not seek IRB approval, because their study used Instagram data without personally identifiable information.

Researchers in 6 studies [31,33,35,52,53,70] reported that the social network datasets collected from participants were anonymized. O’Dea et al [33] removed names, user identifiers, and user identities, and the data collected had to be analyzed after 3 months. Names and usernames in tweets were removed or replaced with other text in 3 studies [31,52,53]. Jamison-Powell et al [70] reported that they removed user identifiers from tweets illustrated in their published article.

The performance of these models is still fuzzy and unstable. As a consequence, none of these studies presented the model’s predicted output to participants themselves. Schwartz et al [28] also noted that mental health predictive models are still under development and not sufficiently accurate to be used in practice, and little research has been done on user acceptability of such tools.

The purpose of this review was to investigate the state of development of research on machine learning techniques predicting mental health from social network data. This review also focused on identifying gaps in research and potential applications to detect users with mental health problems. Despite the thousands of articles collected through our search terms, the results of our review suggest that there is a relatively small but growing number of studies using machine learning models to predict mental health problems from social network data. From the initial set of matched articles, only 48 met our inclusion criteria and were selected for review. Some of the excluded studies focused on analysis of the effects of social media use on mental health and well-being states of individual users, and the influence of cyberbullying in social networks on other users.

From the above results, we observed that the same methods could be adapted to analyze posts in different languages. For example, Tsugawa et al [38] adapted De Choudhury’s methods [67], originally designed for the analysis of English textual content, to Japanese textual content. Both of them achieved similar results, although some outcomes were dissimilar due to differences in contexts and cultures. This example illustrates that same methods can be used to facilitate studies in different languages.

Several sites were used as sources of data. Facebook is possibly the most popular social network platform. However, only a few studies relied on Facebook datasets to predict mental disorders. One reason for this might be that, by default, users on this site do not make their profiles publicly accessible. Another reason is that getting data from Facebook requires consent from users.

From the selected studies, we can acknowledge several benefits and drawbacks of the methods used in the experiments.

The selected articles were largely focused on depression, at around 46% (22/48), while 17% (8/48) focused on suicide. Nearly 15% (7/48) of the articles reported a study of well-being and happiness. The rest of the articles investigated postpartum depression, eating disorder, and PTSD. Worthy of note, there is a lack of models for detection of chronic stress and anxiety disorders. Only 1 study in our sample built a stress state detection model [57]. Therefore, this is likely to become an interesting avenue for future research. If a user has a long period of chronic stress, he or she might be becoming depressed. For instance, Hammen [89] reported that chronic stress is a symptomatic source of depression and can develop into other disorders.

Furthermore, based on the selected articles, no study used social network data from actual patients with mental illnesses clinically identified by a doctor or psychologist. Most of the studies included in our review assessed mental disorders with surveys, which are open to self-identifying biases. It would be interesting to promote a closer collaboration between computer scientists and doctors or psychologists, who could provide access to patients with a diagnosis of mental disorders. This might improve the accuracy and reliability of data, making it possible to build predictive models based on features extracted from real patients’ social networks. However, mental health conditions might only be formally diagnosed in a specific subset of patients with those conditions, which may lead to a different type of bias.

Importantly, this area of research can benefit from the adoption of open science standards [90]. Many of the studies we reviewed were based on an analysis of openly available data from social networking sites or from the myPersonality project. Parts of the materials used in some of the studies are posted online (eg, [28,32]) or available upon request (eg, [61]). However, most of the studies did not share their entire computational workflow, including not only the datasets, but also the specific code used to preprocess and analyze them. Therefore, future studies should comply with the Transparency and Openness Promotion guidelines [91] at level 2 (which requires authors to deposit data and code in trusted repositories) or 3 (which also requires the reported analyses to be reproduced independently before publication). Of course, to avoid the dissemination of sensitive personal information, the datasets should be properly deidentified when necessary.

The next generation of predictive models will include more technical analyses. Most of the selected studies relied on textual analysis. But apart from text mining techniques, other methods can be used to gain insights into mental disorders in collected datasets. For instance, image analysis can be used to extract meaningful features from images posted by users. Users facing mental disorders may post images with specific color filters or contents. Among our reviewed studies, 2 found a significant relationship between emotions and color use [92,93]. Another interesting technique is social network analysis. In this review, we selected 3 studies that used social network analysis to examine mental health. However, only 2 studies analyzed symptoms of mental disorders through social network analysis [24,63], while 1 study explored well-being [71]. One study reported that symptoms of depression can be observed through social networks. In other words, depression can be detected through each person’s friends [94]. These examples show that social network analysis is a promising tool to investigate the prevalence of mental illness among online users.

A wide range of machine learning algorithms were used in the reviewed studies. Only 1 study used deep learning algorithms to build a classifier [57], with the rest relying on SVMs, regression models, and decision trees to build classification models. It is expected that, with the rise in popularity of deep learning techniques, this will be changing soon. However, deep learning models are a black box, as opposed to human-interpretable models, such as regression and decision trees, raising the issue of whether it is possible, or indeed necessary, to have these algorithms validated by clinical experts [95].

As this review showed, it is now possible to detect social network users with mental health problems. However, a supporting methodology needs to be developed to translate this innovation into practice and provide help to individuals. Thus, mechanisms are needed to integrate the data science efforts with digital interventions on social network platforms, such as promoting access to health services, offering real-time interventions [96], delivering useful health information links, and conducting cognitive behavioral therapy [97] (see Figure 2).

Several studies outside the scope of this review are particularly useful in highlighting the importance of ethical issues in this area of research. For instance, researchers from Facebook and Cornell University [98] collected and used datasets from Facebook, without offering the possibility to opt out. According to the US Federal Policy for the Protection of Human Subjects (‘Common Rule’), all studies conducted in the United States are required to offer an opt-out for participants. However, private companies do not fall under this rule [99]. This study was not approved by the Cornell University IRB either, “[b]ecause this experiment was conducted by Facebook, Inc. for internal purposes, the Cornell University IRB determined that the project did not fall under Cornell’s Human Research Protection Program” [99].

Another study collected public Facebook posts and made the dataset publicly available to other researchers on the Internet [100]. The posts were manually collected by accessing authors’ friends’ profiles, and anonymizing them. But even so, the posts could still be easily identified [101].

As a result of privacy issues in research with human subjects, the Association of Internet Researchers and other authors have proposed not only ethical questions to evaluate the ethical implications of a research project before starting, but also specific guidelines to eliminate and deal with these issues [102,103].

Surprisingly, few of the studies focused on ethical issues. Conway [104] provided a taxonomy of ethical concepts to bear in mind when using Twitter data for public health studies. Conway [104] and McKee [105] reviewed and presented normative rules for using public Twitter data, including paraphrasing collected posts, receiving informed consent from participants, hiding a participant’s identity, and protecting collected data. Some ethical issues, including context sensitivity, complication of ethics and methodology, and legitimacy requirements, were explicitly addressed by Vayena et al [106].

Mikal et al [107] focused on the perspectives of participants in using social media for population health monitoring. The authors reported that most research participants agreed to have their public posts used for health monitoring, with anonymized data, although they also thought that informed consent would be necessary in some cases.

One approach to reducing the ethical issues of accessing to and using personal information in this area of research is to anonymize the collected datasets to prevent the identification of participants. Wilkinson et al [103] suggested that researchers should not directly quote messages or the public URLs of messages in publication, because these can be used to identify content creators. Sula [108] provided strategies to deal with research in social media including involving participants in studies (not just collect public contents), not collect personally identifiable information (eg, social network profile names), provide participants with a chance to opt out, and make resulting research findings easily accessible and understandable to participants. In most localities, doing any research that collects private information (including social networking posts) from human participants is required to provide project information to IRBs or ethics committees to obtain approval prior to data collection [102,109].

This review focused on studies building predictive machine learning models to automatically detect mental health conditions from social network data. Some studies linking mental health and other sources of data did not meet our selection criteria but provide interesting insights about research trends in this area. For instance, previous research has tried to predict mental health conditions or suicidal risk from alternative sources of data such as clinical notes [110], voice analysis [111,112], face analysis [113], and multimodal analysis [114]. We excluded other studies from this review because they used social media data to predict different outcomes; for example, Hanson et al [115] used Twitter data to predict drug abuse. Additionally, recent work has investigated reasons behind Twitter users posting about their mental health [116].

The purpose of this review was to provide an overview of the state-of-the-art in research on machine learning techniques predicting mental health from social network data. Most of the selected studies approached this problem using text analysis. However, some studies also relied on image analysis and social network analysis to gain insights into mental health problems from social network datasets. Predictive models and binary classifiers can be trained based on features obtained from all these techniques. Based on our selected articles, there were relatively few studies applying predictive machine learning models to detect users with mental disorders in real social networks. Moving forward, this research can help in designing and validating new classification models for detecting social network users with mental illnesses and recommend a suitable individually tailored intervention. These interventions might be delivered in the form of advertisements, information links, online advice, or cognitive behavioral therapy; for example, Facebook is considering offering users deemed at risk of suicide online help in real time [117]. However, the reliability of the provided social network data and the general desirability of such interventions should be carefully studied with the users.

With advances in smart data capture devices, such as mobile phones, smart watches, and fitness accessories, future research could combine physical symptoms, such as movements, heart signs, or sleep patterns, with online social network activity to improve the accuracy and reliability of predictions. Finally, scholars interested in conducting research in this area should pay particular attention to the ethical issues of research with human subjects and data privacy in social media, as these are still not fully understood by ethics boards and the wider public.