JMIR Publications

ABSTRACT

Background:

There is an increasing number of online support groups where one can get advice and information regarding topics related to mental health.

Objective:

We aimed to map evidence concerning interactions among Internet users with mental disorders and identify gaps in knowledge.

Methods:

We searched MEDLINE, Embase, Cochrane, and Web of Science until July 24, 2021. We included qualitative or qualitative-quantitative studies investigating interactions among Internet users with mental disorders. We used phi coefficient and applied machine learning techniques (decision trees, logistic regression, support vector machines, k-nearest neighbours, and Gaussian Naive Bayes classifier) to predict type of mental disorders from interactions. Our protocol was registered in the Open Science Framework (osf.io/j3azv).

Results:

Out of 11,316 identified records, we included 38, which analysed 79,735 users and 218,156 posts. Most frequently interactions concerned people with eating disorders (19%), depression (19%), and psychoactive substance use (17%). We grouped interactions between users into 42 codes, with ‘network’ being the most common (7%). Most frequently co-existing codes were ‘request for information’ and ‘sharing self-disclosure’ (34 times, φ = 0.57, p <.001). Algorithms that provided the best accuracies in classifying disorders by interactions included logistic regression and decision trees (81%). Included studies were moderately valuable in terms of quality.

Conclusions:

There are peer-to-peer interactions that are characteristic of some mental illnesses. Obtaining more data about the online interactions between people with mental illnesses could help properly apply machine learning methods to create a tool that helps in screening or even supports assessing mental state. Clinical Trial: Our protocol was registered in the Open Science Framework (osf.io/j3azv).