JMIR Publications

ABSTRACT

Background:

Suicide risk assessment usually involves an interaction between doctors and patients. However, a significant amount of people with mental disorders receive no treatment for their condition due to the limited access to mental healthcare facilities, the reduced availability of clinicians, lack of awareness, and the stigma, neglect and discrimination surrounding mental disorders. In contrast to this, Internet access and social media usage has increased significantly, providing experts and patients with a mean of communication that might contribute to the development of methods for the early detection of mental health issues on social media users.

Objective:

This work reports an approach for suicide risk assessment on Spanish speaking users on social media. We explore behavioral, relational and multi-modal data extracted from multiple social platforms; and build machine learning models to detect users at risk.

Methods:

We first characterize users based on their writings, posting patterns in different time periods, relations with other users, and images posted. Later we explore and evaluate several statistical and deep learning approaches to handle multi-modal data for the detection of users at suicide risk. Our methods are evaluated over a dataset annotated by specialized clinicians. To evaluate our models’ performance, we distinguish two types of control groups: users that make use of suicide related vocabulary (focused control group), and generic random users (generic control group).

Results:

We identified significant statistical differences (P<.001) between textual and behavioral attributes of each of the control groups compared to the suicide risk group. Our findings show that the combination of textual, visual, relational and behavioral data from users outperforms the accuracy of using each modality separately. We define text-based baselines models based on Bag of Words and word embeddings, which are outperformed by our models obtaining an increase in accuracy of up to 8% when distinguishing users at risk from both types of control users.

Conclusions:

The types of attributes analyzed are significant for detecting users at risk, and their combination outperforms the results provided by generic exclusively text-based baseline models. Seeing the performance of image based predictive models, we believe that our results can be improved by enhancing the contribution of the textual and relational features. These methods can be extended and applied to different use cases related with mental disorders such as depression, anxiety, or eating disorders.