Jointly predicting affective and mental health scores using deep neural networks of visual cues on the web

Abstract

Despite the range of studies examining the relationship between mental health and social media data, not all prior studies have validated the social media markers against “ground truth”, or validated psychiatric information, in general community samples. Instead, researchers have approximated psychiatric diagnosis using user statements such as “I have been diagnosed as X”. Without “ground truth”, the value of predictive algorithms is highly questionable and potentially harmful. In addition, for social media data, whilst linguistic features have been widely identified as strong markers of mental health disorders, little is known about non-textual features on their links with the disorders. The current work is a longitudinal study during which participants’ mental health data, consisting of depression and anxiety scores, were collected fortnightly with a validated, diagnostic, clinical measure. Also, datasets with labels relevant to mental health scores, such as emotional scores, are also employed to improve the performance in prediction of mental health scores. This work introduces a deep neural network-based method integrating sub-networks on predicting affective scores and mental health outcomes from images. Experimental results have shown that in the both predictions of emotion and mental health scores, (1) deep features majorly outperform handcrafted ones and (2) the proposed network achieves better performance compared with separate networks.