Pre- and Post-Depressive Detection using Deep Learning and Textual-based Features

Abstract

The rise of mental disorders, specifically depression, has shown an upward trend, especially during the COVID-19 pandemic. Previous studies have suggested that it is feasible to learn the textual and behavioral features of a user to identify depression on social media. This paper highlights three new contributions, which are, firstly, to introduce Patient Health Questionnaire-9 (PHQ-9) as the survey-based method to complement the TWINT API to collect Twitter data. Secondly, it is to propose a Bidirectional Encoder Representations from Transformers (BERT)-based model, along with emoji decoding and PHQ-9-based lexicon features for predicting the likelihood that a user will exhibit depressive symptoms. The results are promising, achieving an F1 score of 0.98 on a baseline dataset and an F1 score of 0.90 on a benchmark dataset, outperforming previous researcher’s work of achieving a F1 score of 0.85 using solely textual features. Thirdly, previous researcher’s work focuses on differentiating between depressed and non-depressed users only, while this paper further separates the users in the depressive class into before (pre-) and after (post-) self-reported diagnosis, which can potentially be used to detect early symptoms of depression. It was found that the top TF-IDF scores of the post-depressive class contain more frequently negatively implied words compared to the pre-depressive class.

Keywords

• deep learning
• depressive symptoms
• phq-9
• predictive modelling
• sentiment analysis