BACKGROUND
Depression affects the mental and physical well-being of humans. According to the latest statistics, cases of depression-driven self-harm and suicide are reported at an alarming rate. Although multiple clinical diagnosis methods have been established, the requirements of having skilled medical staff, equipment, and processing time limit their accessibility. Depression detection using social media data has therefore become a growing research area. Social media is used to express emotions and feelings hence analyzing such content could help to detect abnormalities sooner than by clinical diagnosis. Even though there are systems which can detect the presence or absence of depression, such methods cannot determine the severity of depression and many of them were not validated on large corpora.
OBJECTIVE
In this study, we propose a confidence vector based novel deep learning approach for detecting the severity of human depression. A well-balanced and aggregated novel dataset is also introduced as a part of this study to validate our proposed model on a large corpus.
METHODS
We formulated the problem of interest into a multiclass classification problem. The core logic is to generate confidence vectors considering the identified depressive words/phrases of each extracted social media statements and then to obtain the exact depression severity level using a fully connected deep neural network. To validate our method, we aggregated two labelled, unbalanced, and small public corpora for depression severity detection to obtain a well-balanced dataset consisting of more than 40,000 social media statements. Optimal data balancing techniques were applied to maintain a fair number of samples in each depression severity class. To our knowledge, this is the largest corpus for depression severity detection to date.
RESULTS
Experimental results showed that our methodology achieved a new state-of-the-art (SOTA) performance in depression severity detection with Precision, Recall, and F1 Scores of 79%, 77%, and 76%, respectively, over existing baselines.
CONCLUSIONS
The results showed that our method is competitive with baseline models for depression severity detection.