Depression Detection System using Hybrid Deep Learning and Social Media Data

Abstract

Depression is one of the most common mental health disorders, which is increasingly being expressed through online social media platforms. Depression, however, remains untreated and even goes unnoticed. Motivated by the recent advances in affective computing and social media analysis, this paper proposes a depression detection system that combines text, emoji, and image modalities of social media posts using a hybrid machine learning approach. The proposed system builds upon the recent advances in the sentiment pretraining stage and the depression-specific fine-tuning stage. The system employs a multi lingual DistilBERT-based text encoder, along with an emoji aware text preprocessing approach, and an EfficientNetB4-based facial expression model for image analysis. The proposed system combines the power of deep representation learning and metrics based classifiers for detecting sentiment, emotion, and visual affect attributes of the social media posts of the users. Extensive experiments using publicly available sentiment, emotion, and depression datasets have been carried out, which prove the effectiveness of the proposed system. The system achieves good results, with weighted F1-scores greater than 0.87 for multi-class sentiment classification and around 0.90 for binary depression related emotion detection. The results, however, indicate the fea sibility of the proposed system, which is computationally efficient.