Sentiment analysis for depression detection: A stacking ensemble-based deep learning approach

Abstract

Depression is one of the most common mental health issues that seriously affect people's quality of life. The World Health Organization reported that depression overwhelms about 300 million people across the globe. Due to the widespread prevalence of this disorder in society, novel and efficient methods must be developed for effective detection and treatment. In the modern era of social media, individuals often reveal their emotional states by providing daily posts on platforms like X (previously Twitter) and Facebook. The information can be utilized as an essential input for determining whether a person has depression based on their writing content. The disclosure of transformer-based deep learning models provides an opportunity to use pre-trained models to successfully capture complex patterns and nuances in the textual data. This study proposes a novel depression detection method through sentiment analysis by developing a Stacking ENSemble-based Deep learning (SENSDeep) model. The proposed model integrates the capabilities of six pre-trained cutting-edge models, including BERT, RoBERTa, AlBERT, DistilBERT, XLNet, and BART, through stacking ensemble to enhance the predicted performance of the proposed model. The SENSDeep model is evaluated by precision, recall, F1-score, and accuracy. In contrast to other models, the SENSDeep model excels with 96.93 % precision, 97.50 % recall, 97.22 % F1-Score, and 97.21 % accuracy. To our knowledge, SENSDeep is the first deep-learning ensemble model that leverages the capabilities of cutting-edge pre-trained transformer models via stacking, specifically for detecting depression from the textual data.