Text data mining approach for mental health prediction

• Main Author: Chan, Ian Jia Jun
• Other Authors: Vidya Sudarshan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Engineering; Sentiment analysis; Mental health prediction; BERT
• Online Access: https://hdl.handle.net/10356/175167
• Institution: Nanyang Technological University

Depression has been considered one of the most common mental disorders around the world with far-reaching negative impacts on those who suffer from it. On top of that, social media has taken the world by storm, with around 60% of the global population and 92.7% of all internet users being social media users. As such, users tend to express their emotions and thoughts through text on social media platforms. The advancements in Artificial Intelligence, namely Natural Language Processing (NLP), has allowed for the understanding of text and the underlying emotion within. This is called Sentiment Analysis. With the proper Sentiment Analysis model, the detection of depressive text can help identify potential depression and raise awareness of mental health. In this study, this paper compares existing tools, namely Long Short-Term Memory (LSTM), and Bi-directional Encoder Representation from Transformers (BERT), for depression detection and examine the shortfalls of these models. Furthermore, by using existing models as a base and tweaking it, this paper aims to produce something better. Not only that, but this paper will only investigate depression text classification and not a wider range of classes. This is in hopes of producing models that are more accurate and robust in this binary class classification. This paper proposes the use of pre-trained BERT to outperform existing Sentiment Analysis models. Furthermore, this paper proposes a novel modification to the BERT classifier to possibly perform better. This paper examines the performance of these models with a wide range of metrics and aims to produce a model that can accurately detect underlying depression among textual data.