Early Detection of Mental Health Crisis Warning in Online Content

Abstract

This study investigates the early detection of depression in online content through advanced Artificial Intelligence techniques. Whereby, we conducted a comprehensive comparison between Machine Learning, Deep Learning, and Transfer Learning models based on Transformers. Our methodology involves applying several text pre-processing steps and employing different word embedding techniques including FastText to achieve best representation of social media textual data. Empirical results demonstrated the outperformance of large-scale transformer-based models, particularly BERT and ELECTRA by achieving 99.9%, 99.89%, 99.79%, and 99.84% for accuracy, recall, precision and F1-score, respectively. Compared to literature, these two transfer learning models showcased to be notably superior. Thereby, we incorporated ELECTRA model yielding the best performance in chatbot depression detector tool to provide immediate support and resources to individuals identified as at risk of mental health crisis.