BERT-LSTM Hybrid Model for Early Detection of Mental Health Risks

Mental health issues, specifically depression, is an important problem on a global scale. Millions of people around the world suffer from depression due to high-stress living situations, traumatic events, peer pressure and sometimes it may not be detected due to the stigma that follows it. This situation can be improved by early detection and providing the necessary support. This paper proposes a model that combines the contextual understanding of BERT(Bidirectional Encoder Representations from Transformers) with sequential learning in LSTM (Long Short-Term Memory) networks. This hybrid model gave an accuracy of 92.70% with a score of 0.9236 F1 in classifying posts on social media into depressive and non-depressive. Along with this it also gave balanced precision and recall scores for the two labels thus indicating the models robustness across multiple text types. The use of this hybrid model shows the opportunities, text classification applications can gain by combining transformer based models and recurrent neural networks in the mental health field. The high accuracy of this model can play an important role in recognizing severity levels and detecting variations in depression. This research can pave the way for further advancement in online mental health tools and support systems.