Harnessing Deep Learning for Mental Health: A Comparative Study of BERT, LSTM, and GPT-2

Mental health disorders in particular depression threaten the lives of millions of people worldwide, making early detection vital for timely intervention. With the emergence of social media, user-generated posts present an extensive repository of data for studying mental health disorders.The present study examines various deep learning models including BERT, LSTM and GPT-2 for detecting depressive symptoms in social media text. A Mental Health Social Media Dataset made available by Kaggle was used to carry out binary classification of the posts, as either depressed or not depressed.The models were fine-tuned and evaluated based on the metrics of accuracy, precision, recall, F1-score and AUC-ROC.From the experimental results, BERT performed the best, achieving an accuracy of 91.38% and AUC-ROC of 0.97 compared to 90.15% for GPT-2 and 81.72% for LSTM.BERT was excellent in understanding context where GPT-2 could analyze conversation well, and LSTM was best in computational efficiency.This study emphasizes deep learning techniques for mental health prediction and discusses challenges with respect to model interpretability, dataset bias, and real-time deployment. Future work should focus on multimodal learning, bias reduction, and Explainable AI (XAI) for enhanced transparency and ethical deployment within mental health diagnosis.