Adapting Deep Learning Transformer Model for Depression-Specific Named Entity Recognition: A Transfer Learning Approach

Named Entity Recognition (NER) is essential in the biomedical domain, particularly in mental health studies focused on disorders like depression. It helps extract structured information from unstructured text, enabling accurate recognition of key entities such as symptoms, disorder, antidepressants, brands, and therapies. This step is crucial for recommending diagnoses, personalizing treatments, and advancing research. Traditional methods, like rule-based or dictionary-driven approaches, often fall short in flexibility and scalability, as they rely heavily on predefined lexicons and corpora. The proposed work presents an advanced deep learning model aimed at improving NER performance on a customized dataset. Our model focuses on identifying critical entities disorders, brands, antidepressants, and therapies by addressing the limitations of conventional techniques. It uses a hybrid approach that combines a BERT-based model, fine-tuned through transfer learning, with a knowledge-based dictionary (MedDRA). This integration allows for enhanced semantic enrichment, improving the identification and categorization of biomedical entities. By leveraging BERT’s contextual capabilities alongside the structured knowledge from MedDRA, the model delivers better accuracy and classification in mental health data. Preliminary results show high confidence levels, with substantial improvements in key metrics like precision, recall, and F1 score, further advancing NER's role in mental health research.