Multi Class Skin Disease Classification Using T5 and BERT Based NLP Models

Effective and automated diagnostic methods are required due to the rising incidence of dermatological illnesses, especially in clinical settings with limited resources. Despite the wealth of clinical narratives and medical records, text-based categorisation of dermatological disorders is still understudied. Traditional image-based techniques for skin disease classification have received a lot of attention. The present work provides a multi-class skin disease classification framework using advanced NLP models (T5 and BERT). The proposed working method for solving class imbalance problem, semantic diversity problem and having a complex medical language, works with text-level entries of dermatological diseases and assigns them to given classes. In the proposed work, we utilized a large clinical notes dataset that had been annotated by domain-experts across 13 different classes of skin diseases. Tokenization, stop word removal, and domain-specific embedding generation were among the initial pre- processing stages of our method that receive unique attention within the context of dermatological vocabulary. Then we adapted pretrained transformers—namely T5 and BERT—for the procedure of multiclass text classification. We used common metrics like accuracy, F1-score, precision, and recall to measure the performance of the implemented models. These results were compared to the results from both conventional machine learning algorithms and state-of-the-art transformer architectures. As per comparative analysis, our method obtained considerably higher classification accuracy which highlights the power of transformer-based architectures in modelling complex medical textual data. Our work contributes to the rapidly evolving field of AI-augmented dermatological diagnosis through a scalable text-centric classification approach. Such an approach also has the potential to complement existing image-based diagnostic tools and thereby improve overall clinical decision-making abilities. For future work, we want to investigate multimodal learning methods to combine text data with visual data and improve diagnostic accuracy while expanding practical clinical uses.