Dynamic Topic Modeling Techniques for Evolving Medical Texts with UMLS Concepts

In the realm of biomedical research, efficiently extracting and categorising subjects from enormous amounts of medical texts is crucial for knowledge discovery and information retrieval. Traditional topic modelling approaches are useful, but they usually fall short in capturing the intricate semantics of medical terminology. This study investigates the potential benefits of using Unified Medical Language System (UMLS) principles to topic modelling using the MedMentions dataset. We employ four techniques: BERTopic, Latent Dirichlet Allocation (LDA), Hybrid LDA and RNN, and a novel Hybrid BERTopic with Recurrent Neural Networks (RNN). By incorporating UMLS concepts into these models, we hope to improve subject coherence and relevance. According to our research, in terms of clinical relevance and topic coherence.