Aware Topic Modeling in Medical Texts Using A DTM-RNNLSTM Framework with UMLS Integration

Abstract

The exponential rise in unstructured medical text volume in recent years has led to a pressing need for sophisticated topic modeling methods that can capture temporal dynamics and semantic richness. This study suggests a brand-new hybrid framework called DTM-RNNLSTM, which combines the sequential learning powers of Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) networks with Dynamic Topic Modeling (DTM). The model integrates ideas from the Unified Medical Language System (UMLS) to improve semantic relevance, making it possible to identify issues with medical significance. The MedMentions dataset, a sizable corpus annotated with UMLS concepts, is used to assess the efficacy of the suggested model.  Three robust baseline models are compared: the Dynamic Topic Model (DTM), Gibbs Sampling Dirichlet Multinomial Mixture (GSDMM), and Non-negative Matrix Factorization (NMF). Coherence, Perplexity, Precision, Recall, F1-Score, and Accuracy are evaluation measures that address both statistical and semantic performance factors. The findings show that DTM-RNNLSTM outperforms conventional methods in capturing changing topic patterns and greatly enhances semantic coherence.