Assessing mortality prediction through different representation models based on concepts extracted from clinical notes

Abstract

Recent years have seen particular interest in using electronic medical records (EMRs) for secondary purposes to enhance the quality and safety of healthcare delivery. EMRs tend to contain large amounts of valuable clinical notes. Learning of embedding is a method for converting notes into a format that makes them comparable. Transformer-based representation models have recently made a great leap forward. These models are pre-trained on large online datasets to understand natural language texts effectively. The quality of a learning embedding is influenced by how clinical notes are used as input to representation models. A clinical note has several sections with different levels of information value. It is also common for healthcare providers to use different expressions for the same concept. Existing methods use clinical notes directly or with an initial preprocessing as input to representation models. However, to learn a good embedding, we identified the most essential clinical notes section. We then mapped the extracted concepts from selected sections to the standard names in the Unified Medical Language System (UMLS). We used the standard phrases corresponding to the unique concepts as input for clinical models. We performed experiments to measure the usefulness of the learned embedding vectors in the task of hospital mortality prediction on a subset of the publicly available Medical Information Mart for Intensive Care (MIMIC-III) dataset. According to the experiments, clinical transformer-based representation models produced better results with getting input generated by standard names of extracted unique concepts compared to other input formats. The best-performing models were BioBERT, PubMedBERT, and UmlsBERT, respectively.