Predicting patient outcome from clinical journals and biomedical articles: Using the MIMIC-IV database, multiple in-hospital mortality prediction models are created, to which improvements are attempted through the use of word embeddings trained on scientific biomedical literature

Abstract

There exists an immense amount of data within the medical and healthcare field. Two examples of these are the patient’s clinical records and biomedical research literature. The clinical records often contain numerical information suitable for developing AI-driven systems to support medical workers, and with some work, we believe that data from the biomedical literature can also be used to improve these systems. In this work, clinical records from ICU patients in the MIMIC-IV database are used to create multiple prediction models to predict the risk of a subject passing away during their hospital stay (in-hospital mortality prediction). From the clinical records, abnormal observations are categorised into medical terms which are then used to find relevant biomedical research work from the PubMed database. This data is used to train word embeddings using Word2Vec. The word vectors for these categorised observations are then added as a concatenation to the clinical records data, in hopes of improving the predictions. The results show very little difference with the additional information from the word embeddings. These are small improvements but unfortunately, we can not conclude that this can help improve such prediction models. We do however believe it suggest that there is valuable information that can be extracted from biomedical articles and that further work could show this.