Extracting Latent Insights and Tagging Fall Injuries from Clinical Narratives Using Unsupervised Learning

Falls have been one of the leading causes of death in the United States. Any insights about these falls can potentially inform interventions such as education and awareness programs, exercise programs, and home safety measures. As the accumulation of clinical text data by healthcare organizations is growing, it prompts a need for further insightful exploration into the application of machine learning techniques on these data. This paper discusses the use of a state-of-the-art language model to extract concepts from narratives. We propose novel unsupervised approaches such as Latent Dirchlet Allocation (LDA) and K-means to classify the narratives into multiple classes of injury type. In addition, we discuss a human-assisted survey method to evaluate the performance of the models. The proposed model demonstrated the classification accuracy of 83% (LDA) and 75% (K-means) on unlabeled clinical narratives. The approach yielded valuable insights that might not be readily apparent through human observation. Overall, this study demonstrated the feasibility of utilizing unsupervised NLP methods to extract information from clinical narratives.