Fall Detection in EHR using Word Embeddings and Deep Learning
Author: Henrique D. P. dos Santos, Amanda P. Silva, Maria Carolina O. Maciel, Haline Maria V. Burin, Janete S. Urbanetto and Renata Vieira
Abstract: Electronic health records (EHR) are an important source of information to detect adverse events in patients. In-hospital fall incidents represent the largest category of adverse event reports. The detection of such incidents leads to better understanding of the event and improves the quality of patient health care. In this work, we evaluate several language models with state-of-the art recurrent neural networks (RNN) to detect fall incidents in progress notes. Our experiments show that the deep-learning approach outperforms previous works in the task of detecting fall events. Vector representation of words in the health domain was able to detect falls with an F-Measure of 90%. Additionally, we made available an annotated dataset with 1,078 de-identified progress notes for replication purposes.
Keywords: Fall Detection, Electronic Health Records, Text Classification, Word Embeddings, Deep Learning
This project belongs to GIAS at PUCRS, Brazil