Fall Detection in Clinical Notes using Language Models and Token Classifier
Author: Joaquim Santos, Henrique D. P. dos Santos and Renata Vieira
Abstract: Electronic health records (EHR) are a key sourceof information to identify adverse events in patients. The largestcategory of adverse events in hospitals is fall incidents. Theidentification of such incidents guide to a better comprehensionof the event and enhance the quality of patient health care.In this initial work, we compare the performance of Sentence-Classifier (StC) against the Token-Classifier (TkC) with state-of-the-art recurrent neural networks (RNN) to detect fall incidentsin progress notes. Our experiments show that the use of deep-learning algorithms as token-classifier outperforms text-classifier.It improves fall identification using StC from 65% to 92% with TkC (F-Measure). Additionally, the token classifier is able toexplain which words are most important in positive detection.
Keywords: Fall Detection, Clinical Notes, Language Model, Token Classifier
You need install Flair 0.4.3 pip install flair==0.4.3
This project belongs to GIAS at PUCRS, Brazil