Register for: WIMSIG2017

Details of talk

TitleAutomated Classification of Post-Stroke Aphasia by Severity
PresenterTea Kristiane Espeland Uggen (University of Technology, Sydney)
Author(s)Tea Kristiane Espeland Uggen, Tapan Rai and Erin Godecke
SessionContributed Talks
Time14:00:00 2017-09-26

Aphasia is a post-stroke communication impairment identified in approximately
one third of stroke survivors. Assessing the level of severity for patients with
aphasia is necessary in order to determine the optimal rehabilitation pathway
for each patient. This paper aims to predict aphasia severity through natural
language processing and machine learning techniques. Predictive analysis was
conducted in this paper based on 51 patient transcripts from English-speaking
stroke survivors with aphasia. Numerical linguistic measures based on sentence
structure, syntax and characteristics of normal speech were derived from these
transcripts to perform predictive modelling using Support Vector Machines (SVM).
Two different SVM models were conducted to predict aphasia severity. The results
of model 1 show that these linguistic measures are able to predict aphasia
severity (mild, moderate, severe) at approximately a 72.6\% accuracy. In model
2, the relevance of patientsí speech to the task was incorporated to account for
patients who deviate from the topic of the task. This addition further improved
the accuracy by 2.6\% from model 1 to an accuracy of approximately 74.5\%.
Further research involves determining which measures of speech are most relevant
in predicting aphasia severity as well as obtaining other appropriate measures
to be included in the predictive analysis to increase the accuracy of the