Clinical Score Estimation for Determining Oro-Facial Dysfunction Severity

Document Type

Conference Proceeding

Publication Date



Stroke and amyotrophic lateral sclerosis manifest symptoms that affect facial motion in patients. Tracking these movements and assessing the severity of the impairment can be achieved with facial alignment technology and classification algorithms. Using the Toronto NeuroFace Dataset consisting of patients and healthy individuals performing clinical examination tasks, this study focuses on score estimation of clinical examinations to determine oro-facial dysfunction severity. Facial landmarks extracted using the 2D FAN were used to determine features under range of motion, speed of motion, and symmetry. Speech language pathologist scores from the dataset were transformed using ordinal encoding, then oversampled using random oversampling. The features and transformed scores were fed into random forest classifier models to predict a score using a scale of 1 to 4 for each feature category. The results show that the proposed method is able to estimate oro-facial dysfunction severity and classify between healthy individuals and patients. The average performance of the model setups are comparable to that of the baseline in terms of accuracy (<5% difference), accuracy±1 (<2% difference), binary accuracy (<3% difference), and specificity (<7% difference).

This document is currently not available here.