Face Validity: Digital Facial Analysis Recapitulates Clinical Findings of Neurological Disease
Alyssa Nylander1, Kyra Henderson1, Kanishka Koshal1, Nikki Sisodia1, Jaeleene Wijangco1, Shane Poole1, Jim Rowson1, Cathra Halabi2, Jill Ostrem1, Simon Little1, Ethan Brown3, Adam Staffaroni1, Riley Bove3
1UCSF, 2UCSF Department of Neurology, 3University of California, San Francisco
Objective:
To distinguish characteristic facial phenotypes of neurologic disease using computer vision-based facial movement analysis.
Background:
Neurological diseases can present with differences in facial expressions and movements that are often characteristic. Facial expression recognition and monitoring could assist in telehealth diagnostic processes. Digital facial expressivity is an objective and unbiased metric to assess facial movements.
Design/Methods:
Using untargeted recruitment, 124 individuals with multiple sclerosis (MS) and Parkinson’s disease (PD), and healthy controls (HC) are recruited to an ongoing digital phenotyping study. The task evaluated involves video recording of participant faces while engaging in a spontaneous language task. Videos are processed using OpenFace 2.0, an open-access digital tool pre-trained for facial landmark detection and facial action unit (AU) recognition. Average instances of AU activation were averaged across number of frames. Current analyses explored smiling (“cheek raiser,” “lip corner puller”) and eye expression (“brow furrow” and “blink”) AUs. Student’s t-tests were used to compare means between groups.
Results:
Mean age was 47 (SD 12.6) for MS (n=98), 67.3 (SD 10.8) for PD (n=15), and 58.5 (SD 25) for HCs (n=11). Analysis of video metrics revealed significant differences between the groups. Individuals with PD had increased brow-furrowing compared to both MS (difference 0.44, 95% CI 0.29-0.61, p<0.001) and HCs (difference 0.37, 95%CI 0.13-0.61, p=0.002), as well as decreased eye-blinking (MS: difference 0.11 95% CI 0.06-0.15, p<0.001; HC: difference 0.11, 95%CI 0.05-0.17, p=0.0003). Individuals with MS smiled more than those with PD (difference 0.20, 95%CI 0.04-0.36, p=0.01) and also crinkled their eyes less than HC (difference 0.19, 95%CI 0.01-0.37, p=0.04). Recruitment is ongoing. Machine learning will be further applied to the data to increase specificity.
Conclusions:
Digitally identified facial phenotypes recapitulate known clinical characteristics of PD, and differentiate between PD, MS, and HC.