Using AI to Detect Early Signs of Parkinson Disease Through Facial Analysis

Researchers at the University of Rochester are analyzing facial muscle movements through videos taken over webcams or smartphones to train a machine learning algorithm with the hope that it can predict Parkinson disease at an earlier stage.

Hypomimia, where facial expressions are reduced, frozen, or stiff, is a common symptom in Parkinson disease. Writing in npj Digital Medicine, scientists at the University of Rochester described early results of a study examining facial micro-expressions of those with PD and without to create a digital biomarker to predict diagnosis of the neurological disease.

Patients with PD were recruited from the University of Rochester medical center. Those without PD were recruited from online advertisements.

In total, investigators examined 1812 videos self-recorded from 604 individuals and collected through a web-based tool; 61 had PD and 543 did not have PD. The mean age of the participants was 63.9 (7.8).

They were asked to record 3 facial mimicry tasks, which involved making 3 facial expressions 3 times, followed by a neutral face after every expression.

Participants were asked to smile, look disgusted, and look surprised. The videos were analyzed with facial recognition software that computed the facial action units (AU) of each frame to measure the variance of the facial muscle movements. Each AU corresponds to a specific set of facial muscles under a coding system published in 1978.

Logistic regression analysis showed that participants with PD have fewer facial muscle movements compared with those without PD; they had less variance in AU6 (cheek raiser, associated with smiling), AU12 (lip corner puller, also linked with smiling), and AU4 (brow lowerer, associated with looking disgusted or surprised).

The authors said the prediction accuracy using this method was comparable to methodologies, such as wearable sensors, that utilize motor symptoms but are more expensive, require active participation on behalf of the individual, and are not scalable.

Next, an automated classifier using support vector machine was trained on the variances and achieved 95.6% accuracy; while the work needs further study, the researchers said “that an algorithm’s ability to analyze the subtle characteristics of facial expressions, often invisible to a naked eye, adds significant new information to a neurologist.”

Another benefit of creating an easy-to-use, less expensive digital biomarker, the researchers said, is that video analysis would overcome hurdles posed by COVID-19, rural areas that lack access to neurology services, patients who are immobile, or potentially provide access to underserved populations.


Ali, M.R., Myers, T., Wagner, E. et al. Facial expressions can detect Parkinson’s disease: preliminary evidence from videos collected online. npj Digit Med. Published online September 3, 2021.