Ehsan Hoque and his team at the University of Rochester have found a way to analyze selfies and detect the likelihood of Parkinson’s disease in a patient. The novel approach, which involves computer vision software, has attracted $500,000 in funds from the Gordon and Betty Moore Foundation.
“The foundation wants us to validate the feedback that we would give people if they did, indeed, show early signs of Parkinson’s—especially if they are performing the test at home,” says Hoque, associate professor of computer science. “The challenge is not only validating the accuracy of our algorithms but also translating the raw machine-generated output in a language that is humane, assuring, understandable, and empowering to the patients.”
Though ethical and technological considerations need to addressed, the tool examines a major symptom of Parkinson’s: stiffness in facial muscle movements and reduced facial expression. Called hypomimia, it is considered a sensitive biomarker for the disease and a way to diagnose the disease early on.
In a recently published paper in Nature Digital Medicine, Hoque and collaborators shared analyses of 1,812 videos using an online tool. They studied facial muscle movements using a coding system when participants demonstrated a facial expression. By computing variances in muscle movements, the authors were able to show that the measure of micro-expressions has the potential to be a digital biomarker for Parkinson’s disease.
Patients with the disease have fewer facial muscle movements than those without the affliction. The disease predication accuracy (95 percent) using facial muscles is comparable to existing state-of-the-art video analysis that uses limb movements (93 percent), the study states.
In collaboration with Ray Dorsey—an expert in Parkinson’s disease and the David M. Levy professor of neurology—and UR’s Morris K. Udall Parkinson Disease Research Center, the researchers have developed a five-pronged test. Neurologists can administer the test to patients sitting in front of their computer webcams, UR says.
Patients taking the test are asked to read aloud a complex written sentence, touch their index finger to their thumb 10 times as fast as possible, make a disgusted look along with a neutral expression three times and raise their eyebrows as high as possible and lower them as far as they can, three times slowly. These test takers also are asked to smile and alternate them with neutral expressions three times.
Within minutes, using machine learning algorithms, the software can generate a percentage likelihood, from each of the tests, that the patient is showing symptoms of Parkinson’s disease or related disorders. For example, the software can detect whether patients show less control over their facial muscles while making a smile, a symptom of Parkinson’s.
“One thing about Parkinson’s is that you don’t show all the symptoms all the time, and not every symptom is shown in every part of your body,” says Rafayet Ali, lead author of the paper and former post-doctoral student in Hoque’s lab. “For example, you may not have hand tremors, but you may show a significant level of deviation in your smile.”
That underscores the importance of testing other expressions and movements, adds Ali, an associate data scientist at Sysco.
UR says both Hoque and Ali have personal reasons to help people with the debilitating disease. Their mothers both have suffered from the disorder. Hoque’s late mother in Bangladesh was prescribed a medication for Parkinson’s. Though drug helped take away tremors for a while, they eventually returned.
A discussion with Dorsey a few years ago led to a move away from pen and paper forms, introducing automation.
“Objective, digital assessments of Parkinson’s disease can help us diagnose people with the condition and evaluate new therapies for the condition faster,” says Dorsey, an author of “Ending Parkinson’s Disease.”
Before Hoque and researchers can begin asking for patients’ permission to analyze selfies and use the test, the algorithm needs to be refined to ensure high accuracy, while recognizing that automation can never be always on the mark.
“An algorithm will never be 100 percent accurate,” Hoque says. “What if it makes a mistake? We want to be very careful and follow guidance from the FDA if we want anybody from any part of the world to try this and get an assessment.”
Still, the approach could be transformative in that it could have a wide reach. Patients who are immobile, quarantined or live in areas with limited access to a neurologist could find answers, for instance.
Also, movement disorders related to Parkinson’s disease could be next. Diseases such as Huntington’s share similar tremor symptoms but are difficult to distinguish.
“We can’t tell that just yet,” Hoque says. “But we are in a pursuit of differentiating those tremors using (artificial intelligence) to prevent the potential harm of misdiagnosis while maximizing benefit.”
For now, early detection of Parkinson’s disease could also help with related factors such as reduced social wellbeing and depression among patients.
Smriti Jacob is Rochester Beacon managing editor.