Subscribe to eTOC

Facial Analysis Software Helps Screen for Parkinson's Disease
Where It Is Now and Where It Needs to Go

Article In Brief

A new study shows that facial analysis software can distinguish people with Parkinson's disease (PD) from healthy controls with better than 95 percent accuracy, working only with three short videos of individuals evincing surprise, disgust, and smiling. The software is one of a suite of tools for remote analysis of PD, including video analysis of motor tasks and audio analysis of speech.

figure1

The software is one of a suite of tools for remote analysis of PD, including video analysis of motor tasks and audio analysis of speech.

Perhaps the most important tool in the neurologist's toolkit is a keen power of observation: the ability to sense a slight weakening of the grip, a small asymmetry in the swing of the arms, a subtle decrease in the speed of finger-tapping. Such observations are the key to early clinical diagnosis, especially in diseases without validated biomarkers, such as Parkinson's disease (PD).

But even in a well-off country like the United States, there are tens of thousands, perhaps hundreds of thousands, of people at risk for or in the early stages of PD without ready access to a neurologist; worldwide, the gap between need and access is far greater.

Computer-assisted screening tools have the potential to partly fill that gap, according to Ehsan Hoque, PhD, associate professor of computer science at the University of Rochester in New York, and co-author of a new study showing that facial analysis software can distinguish people with PD from healthy controls with better than 95 percent accuracy, working only with three short videos of individuals evincing surprise, disgust, and smiling. The software is one of a suite of tools for remote analysis of PD, including video analysis of motor tasks and audio analysis of speech.

“We want to enable screening that is available anytime, anywhere, so anyone in any part of the world can go to a website, perform a subset of the United Parkinson's Disease Rating Scale tasks, and immediately get a referral if their results indicate a high probability of PD,” Dr. Hoque said.

It Screens, Not Diagnoses

The software does not yield a diagnosis, Dr. Hoque stressed. “We don't use the word ‘diagnosis’—that's a very strong expectation. We are screening patients. There is no way that we can guarantee that the algorithm will be 100 percent accurate. Some mistakes are likely.”

But of course human neurologists make mistakes too, he pointed out, adding that the estimates of misdiagnosis of PD range from 10 to 30 percent, with essential tremor and atypical Parkinson syndromes topping the list.

Hypomimia—reduced facial expression—is a characteristic feature of PD, and in many patients is observable very early in the disease. The motion analysis software Dr. Hoque and his team have developed can detect even the most minute expressions of hypomimia, he said. “We can quantify these movements, which may be invisible to the human eye, and compare them to someone who doesn't have Parkinson's disease. The computer algorithm can pick up on subtle differences that a neurologist might not see.”

And unlike early forms of motion-capture technology, which relied on applying reflective dots to the skin and tracking their movements, “the technology has progressed so far thar we now have no need for that,” Dr. Hoque said. A simple web camera or smart phone, even in low light, provides all the input the system needs.

figure2

“This is a really exciting study, and has some important applications, including perhaps, with further development, the ability to help less experienced clinicians distinguish PD from other parkinsonian syndromes.”—DR. DEANNA SAYLOR

Study Details

In the study, published in the Nature-affiliated journal npj Digital Medicine in September, the researchers uploaded and analyzed videos from 61 individuals with PD and 543 matched controls. Participants were asked to make three facial expressions—surprise, disgust, and smiling—with a return to a neutral face in between each. The program analyzed the variance, or degree of change from neutral, in the movement of multiple individual “facial action units,” such as “cheek raiser” or “lip corner puller,” which correlate with small sets of individual muscle movements, and which have been shown to be consistently involved in the evoked expressions. The cheek raiser, for instance, involves contraction of orbicularis oculi and pars orbitalis.

“Our analysis showed that individuals with PD have fewer facial muscle movements than those without PD, Dr. Hoque said, with less variance in almost all of the facial action units involved in each of the three expressions.

“We found that the most important expression for prediction was the smile,” which corroborates previous studies of hypomimia in PD, he said.

That finding is particularly exciting, Dr. Hoque added, because there are 93 million new selfies taken and uploaded every day by the Android system alone, most of which involve smiling, potentially providing (with appropriate permissions) a vast database for population-wide screening.

figure3

“We want to enable screening that is available anytime, anywhere, so anyone in any part of the world can go to a website, perform a subset of the United Parkinsons Disease Rating Scale tasks, and immediately get a referral if their results indicate a high probability of PD.”—DR. EHSAN HOQUE

Unlike genetic testing, which is widely available from private companies without a prescription, PD screening is unlikely to emerge unregulated and outside the confines of the current medical system, Dr. Hoque noted. An individual who learns they may have incipient PD “needs to receive that information in a way that empowers them,” through seeking a referral for definitive diagnosis, for instance, he said.

Beyond screening, the suite of analytic software, once fully validated, may provide the neurologist a way to monitor PD patients remotely, as their disease progresses or as new medications are introduced.

“If the patient has PD, the system may allow the United Parkinson's Disease Rating Scale to be performed at home, without making a trip to the clinic,” Dr Hoque said.

“In the United States, 40 percent of people over age 65 who have PD cannot see a neurologist,” due to distance, insurance, or immobility, Dr. Hoque said. “This could be a way for neurologists to care for more patients, given the limited time they have, to reduce inequities in our health care system.”

Expert Commentary

The remote evaluation of PD patients has promise for under-resourced countries, said Deanna Saylor, MD, an assistant professor of neurology at Johns Hopkins University. “This is a really exciting study, and has some important applications, including perhaps, with further development, the ability to help less experienced clinicians distinguish PD from other parkinsonian syndromes.”

Addressing inequities in the global health system is part of what led Dr. Saylor to Zambia in the center of southern Africa, a country of 18 million people. When she arrived four years ago, there was not a single Zambian neurologist, and only four ex-patriate American and European neurologists who provided outpatient neurological care one day per week. Dr. Saylor began working in a public hospital in the capital, Lusaka, focused on inpatient care and training of Zambian physicians in neurology clinical care and research.

Given that we lack ancillary tests such as DaTscans to confirm the diagnosis of Parkinson's disease in our setting, she said, this screening software could potentially be repurposed as a clinical test to provide additional supportive data for the diagnosis of Parkinson's disease in our setting and provide feedback to trainees about the likelihood of Parkinson's diagnosis that they are considering.

However, she noted, the introduction of a screening tool for PD into a country hugely underserved by neurologists may just move the problem further down the line. “My concern is that if someone screens positive for a high risk of Parkinson's disease, what is the next step? For a general practitioner with a very limited knowledge of neurology, how will they manage the patient?”

At the same time, Dr. Saylor said, technology has played an important role in expanding the neurology capability out from her clinic. Her center is linked to two other hospitals away from the capital, through which she and a handful of colleagues can see patients remotely. They also developed a basic teleneurology clinic using primarily phone calls to provide care to patients from their own clinic when it was closed during COVID-19 surges.

“We are starting to experiment with teleneurology as a forum for training as well, so that we will see the patient with an internal medicine physician in the room. That physician will assist with the evaluation, and then talk through the differential diagnosis and management of the case with us, thereby increasing their neurology skills. This is why training is so important—it's not just the patients I can see, but the patients those I've trained can see.”

“This kind of [computer-assisted screening] system could be very helpful for epidemiological studies,” commented Alberto J. Espay, MD, FAAN, professor of neurology and chair of the Gardner Center for Parkinson's Disease at the University of Cincinnati, and former head of the International Parkinson's and Movement Disorders Society Task Force on Technology. “It would be a valuable way to understand the population prevalence of Parkinson's disease, but perhaps not for trial recruitment purposes,” he said, as hypomimia itself can be a feature of other parkinsonian disorders as well.

The entire suite of tools developed by the Rochester team might also find use in the general neurologist's office, he said. “It could help narrow down the set of tests one would consider doing at the clinic, and that could be helpful.”

Nonetheless, Dr. Espay noted, the technology in this case is only echoing the clinical exam, and its final result has the same limitations. “We know that no clinical feature of PD correlates with any specific pathologic abnormality,” he said. “Two people with the same symptoms may have different pathology, and two people with the same pathology may have different symptoms. In this way, the clinical category of “Parkinson's disease” is “an artificial, convenient creation that helps with classification efforts but not with disease-modifying treatments, the next level of therapeutic endeavor.”

Dr. Espay said one unrealized promise of technology would be “to question our beliefs, and probe into living biology, not to correlate with behaviors or to confirm the fairy tale of Parkinson's disease that we created two centuries ago, but to identify the many biological expressions of Parkinson's in those affected.”

Disclosures

Drs. Hogue and Saylor had no disclosures. Dr. Espay has received grant support from the NIH and the Michael J Fox Foundation; personal compensation as a consultant/scientific advisory board member for Neuroderm, Neurocrine, Amneal, Acadia, Acorda, Bexion, Kyowa Kirin, Sunovion, and Supernus (formerly, USWorldMeds); honoraria from Acadia, Sunovion, Amneal, and Supernus; and publishing royalties from Lippincott Williams & Wilkins, Cambridge University Press, and Springer. He cofounded REGAIN Therapeutics (a biotech start-up developing nonaggregating peptide analogues as replacement therapies for neurodegenerative diseases) and is co-owner of a patent that covers synthetic soluble nonaggregating peptide analogues as replacement treatments in proteinopathies. He serves on the editorial boards of the Journal of Parkinson's Disease, Journal of Alzheimer's Disease, European Journal of Neurology, and JAMA Neurology.

Link Up for More Information

• Ali MR, Myers T, Wagner E, et al. Facial expressions can detect Parkinson's disease: Preliminary evidence from videos collected online https://www.nature.com/articles/s41746-021-00502-8. NPJ Digit Med 2021;4(1):129.