A quick, speech-based AI tool offers a new way to screen for a key indicator of the neurodegenerative disease.
Computer scientists at the University of Rochester have developed an AI-powered, speech-based screening tool that can help people assess whether they are showing signs of Parkinson’s disease, the fastest growing neurological disability in the world. A study published in the journal npj Parkinson’s Disease introduces a web-based screening test that asks users to recite two pangrams—short sentences using all 26 letters of the alphabet. Within seconds, the AI analyzes the voice recordings for subtle patterns linked to Parkinson’s, with nearly 86 percent accuracy.
Parkinson’s disease is typically diagnosed by movement disorder specialists—neurologists with specific training to evaluate complex motor symptoms—using a combination of family history, neurological examinations, and brain imaging. While the study’s authors emphasize that their AI-based tool is not a substitute for a clinical diagnosis, they see it as a fast, low-barrier, and accessible way to flag people, especially in remote areas, who might be living with the condition and encourage them to seek more thorough clinical evaluations.
“There are huge swaths of the US and across the globe where access to specialized neurological care is limited,” says Ehsan Hoque, a professor in Rochester’s Department of Computer Science and co-director of the Rochester Human-Computer Interaction Laboratory. “With users’ consent, widely used speech-based interfaces like Amazon Alexa or Google Home could potentially help people identify if they need to seek further care.”
To train and validate the tool, the researchers collected data from more than 1,300 participants—with and without Parkinson’s—across diverse environments, including home settings, clinical visits at the University of Rochester Medical Center, and the InMotion Parkinson’s disease care center in Ohio.
Using the computer’s microphone, users simply read aloud two sentences: “The quick brown fox jumps over the lazy dog. The dog wakes up and follows the fox into the forest, but again the quick brown fox jumps over the lazy dog.” By leveraging the power of advanced semi-supervised speech models trained on millions of digital audio recordings to understand the characteristics of speech, the tool can glean enough vocal cues from those two short sentences to flag warning signs.
“These large audio models are trained to understand how speech works; for example, the way someone with Parkinson’s would utter sounds, pause, breathe, and inadvertently add features of unintelligibility is different in someone without Parkinson’s,” says Abdelrahman Abdelkader, a computer science master’s degree student in Hoque’s lab and one of the two lead authors of the study. “If a person is saying the pangram that contains the full spectrum of the alphabetical variability and trails off at certain points, the model can tell if that’s different from the typical representation and flag it.”
The tool was 85.7 percent accurate when tested, providing a strong indication of whether someone may have Parkinson’s. But it is a multifaceted disease, and while some people demonstrate symptoms through speech, they can also display signs through motor tasks or facial expressions. Over the last decade, Hoque’s lab has pursued clever algorithms to combine multiple indicators and produced state-of-the-art results.
“Research shows that nearly 89 percent of people with Parkinson’s have a deformity in their voice that can be indicative of the disease, making speech a strong starting point for digital screening,” says Tariq Adnan, a computer science PhD student affiliated with Hoque’s lab and another lead author of the study. “By combining this method with assessments of other symptoms, we aim to cover the majority of people through our accessible screening process.”
An interactive demo of the lab’s three screening tests, including the speech test outlined in the paper, is available online.
The other authors of the paper include PhD students Md. Saiful Islam, who co-supervised the work with Hoque, Zipei Liu, Ekram Hossain, and Sooyong Park.
The study was funded by the National Institute of Neurological Disorders and Stroke of the National Institutes of Health, the Gordon and Betty Moore Foundation, a Google Faculty Research Award, and a Google PhD Fellowship.