(© Ivelin Radkov - stock.adobe.com)
Research suggests AI assistants like Alexa or Google Home could be pivotal in early detection of the neurological disease.
In A Nutshell
- AI can detect Parkinson’s disease by analyzing how someone reads the sentence “the quick brown fox jumps over the lazy dog.”
- The system achieved 85.7% accuracy, surpassing non-specialist doctors and approaching expert performance.
- It works using a simple web-based microphone recording and could dramatically expand early screening access.
- Limitations include language constraints, internet access needs, and a 25% false negative rate.
ROCHESTER, N.Y. — Artificial intelligence can now spot Parkinson’s disease just by listening to someone read “the quick brown fox jumps over the lazy dog” aloud. The breakthrough technology correctly identified the neurological condition 85.7% of the time, potentially offering hope to millions who lack access to specialized medical care.
University of Rochester researchers discovered that subtle changes in speech patterns may reveal Parkinson’s before the telltale tremors and stiff movements appear. Their AI system analyzes vocal quality, rhythm, and pronunciation when people recite the familiar sentence that contains every letter of the alphabet. Findings from their research are published in npj Parkinson’s Disease.
The timing couldn’t be more urgent. Parkinson’s cases are expected to double by 2030, yet entire regions face severe shortages of neurologists. In 2014, Bangladesh had only 86 neurologists for over 140 million people, while some African nations had one neurologist per three million residents.
“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, in a statement. “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.”
How AI Listens for Early Signs of Parkinson’s
Speech problems affect as many as 89% of people with Parkinson’s, according to previous studies cited by the researchers. These vocal changes often emerge before physical symptoms become obvious. The team collected recordings from 1,306 people — 392 with Parkinson’s and 914 without the condition — across homes, clinics, and care facilities.
Instead of listening for what people said, the AI focused on how they said it. The system combined three powerful speech recognition technologies, each trained on massive amounts of human speech data. When working together, these programs could detect voice patterns invisible to human ears but characteristic of neurological changes.
Participants simply used a web-based platform to record themselves reading the sentence on any computer with a microphone. The AI automatically analyzed their speech for Parkinson’s-related vocal signatures.
AI Outperforms Many Human Doctors
The AI performed better than many human doctors. According to a study cited by the researchers, non-specialist physicians correctly diagnose Parkinson’s about 73.8% of the time, while movement disorder specialists reach 79.6% accuracy. By comparison, the AI system achieved 85.7% accuracy on its internal test dataset.
Importantly, the researchers found no statistically significant bias across key demographic groups such as sex, age, or ethnicity. However, deeper error analysis revealed elevated misclassification rates in certain groups — particularly older women and men in transitional age ranges — indicating areas where the model’s performance could be improved.
When researchers tested the AI on entirely new datasets from medical facilities it had never seen before, performance remained clinically meaningful. Accuracy dropped slightly to between 70% and 75%, and AUROC scores ranged from 78% to 82%, suggesting the technology could still function effectively outside controlled research settings.
Revolutionary Impact on Global Healthcare Access
The economic burden of Parkinson’s reached $52 billion annually in the United States in 2017, with costs expected to exceed $79 billion by 2030. Early detection could dramatically reduce these expenses while improving patient outcomes, since treatments work best when started early.
The web-based design means the technology could theoretically reach anyone with internet access. Someone experiencing subtle speech changes could complete the screening at home and receive results that might prompt them to seek medical evaluation.
Rural communities and developing countries stand to benefit most. Rather than traveling hundreds of miles to see a specialist, people could receive preliminary screening locally, with positive results triggering referrals for comprehensive evaluation.
The system has important limits. It currently works only with English speakers, though researchers plan to expand to other languages. It also requires computer access and stable internet connections, potentially limiting use in the communities that need it most.
More fundamentally, not all Parkinson’s patients develop noticeable speech changes, especially in early stages. The technology missed about 25% of people with the disease, meaning it works better as a screening tool than a definitive diagnostic test.
The AI also struggled with certain age groups where normal aging affects voice quality. These limitations mean any real-world deployment would need careful safeguards to prevent misdiagnosis and ensure people understand the results aren’t final medical verdicts.
Rather than replacing doctors, this technology could serve as a powerful early warning system in the global fight against Parkinson’s disease. As populations age worldwide, having a simple speech test could help catch this devastating condition before it’s too late to make a difference.
Paper Summary
Methodology
Researchers collected speech data from 1,306 participants (392 with Parkinson’s disease, 914 without) across three different recording environments: participants’ homes using a web-based platform, clinical settings, and a Parkinson’s care facility. Participants read aloud the English pangram “the quick brown fox jumps over the lazy dog” while being recorded. The team used three advanced AI speech models (WavLM, ImageBind, and Wav2Vec 2.0) to extract complex speech features, then developed a novel “projection-based fusion architecture” that combines insights from these models. The data was split into training (70%), validation (15%), and testing (15%) sets, with extensive bias analysis across demographic groups.
Results
The best-performing fusion model achieved 85.7% accuracy and 88.9% AUROC (Area Under Receiver Operating Characteristic curve) in detecting Parkinson’s disease. The system showed 75% sensitivity, 91.08% specificity, and demonstrated no statistically significant bias across sex, ethnicity, or age groups. When tested on external datasets from different clinical environments, the model maintained respectable performance with AUROCs of 82.1% and 78.4%. An additional generalization test using continuous speech resulted in 74.1% accuracy and 77.4% AUROC. Notably, the AI’s accuracy compared favorably to non-expert clinicians (73.8%) and approached the performance of movement disorder specialists (79.6%).
Limitations
The study had several important limitations. The technology currently only works with English speakers, limiting global applicability. The dataset was imbalanced with significantly more control participants than those with Parkinson’s disease. The system showed a 25% false negative rate, meaning it missed one in four people with the condition. Performance varied across specific demographic subgroups, with higher error rates in certain age ranges for both men and women. The technology requires access to computers and stable internet connections, potentially limiting use in underserved areas where it’s most needed.
Funding and Disclosures
The research was supported by the National Institute of Neurological Disorders and Stroke of the National Institutes of Health under award number P50NS108676, the Gordon and Betty Moore Foundation, and a Google Faculty Research Award. One author was supported by a Google PhD Fellowship. All authors declared no competing interests.
Publication Information
This study was published in npj Parkinson’s Disease, Volume 11, Article 176, in 2025. The paper was titled “A novel fusion architecture for detecting Parkinson’s Disease using semi-supervised speech embeddings” and was authored by Tariq Adnan, Abdelrahman Abdelkader, Zipei Liu, Ekram Hossain, Sooyong Park, Md Saiful Islam, and Ehsan Hoque from the University of Rochester and other institutions. The study was approved by the Institutional Review Board of the University of Rochester and University of Rochester Medical Center.
