For reasons not fully understood, Parkinson’s disease is the fastest growing neurological disease in the world. That is bad news considering how debilitating the condition is, and that there is no curative treatment available. Fortunately, the future looks brighter, with research in the field adding to our understanding of the disease by the day. But until a cure is finally found, physicians and their patients will have to manage Parkinson’s disease the best way we presently know how.
To give patients the best chance at improving their quality of life, the key is early detection. That is not as easy as it sounds, however. Diagnosis of the disorder is challenging, and requires a physical evaluation by a specially trained neurologist. The time and cost — and in many cases, lack of available specialists — can be a significant barrier standing in the way of someone seeking a diagnosis, especially those that are not yet symptomatic. As such, people are frequently diagnosed only after the disease has already progressed significantly.
A high-level overview of the approach (📷: T. Adnan et al.)
But those barriers may disappear in the near future, thanks to the work of a group of researchers at the University of Rochester. They have developed an algorithm that can detect Parkinson’s disease with a high level of accuracy just by listening to a person speak for a few seconds. And they have made this tool available for use via the web, so no trip to the clinic or hefty bill from a neurologist is involved.
Past research has shown that nearly 89% of those with Parkinson’s experience subtle changes in their voice, such as altered speech patterns, reduced clarity, or different breathing and pausing habits before obvious symptoms appear. So the team designed a web-based screening system that prompts users to read two short pangrams (sentences that contain every letter of the alphabet), such as “The quick brown fox jumps over the lazy dog.”
Once the user recites the sentences using a computer’s microphone, an artificial intelligence (AI) model analyzes the recording for speech characteristics associated with Parkinson’s. This model was trained and validated on a dataset collected from more than 1,300 participants, both with and without Parkinson’s. Validation studies showed that the system achieved an accuracy rate of 85.7% in diagnosing the disease.
The performance of the system on external data sets (📷: T. Adnan et al.)
The team achieved this result by leveraging semi-supervised deep learning models such as Wav2Vec 2.0, WavLM, and ImageBind. These advanced models, trained on millions of audio samples, can detect subtle speech patterns that traditional methods might miss. By fusing multiple speech embeddings into a unified representation, the group created a model that outperformed existing diagnostic techniques.
While this AI-based test is not meant to replace a formal diagnosis, it could serve as an accessible, low-cost way to flag individuals who should seek further evaluation. This is particularly valuable in areas with limited access to neurologists, but can be used by anyone, anywhere. If you would like to take the test for yourself, check out the researchers’ website for more information.