26 June 12 | By Liat Clark
A UK mathematician has made a public appeal for people to phone a dedicated number so data can be gathered to hone a tool that can diagnose Parkinson’s disease by analysing voice patterns.
Max Little, a research fellow at the Massachusetts Institute of Technology, made the announcement during the opening of the TEDGlobal conference in Edinburgh, 25 June. While studying at Oxford University, Little developed an algorithm that identifies the unique characteristics present in the voice of a Parkinson’s Disease sufferer. He setup the Parkinson’s Voice Initiative in order to improve upon the machine learning system — the algorithm is built to adapt when new information is introduced and, by widening the pool (it’s hoped, with 10,000 phone calls form the public), it should become a more accurate diagnosis tool, able to identify specific symptoms amid numerous variants of speech.
"This raises a very interesting possibility," Little says in a promotional video. "If we could use the entire existing telephone network then we could scale up the screening of Parkinson’s disease to the entire population, and do it at very minimal cost."
Other than the UK, there are phone numbers on the Parkinson’s Voice Initiative website for people in the US, Brazil, Mexico, Spain, Argentina and Canada. Parkinson’s sufferers and non-sufferers are both encouraged to call in anonymously. The calls should only last around three minutes. By getting non-sufferers to call in, the system can learn to weed out and discard unnecessary voice patterns, such as those brought on by a cold or heavy smoking.
Around 70-90 percent of sufferers report instances of vocal impairment following the onset of the disease. Little’s proposal therefore presents opportunities for widespread remote diagnosis.
He first presented the diagnosis tool’s successful testing in a paper published earlier this year in the IEEE Transactions journal. Little and co-author Athanasios Tsanas explained how 43 candidates were asked to hold one sound frequency for as long as possible. They collected 263 data samples in this way, and from this extracted 132 different vocal impairments. Using only ten of these recorded impairments, the algorithm could diagnose Parkinson’s speech markers accurately 99 percent of the time. The system is trained to identify the anomalies in the speech.
By collating more data in the future, the range of these vocal features could be widened, lessening the margin of error even more.
The paper suggests that in the future, data could be collected using Intel’s At-Home Testing Device, a telemonitoring system. It would then be sent to a clinic where the algorithm processes it and maps out the speech, identifying markers on the Unified Parkinson’s Disease Rating Scale (UPDRS) so that the severity of the illness is known. In this way, the system could not only be used to diagnose, but to monitor the progression of the disease.
Voice recognition could be a cheap and efficient alternative to having patients’ head to their GP for a twenty-minute diagnosis session. There is currently no simple diagnosis tool — no blood test that can identify Parkinson’s — and vocal tremors, breathiness and reduced speech volume are some of the first symptoms recorded in nearly all patients. These can be very subtle at the start, however, and systems such as Little’s could conceivably pick up the slightest abnormal intonation.
Parkinson’s Disease is the second most common neurodegenerative disorder after Alzheimer’s and, since it can only be treated with drugs or surgery and cannot be cured, early diagnosis can massively effect an individual’s quality of life.