I am a very sceptical person. Particularly, it has to be said, when I get a comment on my blog that effuses about how wonderful I am, and would I like to watch a video?
I also don’t particularly like watching videos. So I didn’t.
I do, however, like reading stuff, and there was enough in that comment that made me think it might be for real, even if the author of the comment did seem to be working for the same Web site to which she was directing me. (The site – called Newsy – seems kosher – MacWorld gave it a nice review). So I plugged a few key terms into a search engine.
I found a BBC article. Being British, I’m quite inclined to trust the BBC, within limits. And – bonus! – it had words to read.
[Mathematician Max] Little has discovered that Parkinson’s symptoms can be detected by computer algorithms that analyse voice recordings.
Apparently, his software can detect tremors in the voice – or, more specifically, in the vocal chords. He is building up a database of voices to maximise the system’s accuracy. Little is reported as saying:
“This is machine learning. We are collecting a large amount of data when we know if someone has the disease or not and we train the database to learn how to separate out the true symptoms of the disease from other factors […] It is not as simple as listening for a tremor in the voice. That tremor has to be in context of other measures and the system has to take in other factors such as if someone has a cold.”
Note that “we are collecting”. This is at the root of the publicity. They want YOU to contribute. With or without Parkinson’s. There are local ‘phone numbers in 10 countries that you can call, anonymously, to leave a voice sample.
“The more people that call in, the better,” [Little] said.
That part of me that is still a scientist needs more than an article in the mainstream media. Where’s the peer-reviewed research paper? Where’s the hard stuff – the abstract? I found a MedicalNewsToday article that quoted the BBC and that mentioned “a paper published earlier this year in IEEE Transactions on Biomedical Engineering“. And that hint led me to it. Not just the abstract. The whole paper! In a handy PDF.
It tells you quite a bit more about how the system works. Probably more than you care to know. I discovered that they are measuring dysphonia in sustained vowel sounds. There were some complicated mathematical terms, and the assertion that this technique could be “an important step towards non-invasive diagnostic decision support in PD.” And that was just the abstract. (Feel free to skip the next bit, which is me trying to make sense of the rest of the paper – the simplified version can be found at the end of this rather long post).
Elsewhere in the paper, the researchers note that:
Research has shown that speech may be a useful signal for discriminating PWP from healthy controls, building on clinical evidence which suggests that the vast majority of PWP typically exhibit some form of vocal disorder. In fact, vocal impairment may be amongst the earliest prodromal PD symptoms, detectable up to five years prior to clinical diagnosis. In our own research, we have also presented strong evidence linking speech to average Parkinson’s disease symptom severity. Collectively, these findings reinforce the notion that speech may reflect disease status, after appropriate processing of the recorded speech signals.
A. Tsanas, M.A. Little, P.E. McSharry, J. Spielman, L.O. Ramig (2012)
Novel speech signal processing algorithms for high-accuracy classification of Parkinson’s disease
IEEE Transactions on Biomedical Engineering, 59(5):1264-1271
I took the liberty of removing the endnote references; they are extant in the PDF.
What kind of dysphonia indicates Parkinson’s, then? It was difficult to work this out, what with all the talk of feature selection algorithms and extensive use of acronyms, but I gather that the following features are the most typically indicative:
- Vocal fold excitation ratio (VFER): Extent of noise in speech using energy, nonlinear energy, and entropy concepts
- Mel Frequency Cepstral Coefficients (MFCC): Amplitude and spectral fluctuations
- Recurrence period density entropy (RPDE): Uncertainty in estimation of fundamental frequency
- Detrended fluctuation analysis (DFA): Stochastic self-similarity of turbulent noise
- Shimmer variants: Amplitude perturbation
In the discussion section, the researchers say:
The pathophysiological importance of signal to noise ratio measures is well-known: it is most likely the effect of amplified aeroacoustic noise due to increased airflow turbulence, ultimately generated by incomplete vocal fold closure.
Now, I can’t say that I completely understood all of that, but what I think they’re saying is that your “vocal folds“, also known as vocal chords, tend not to close properly in Parkinson’s, and that this causes certain types of “noise” (which I interpret as audio frequency interference) to manifest in the voice.
All in all, it sounds very convincing and worthwhile. I particularly like the suggestion (made in the BBC article) that analysis of a simple voice recording could, potentially, be used as a measure of the severity of symptoms – and so as an objective measure of how well medication may be working.
Excuse me. I have a ‘phone call to make.