Early last week, we posted a TED talk by Max Little, Director of the Parkinson’s Voice Initiative. The effort seeks to increase the accuracy of detecting Parkinson’s disease using a phone-based voice diagnostic. We had the chance to sit down with Dr. Little last week to discuss this initiative.
Ravi Parikh, Medgadget: We recently featured the TEDGlobal talk you gave this past year on the Parkinson’s Voice Initiative. Could you give us a brief overview of the technology behind it and how it came about?
Dr. Max Little: I was working on my PhD at Oxford in clinical voice analysis using advanced mathematical algorithms. I had a chance encounter with a researcher at Intel Corporation. One of the co-founders of Intel was diagnosed with Parkinson’s and, realizing that there are no objective methods that could be used to score the symptoms over time, he put some of his own money into Intel to develop a testing device that could do just that. They recruited about 50 people into a 6-month trial to test the effectiveness of the device, where the subjects took the tests every week. However, they did not really know what to do with the nearly 7,000 voice recordings that they captured. Using some voice data from subjects that was blinded as to whether they had Parkinson’s or were healthy controls, we used my algorithms to see whether we could predict which had the disease. We found that my algorithms were around 86% accurate without any modification. At that point we decided that maybe there was something useful to the algorithms outside of voice disorders!
The basic algorithms function by first extracting characteristics of the digital voice signal. We exclusively use recordings of the vowel sound ‘aaaaah’. The characteristics include measures of vocal tremor, vocal weakness and breathiness, and measurements of fluctuations in the jaw, tongue and lips. These characteristics have been shown to have correlates in the typical limb tremor, weakness and rigidity of Parkinson’s. Then, statistical machine learning algorithms learn to associate these characteristics with the clinical outcome for that subject (e.g. whether the subject has Parkinson’s or not, or symptom severity measured on a standard clinical scale such as the Unified Parkinson’s Disease Rating Scale). The machine learning algorithms are sufficiently flexible that they can adapt to known confounding factors: whether someone is a heavy smoker or has had vocal fold surgery, for example, conditions that are known to affect voice production. To make a prediction about a clinical outcome for a new subject, these characteristics are extracted from the digital vowel sound ‘aaah’, and the machine learning algorithm outputs a clinical prediction.
Currently, we achieve 98.6% overall accuracy using lab-based recordings, broken down into 99.2% sensitivity and 95.1% specificity.
Medgadget: The concept of a simple, non-expert based test is truly innovative. Can you tell us why a voice-based diagnostic is better than a clinical diagnosis by a neurologist?
Dr. Little: I wouldn’t claim that these are better than a neurologist’s diagnosis. Ultimately the diagnosis rests with the person who has to make clinical treatment decisions, so what we are doing is making highly accurate clinical decision support tools. However, these tools are very cheap and can be administered remotely from the clinic, using ubiquitous hardware (the telephone). This makes it possible to automate the process of clinical decision-making about diagnoses and treatments. This automation could have a number of uses, including: reducing logistical difficulties for patients who can radically reduce the need to visit the clinic for checkups; and reducing the cost of mass recruitment into clinical trials for new treatments because the follow-up stage can be done remotely. There are also situations where the ratio of neurologists to patients is very low – in the developing world for example – and here the ability to detect the disease remotely and cheaply using an existing technological infrastructure (the telephone, accessible to 75% of the world’s population), could dramatically reduce the total number of undiagnosed sufferers of this manageable disease.
The technology’s outputs correlate extremely well with externally validated Parkinson’s symptom scales. Our technology predicts the total and motor components of the Unified Parkinson’s Disease Rating Scale, used by neurologists as a standard symptom scale, within about 2% of the neurologist’s score. We are thus confident that the device can also be used to monitor disease progression.
Medgadget: Give us an overview of the goals of the Parkinson’s Voice Initiative. Where do you stand now in terms of meeting those objectives, and what do you need from the public/academic community to achieve them?
Little: The Parkinson’s Voice Initiative seeks to test out the hypothesis that the lab-based algorithms we have developed can be applied in the real world, using the telephone. We conducted a feasibility study in which we simulated the typical audio bandwidth reduction and channel distortion of the mobile telephone and this seemed to have little negative impact on our ability to accurately determine the symptoms of the disease. So this gave us encouragement that the methods might work in the real world over the standard telephone network.
What we need from the public are participants: this is a simple trial in which subjects, both healthy and with Parkinson’s, call up a local number in their country and donate a voice recording. So far, we have had around 8,000 contributors in around 7 weeks since launching the project, so we are well on our way to reaching the 10,000 we think we need in order to detect statistically significant and clinically important differences between healthy controls and those with Parkinson’s.
In the academic community we are interested in reaching out to clinician-scientists such as neurologists, who might want to test out the technology in their own clinical settings and research trials. We are of course interested in talking with pharmaceutical organizations about the potential for collaboration on clinical trials.
Medgadget: You spoke in your TED talk about scaling this technology to mobile phones. What is the status of that effort, and are there certain formats the voice file must be in to be analyzed?
Little: It’s quite important to realize that this technology does not require a specific kind of audio recording hardware or software. It could just as easily function in a smart phone as in a standard telephone system. But smart phones can do a whole lot more than just record the voice, they also have the potential to carry out some, if not all, of the data analysis on the phone. In terms of scale, it is really just a matter of accessibility to those with the hardware and software necessary to record their voices. We can send the data anywhere and so do the analysis anywhere if necessary, perhaps on a server located on the other side of the world.
Medgadget: And how do you control for environmental noise and general background noise?
Little: Actually, with voice-based recordings over the telephone, background noise isn’t so much of an issue unless you are in an extremely noisy environment. The reason is that telephones are generally close-mic (that is, the microphone is directionally sensitive and very close to the mouth), which means that you don’t need noise-reduction technology for mobile phones. You would, however, need noise-reduction if you were using a computer, which uses an ambient microphone.
Medgadget: Are there other diseases that could benefit from a widely scalable voice-based diagnostic? Do you have plans to expand the scope of the technology in terms of diseases treated?
Little: There are many other neurological disorders that affect movement, all of these might benefit in the same way as Parkinson’s. It’s not clear at this stage exactly how, because research into voice impairment in diseases such as Multiple System Atrophy, Multiple Sclerosis, and Major Depression (see here), is relatively underdeveloped. Nonetheless, if voice is affected in these disorders, it is really only a matter of extracting the same set of characteristics (we have currently tested about 130 of them), and using the same machine learning algorithms to learn the relationship between these characteristics and the clinically-relevant outcomes.
Speech analysis algorithms have also seen some recent impressive successes in detecting and characterizing mood disorders such as depression and anxiety. There is, in my view, a minor revolution under way in the use of clinical voice signal processing in practical applications, made possible by the near total ubiquity and low-cost of communication and computing hardware.
Medgadget: On the PVI website, your biography states that your are “an applied mathematician with a background in digital signal processing and video games coding.” How did you become interested in Parkinson’s and medical technology in general?
Little: I started my career working 10 years in the video games industry as an audio engineer and programmer: I even wrote music for television and radio at the same time. I switched to a career in academia through a mathematics degree studied in night class. I achieved nearly the highest possible 1st-class score in this degree program. I didn’t really know that I had an aptitude and passion for mathematics until then (which runs contrary to what is generally thought true about mathematics: that people find out that they are good at it early on in their schooling). This then won me a PhD scholarship to Oxford where I started concentrating on novel mathematical algorithms for voice signal processing, and gradually developed an interest in clinical applications of voice analysis, because it was there that the algorithms I had been developing (which attempt to characterize phenomena such as chaos and turbulence in the voice) were, it seemed, the most immediately applicable.
I think that in my career I have always been interested in doing highly intellectually demanding work that has a practical application at the end of it. I’m not necessarily that comfortable with doing theoretical research for the sake of it (although I understand that this is what motivates many mathematicians, who are more attracted to the raw beauty of the subject, rather than how relevant that beauty is to the real world). These days I see the potential to really make a difference to the world through the application of advanced mathematics.
____________________________
To illustrate the voice differences between healthy patients and those with Parkinson’s, please check out the voice samples below:
Normal: WAV audio file…
Parkinson’s: WAV audio file…
Flashback: Detecting Parkinson’s with a Phone Call: TED Talk by Max Little