By way of introduction, could you give us a brief introduction into your research interests and role at The Oxford Centre for Industrial and Applied Mathematics
In brief I am interested in using and developing new mathematical tools to analyse data. The data may be in the form of signals collected over some period of time, images, or simply numbers: often we have a mixture of all these. Although I am primarily interested in medical applications, interestingly, the data may come from any domain – the mathematical techniques to process the data are often very similar. My research is inherently multi-disciplinary by nature: I use statistics, signal processing, and artificial intelligence tools to discover concealed relationships amongst the measured quantities.
I was a D.Phil. student (2008-2012) until recently in the Oxford Centre for Industrial and Applied Mathematics. I have now started my post-doctoral research fellowship based jointly in Applied Mathematics and the Institute of Biomedical Engineering, here at Oxford. At this point let me thank the Engineering and Physical Sciences Research Council (EPSRC) who funded both my D.Phil. and now post-doctoral research, and also Intel Corporation, who has partly funded my D.Phil. research.
You have recently achieved excellent results in monitoring and diagnosing Parkinson’s disease – can briefly explain how this technology works?
This work builds on the premise that speech signals carry clinically useful information (a) to differentiate people with Parkinson’s disease from healthy controls, and (b) the severity of the symptoms is reflected in speech performance degradation. I have to admit I was very sceptical how well this could work in practice when I started my D.Phil. four years ago!
We all have an idea of the characteristic Parkinsonian tremor: this is caused because the orchestration of muscle movement is affected, mainly as a result of impaired neuron function in an area of the brain called basal ganglia. Similarly, the co-ordination of the muscles responsible for producing speech is affected. The subtle changes in the voice may not be as pronounced as the tremor observed at the limbs, and they are often very difficult to detect perceptually; we have developed some mathematical algorithms which can accurately determine patterns in the speech signals which reflect those pathological effects. Then, we use some machine learning tools to determine the probability that someone’s voice has patterns similar to a Parkinson’s disease subject or a healthy subject. Similarly, we can use the computed patterns from someone’s voice to determine how closely they resemble patterns from a large database of voices that are collected from people on various Parkinson’s disease stages.
Your work highlights the use of telemonitoring, where the patient is not based at a medical clinic and communication between them and a medical expert is conducted via a mobile. Do you think that this could have potential to be implemented in third world countries where people do not have easy access to a healthcare facility?
I have to say that during my D.Phil. I looked at high quality speech signals collected using a dedicated telemonitoring device or collected at acoustically controlled environments; only recently I have started looking at more inexpensive ways of collecting data (not only speech), for example using mobile phones and tables such as ipad. Currently, I have only computer simulations which demonstrate that using speech signals and the current techniques may work well in a setting where the subject uses a mobile phone instead of the dedicated telemonitoring device. However, these simulations make certain assumptions, and we will need to conduct a detailed study to determine whether how well those findings generalize in practice.
I am optimistic and believe the developed technology could be useful in cases where access to specialised healthcare is difficult: in addition to third world countries, people who live far away from health centres, or those who find it awkward to physically visit the clinic. Nevertheless, I would like to stress that we need to conduct more long term studies and have a fairly large number of questions to answer before this technology becomes clinically reliable and useful in practice. For example, we need to study subjects who may be suffering from diseases that have symptoms similar to Parkinson’s disease.
What were the major obstacles you encountered in your research – funding, finding subjects etc?
I was extraordinarily privileged in all those respects: the project was excellently planned by an international group of experts in the US who took care of details that I only became aware of when reading their reports. I am very happy the data collection was completed at the time I started my research – I only had to work on the mathematical analysis and extracting information from the data. And of course, quite importantly for a D.Phil. student, the funding was secured for four years, which allowed me to focus exclusively on my work. Moreover, my supervisors (in particular Max) and collaborators were extremely friendly and patient. I have come to appreciate all these critical aspects in the process of my D.Phil. when seeing other students struggling e.g. with data collection, faulty equipment, and insecure funding.
I think the greatest obstacle for me was mastering topics in different disciplines and studying the literature spread across different research communities: for better or for worse it is very difficult to find someone who mastered all the topics involved in this project. At first I had to get a solid background on the basic neurophysiology, then I read a lot of studies about speech and processing speech signals, then it was mathematics, signal processing, and statistical machine learning. During all those years, I had to improve my programming skills as well, and quite often I had been frustrated when some results and experiments did not go as well as I had expected. Also, when I started my D.Phil., I had some trouble with the terminology and how specific words were used to mean different things to different disciplines – or even to people from the same field. But I am very glad it all worked out in the end!
Where do you think the emphasis lies in future research in voice analysis technology/Parkinson’s disease?
It is quite difficult to say, but I believe many recent studies, including my D.Phil., make a very compelling case for using nonlinear tools in studying pathological voices in general. Perhaps this is getting too detailed, but it appears that patterns on energy in different frequency bands in speech signals appear to give clinically important information that some standard speech processing algorithms had hitherto failed to capture. I am sure we (applied mathematicians/engineers/statisticians) can further improve the arsenal of tools in order to extract more useful information out of the collected signals.
Nowadays we can collect high quality signals using mobile phones or smart tablets – I think researchers are becoming increasingly aware of the vast opportunities both in biomedical applications and other fields. We are developing a new monitoring system for Parkinson’s disease based on such technology, but I would not like to reveal more details now. It is possible that in the not too distant future smart phones and tablets will substitute frequent physical visits to the clinic!