Using Speech Markers to Identify Cognitive, Motor Deficits in Parkinson’s

Full video transcript:

Hi, I am Kara Smith, a movement disorders physician at the University of Massachusetts in Worcester, and I’m going to talk to you about my poster that I presented today about speech markers to detect cognitive impairment in Parkinson’s disease.

So the reason we’re interested in this is that speech markers have been used in many different conditions, and in Parkinson’s disease to detect motor symptoms. However, we know that speech is a cognitively demanding task, and the cognitive impact on the speech markers in Parkinson’s has never really been discussed, so we tried to address this relatively understudied area, and we thought that it would have great potential looking to the future since cognitive impairment is so predominant in Parkinson’s, and has such a dramatically negative impact on our patients living with this disease. However, we do not have adequate tools to detect and monitor the cognitive symptoms in Parkinson’s, and we really need better tools. Speech markers could help fill that large gap.

So the objective of the study was to look at the voice and see if we could detect markers of cognitive impairment in [Parkinson’s disease] PD for the first time. So we had 35 patients with Parkinson’s (PD) who were not demented, and they provided brief speech samples, or free speech. They were asked to describe what’s known as the “cookie theft picture,” which is a very commonly known picture in neurology, which is a structured speech task, and then the patients are also examined for their motor symptoms. They underwent the Montréal Cognitive Screening Assessment, or MoCA, and the Geriatric Depression Scale, or GDS, to look at their other symptoms as well. The patient cohort was relatively mild to moderate disease stage, and they had a normal MoCA score — the average was 27 — and relatively few depressive symptoms.

So we found, actually, that there were motor and cognitive associations with different speech markers. So we looked at the voice in a signal processing approach, and using machine learning technology, we extracted certain features or characteristics that pertained to the resident frequencies of the voice and how those change over time in a dynamic fashion. And what we found was that there were certain resident frequencies, or formants, that were associated with both cognitive and motor symptom severity, but then when we looked at other speech features, such as the phonemes, we found that phoneme rate had a different pattern between cognitive symptom association and motor symptom association. So we found, for example, that a certain phoneme that required a large amplitude jaw movement, like “ah” or “aah,” was more associated with motor symptoms, but much less associated with cognitive symptoms, but then there was a group of phonemes that was more associated with cognitive symptoms compared to motor symptoms, and these were phonemes that required lingual articulation such as a “z” or a “j” sound.

We then created models where we combined all of the different vocal acoustic features that we looked at, and were able to accurately predict a patient’s MoCA score or a patient’s UPDRS score using these models. And so the basic summary of our results is that we were able to identify vocal markers of both motor and cognitive symptoms, but that there were some subtle differences in the patterns or clusters of features that we saw, which suggests that as we do more of this work in the future, we might be able to calculate a patient’s motor score and cognitive score just based off a very brief free speech sample.

So the real area that this is greatly needed for is for early detection of cognitive impairment in Parkinson’s, and very close assessment or monitoring of cognitive impairment in Parkinson’s. So we know that even early on in the disease course, patients with PD have cognitive impairments, but I think we’re missing a lot of the very early, very subtle impairment that happens in our clinical measures just because they’re not really designed to detect early subtle changes. And so I think that the cool thing about our study was that these are non-demented PD patients. They had a normal score on the MoCA, and yet we were able to predict if they would get high normal range or normal range just using their voice. And so I think this might have a place in early detection of cognitive function and cognitive decline, and then to really monitor sensitive changes over time. And that’s needed to really empower our clinical research to develop new strategies to treat these patients with cognitive impairments. I think having a really powerful way to get frequent, remote, detailed, and quantitative measure of cognitive function through voice and through speech could really be a huge boost to these future clinical trials.

For more coverage of MDS 2017, go here.