Development of Speech-Based Neurological Assessment Tools and Biomarkers

Speech production is a very complex process, involving the coordinated involvement of several parts of the brain, along with the proper activation of around a hundred muscles. Connected as it is to many components of the neuromuscular system, it has been shown to be impacted in different ways by several types of neurological conditions. If their impacts on speech have been demonstrated by previous works, their exact nature is still not entirely well defined. Furthermore, the ability to utilize the specific changes certain conditions induce on speech to be able to detect or monitor these conditions has yet to be implemented on a wide scale. Such a solution would consist of a protocol to define the speech samples to collect, an apparatus to allow this capture, a system to extract speech features from the samples, and finally, using these features, a method to classify the different samples by condition.

In this work, we present two different approaches, one to detect mild Traumatic Brain Injuries (mTBI), and another to detect two neurodegenerative conditions: Amyotrophic Lateral Sclerosis (ALS), and Parkinson's. Similar underlying principles are behind both approaches, with one building upon the other. These approaches strive to be portable and accurate and to rely upon well researched speech sampling methods and large data collection efforts.

A series of tests, whose designs are informed by speech language pathology research, is presented to collect feature-rich speech samples. In order to be portable, a mobile application is also created to facilitate the collection of speech samples and associated metadata from over 2,500 student athletes in the context of an analysis of mTBI's impact on speech. We then extract time and spectral domain features, and use logistic regression to create a classification model. With this large dataset of samples, we are able to obtain good classification accuracy between the control and mTBI populations (Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve >0.85).

We use the basic design of the mTBI tests as a starting point and develop them further as we pivot to diagnosing neurodegenerative conditions. A new application is then introduced for the data collection based on these new tests. It includes several improvements in the way the tests are administered, and the data stored, transferred and analyzed. Using this application, we collect speech samples from over 80 participants. We then use XGBoost to create classification models to diagnose the speech samples. With these models, we are able to reach an accuracy ≥80\% in classifying the participants in each of our three classifications (Control vs ALS, Control vs Parkinson's, and Parkinson's vs ALS).