Description:

ERIKA EXTON, Linguistics, Psychology, Psychiatry & Behavioral Science

"Balancing Interpretebility, Objectivity, and Automaticity When Using Speech to Study Major Depressive Disorder"

Abstract:

When studying cognition, particularly in the context of sensitive mental health data, it is critical that our results be interpretable — transparently linked to hypothesized cognitive mechanisms. How can we respect interpretability while capitalizing on the efficiency and objectivity afforded by machine learning tools? I’ll discuss these issues in the context of my current research (bridging linguistics, communication science and disorders, and psychology). Using machine learning tools, I objectively and efficiently study speech in individuals with depression in a way that is interpretable, constrained, and theory-driven. Instead of taking a black-box approach to detect whether an individual is depressed from a short sample of speech, I use machine learning tools to extract interpretable and mechanistically-meaningful speech features, relating them to disruptions to motor control that occur in depression (i.e. slowed, jerky movements). I used a custom deep learning algorithm to automatically assess speech features related to motor control and compared those patterns to performance in manual motor tasks. Results indicate that these measures are correlated in individuals who are currently depressed, suggesting that motor control dysfunction in depression may impact speech and manual motor behaviors similarly and that we can objectively and efficiently measure motor dysfunction using speech-based deep learning methods. The next stage of this research is to explore depression-related disruption to other aspects of cognitive functioning and speech using machine learning/AI approaches to efficiently measure a range of linguistic features at scale.