Speech-based Traumatic Brain Injury (TBI) Assessment using Deep Learning Methods with Limited Labeled Data
WPI – Computer Science
Monday, April 26, 2021
Time: 4:00 p.m. – 5:00 p.m.
Zoom link: https://wpi.zoom.us/j/99694629401
Advisor: Prof. Emmanuel Agu
Co-Advisor: Prof. Adam Lammert
Reader: Prof. Elke Rundensteiner
People afflicted with Traumatic Brain Injury (TBI) experience long-term impairments, which require follow-up assessments or rehospitalization in some cases. To promote recovery, continuous monitoring of TBI patients with long-term impairments is an area requiring urgent research in public health care. This master thesis proposes a Deep Neural Network (DNN) system for non-invasive, speech-based assessment of long-term impairments following TBI that runs passively on smartphones. Notably, we tackle the overfitting problem that arises from an insufficient amount of TBI speech, which is infrequently collected and expensive to acquire. Overfitting prevents DNN from learning generalized features of TBI speech and hinders the TBI detection accuracy.
In this master's thesis, we investigate three learning methods that integrate knowledge from other domains to TBI detection to improve our DNN-based TBI assessment accuracy without augmenting the samples. Specifically, we investigate transfer learning, multi-task learning, and meta-learning for improving our proposed cascading DNN for TBI speech assessment that combines sequential features from a backbone model. The results indicate that all three limited labeled data learning methods mitigate the overfitting problem and improve the TBI classification accuracy by 34% and TBI regression error by 31%. Moreover, the results can be extended to estimate data size necessary to obtain a decent performance for each learning method.