Advisor: Gari Clifford, DPhill – School of Biomedical Engineering, GT

Committee:

Eva Dyer, PhD – School of Biomedical Engineering, GT

Thad Starnder, PhD – School of Electrical & Computer Engineering, GT

Hua Wang, PhD – Department of Information Technology, ETH Zürich

Reza Sameni, PhD – Department of Bioinformatics, Emory

METHODS FOR GENERALIZED LOW-DIMENSIONAL EEG ANALYSIS
USING TRANSFER LEARNING

Polysomnography (PSG) is a widely used procedure for diagnosing sleep disorders such as narcolepsy and sleep apnea, however its invasive and time-consuming nature make it infeasible for any form of long-term monitoring. It requires patients to sleep in a hospital setting for several nights while their EEG and other vitals are continuously recorded, after which a trained human clinician must manually score every 30-second block in the entire recording. Long-term monitoring of treatment effectiveness or disease progression thus needs to be conducted using less reliable methods such as wrist actigraphy, sleep diaries, or subjective surveys of sleep quality. The lack of effective methods for long-term monitoring is also problematic for longitudinal studies examining the interaction between sleep quality and other pathologies such as Alzheimer's. There is thus considerable interest in automated sleep staging using at-home wearable sensors. 

A problem in developing automated sleep staging algorithms however is the lack of data available for wearable sensors. There is plenty of data available for in-hospital PSG, but very little for wearable sensors, as they aren't normally used in clinical practice. A possible solution however is the use of transfer learning - a method of boosting machine learning performance on one task using data from a similar task. 

In this thesis, we use transfer learning to make several advances in the field of sleep staging with wearable sensors: 1) We test a variety of transfer learning techniques under a variety of conditions and neural network architectures to determine which transfer learning method is most effective. 2) We develop a novel transfer learning algorithm augmenting training data with synthetic EEG generated using electrophysiological models designed to output data resembling that of the targeted wearable sensor. 3) We used transfer learning to develop a sleep staging model specifically designed for use on mild cognitive impairment patients which is far more effective than models trained on healthy subjects. 4) We used transfer learning to automate sleep staging with an experimental in-ear wearable which is far more comfortable and user-friendly than scalp wearables yet achieves superior sleep staging performance to some commercially available scalp wearables.