Source-free Domain Adaptation for Sleep Stage Classification
Abstract
The popularity of machine learning algorithms has increased in recent years as data volumes have risen, algorithms have advanced, and computational power and storage have improved. EEG-based sleep staging has become one of the most active research areas over the last decade. Labeling each sleep stage manually is a labor-intensive and time-consuming process that requires expertise, making it susceptible to human error. In the meantime, training models on an unseen dataset remains challenging due to physiological differences between subjects and electrode sensor configurations. Unsupervised domain adaptation approaches may provide a solution to this problem by borrowing knowledge from a labeled dataset to train an unlabeled dataset. A source-free unsupervised domain adaptation methodology is employed in this thesis to solve the problem of automatic single-channel EEG sleep stage classification. Our study shows that pre-training source domain models followed by supervised fine-tuning improve the learned representations when applied to EEG sleep signals. We further develop weighted diversity loss in order to achieve a model that outperforms state-of-the-art unsupervised domain adaptation techniques without access to source domain data.