Advanced Deep Learning Techniques for EEG based Sleep Signal Analysis

Abstract

Sleep signals are crucial for diagnosing sleep disorders and analyzing sleep patterns. Polysomnography uses EEG and other physiological measurements for sleep staging and disorder detection but relies on manual analysis requiring expert knowledge, time, and cost.

While machine learning shows promise in automating sleep analysis, three gaps remain: (1) no research combines traditional signal processing theory with deep learning, (2) scarce public sleep disorder datasets limit model development, and (3) current staging focuses on macrostructure, ignoring microstructure.

This thesis addresses these gaps using single-channel EEG signals. We integrate the Hilbert transform into deep learning architecture for improved sleep staging model design. This architecture is applied to insomnia characterization, extracting features from large datasets and transferring them to smaller insomnia datasets via domain adaptation. We integrate cyclic alternating patterns (CAP), a microstructure biomarker, with sleep stages to capture finer-grained dynamics.

This thesis contributes to automated sleep staging and disorder detection by integrating signal processing theory and sleep microstructure, enabling scalable and precise tools for personalized sleep medicine.