Automatic detection of sleep stage delineation using deep learning and visual representation of time-frequency analysis of EEG signals

Abstract

Sleep staging is a fundamental component of polysomnography and the clinical diagnosis of sleep disorders, yet it remains a time-consuming, subjective, and costly process. Despite decades of refinement, the current gold standard for manual sleep scoring continues to suffer from unacceptable inter-scorer variability, driven in part by the complexity and redundancy of polysomnography data and the subjective nature of visual electroencephalogram interpretation. While automated and rule based systems have attempted to standardise scoring using guidelines such as those of the American Academy of Sleep Medicine, these approaches lack adaptability and fail to resolve fundamental disagreements between scorers. More recent data driven machine learning and deep learning approaches have improved performance, but still present limitations in interpretability, clinical trust, and agreement levels comparable to expert consensus.

This thesis addresses these limitations by proposing novel methodologies for sleep stage transition delineation, electroencephalogram data simplification, feature extraction, and interpretable automated sleep staging. First, the thesis situates current sleep scoring practice within its historical and methodological context, reviewing state-of-the-art signal processing and classification techniques and evaluating widely used datasets. The work highlights persistent challenges associated with noisy and redundant electroencephalogram data, motivating the need for robust data reduction strategies.

Building on limitations in current scoring standards, the thesis introduces a transition-centric perspective on sleep staging by explicitly modelling the temporal occurrence of sleep stage transitions, an aspect not captured by conventional epoch based scoring. Time–frequency ridge analysis characterisation reveals consistent and interpretable frequency patterns at N2 ↔ N3 transitions across large, independent datasets comprising approximately 13,888 hours of sleep data. This approach provides a solid framework for generating interpretable features that characterise transitional dynamics and aims to improve inter-scorer reliability.

A novel time–frequency ridge-based electroencephalogram simplification and visualisation framework is then introduced to reduce redundancy while preserving physiologically meaningful information. Using identical statistical features and classification models, a direct comparison between raw electroencephalogram features and ridge-derived features demonstrates significant improvements in mean sensitivity (12.68%), specificity (3.12%), and positive predictive value (12.02%) on a large, balanced test set. These results indicate that targeted data simplification can reduce subjectivity and improve both automated and clinician-led sleep stage classification.

The thesis further explores phase-based electroencephalogram features to address emerging evidence that different brain regions may occupy different sleep stages simultaneously. Phase relationships across inter-electrode and sub-band electroencephalogram signals are shown to achieve classification performance (80.4% accuracy& Cohen’s Kappa 0.75) comparable to upper bounds of reported inter-scorer reliability, without over-fitting.

Finally, the thesis presents a methodological proof of concept for integrating simplified, multi-channel electroencephalogram representations into deep learning frameworks. Techniques such as colour-map transformation, pseudo-RGB encoding, and image packing are shown to preserve inter-channel phase information while enabling efficient convolutional processing. By embedding explainable model outputs back into standard clinical data formats, the work outlines a pathway toward interpretable, clinically compatible AI-assisted sleep scoring systems.

Overall, this thesis contributes novel, interpretable methodologies that address long-standing challenges in sleep staging, offering a foundation for improving reliability, efficiency, and clinical trust in automated sleep analysis, with broader implications for multi-signal biomedical data interpretation.

Thesis embargoed until 31 May 2028

Date of Award	May 2026
Original language	English
Sponsors	European Sleep Research Society & Department for the Economy
Supervisor	Raymond Bond (Supervisor), Dewar Finlay (Supervisor) & Pardis Biglarbeigi (Supervisor)