Time–frequency ridge characterisation of sleep stage transitions: Towards improving electroencephalogram annotations using an advanced visualisation technique

• Published in: Expert systems with applications 2025-03, Vol.262, p.125490, Article 125490
• Main Authors: McCausland, Christopher, Biglarbeigi, Pardis, Bond, Raymond, Yadollahikhales, Golnaz, Kennedy, Alan, Islind, Anna Sigridur, Arnardóttir, Erna Sif, Finlay, Dewar
• Format: Article
• Language: English
• Subjects: Automatic scoring; Electroencephalography; Interpretable AI; Sleep staging; Time–frequency analysis
• ISSN: 0957-4174
• DOI: 10.1016/j.eswa.2024.125490

Manual sleep stage scoring of polysomnography recordings is an expensive and time-consuming process, further complicated by inconsistent sleep stage agreement among sleep experts (clinicians and sleep technologists). Hence, development of automated sleep scoring algorithms are an emerging topic of interest. Automation typically mimics the clinical decision path by implementing a series of predefined rules, such as the American Academy of Sleep Medicine’s (AASM) scoring manual. Recently, data driven methods have emerged using machine or deep learning. Both manual and automated methods of scoring have known limitations; primarily, unacceptable variation in agreement between different scorers and algorithms. Within the literature, electroencephalogram (EEG) frequency is an important feature considered by both sleep experts and automated approaches for classifying sleep stages. This study presents a novel approach to sleep stage analysis, by developing a methodology to precisely determine the temporal location of sleep stage transitions. The current gold standard fails to identify such transitional changes, which leads to poor inter-scorer reliability. Therefore, development and implementation of such methodologies is a crucial, but overlooked, step in improving the consistency of scoring within sleep studies. In this work, EEG time–frequency ridge analysis was used to characterise the dominant frequency component of EEG signals in time, at the point of sleep stage transition. An in-depth analysis of N3 → N2 and N2 → N3 transitions in the 2018 PhysioNet challenge “You Snooze, You Win” and the Wisconsin Sleep Cohort (WSC) datasets (n = 994, n = 742; approximately 13,888 h of sleep data) showed consistent time–frequency patterns at the point of transition, from one sleep stage to another. This methodology allows simple and ‘interpretable’ features to be generated in future work, to precisely identify the temporal location of sleep stage transitions with the aim of improving inter-scorer reliability. •No current methodology to delineate sleep stage transition, results in poor agreement.•Characterised sleep stage transitions based on time–frequency ridge methods.•Identified characteristics to delineate sleep stage transitions in independent data.•Implication: Improved inter-scorer agreement by adoption into sleep scoring manual.•Future work will apply the approach to automated sleep staging algorithms.