Time-Frequency Ridge Analysis of Sleep Stage Transitions

The development of automated sleep apnea detection algorithms is an emerging topic of interest [1], [2]. The main aim of automation is to reduce the time and cost associated with manually scoring polysomnogram (PSG) tests [3]. To automate the process, traditional algorithms attempt to mimic the human observer by implementing a series of predefined rules, such as the American Academy of Sleep Medicine's (AASM) scoring guidelines [4]. Recently, data driven methods have emerged [5]. Electroencephalogram (EEG) frequency is known to be an important feature for both the human observer and data driven methods for sleep staging classification. This study presents the initial findings for a novel approach to sleep stage analysis. EEG time-frequency analysis is used to characterise the dominant frequency with respect to time, specifically at the point of sleep stage transition. Poor inter-scorer agreement at sleep stage transitions is a noted limitation of current manual and automated methods as the point of transition is poorly defined [6]. The goal of this study is to further discuss on the topic of sleep staging automation and explore alternative and novel features to improve the inter-scorer reliability of sleep staging.

Clinically annotated PSG data were acquired from the “You Snooze You Win: The PhysioNet/Computing in Cardiology Challenge 2018” [7, 8]. EEG channel OI-M2 was selected for initial investigation. The dataset contains 994 overnight PSG recordings of approximately eight hours each, with clinical annotations provided. Six sleep stage annotations were possible (N1, N2, N3, R, U, W) Stages N1, N2 and N3 represent progressively deeper sleep, R represents REM sleep, U represents an undefined sleep stage, used for clinical ease and W represents wakefulness. All 994 recordings were used in this study. Sleep stages were scored in 30 second windows as recommended [4]. Given the annotations provided, six sleep stages were possible, for a total of 36 transition categories. Sleep stage transitions were grouped by category as shown in Table 1. It should be noted that successive annotations occurred e.g., NI→ NI in the data, however, as no sleep stage transitions occurred these instances were excluded from analysis. Table 1. Exhaustive sleep stage transition categories

	N1	N2	N3	R	U	W
N1	N1 →N1	N1 →N2	N1 →N3	N1 →R	N1 →U	N1 →W
N2	N2 →N1	N2 →N2	N2 →N3	N2 →R	N2→ U	N2→W
N3	N3 →N1	N3 →N2	N3 →N3	N3→ R	N3 →U	N3→W
R	R→N1	R→N2	R→N3	R→R	R→U	R→W
U	U→N1	U→N2	U→N3	U→R	U→U	U→W
W	W→N1	W→N2	W→N3	W→R	W→U	W→W