Detecting REM sleep from the finger: an automatic REM sleep algorithm based on peripheral arterial tone (PAT) and actigraphy

• Published in: Physiological measurement Vol. 28; no. 2; pp. 129 - 140
• Main Authors: Herscovici, Sarah, Pe'er, Avivit, Papyan, Surik, Lavie, Peretz
• Format: Journal Article
• Language: English
• Published: England IOP Publishing 01.02.2007
• Subjects: Adult; Algorithms; Arteries - physiology; Artificial Intelligence; Female; Fingers - blood supply; Fingers - physiology; Humans; Male; Middle Aged; Monitoring, Ambulatory - methods; Movement - physiology; Muscle Tonus - physiology; Muscle, Smooth, Vascular - physiology; Polysomnography; Regional Blood Flow - physiology; Reproducibility of Results; ROC Curve; Sleep Apnea Syndromes - diagnosis; Sleep Apnea Syndromes - physiopathology; Sleep, REM - physiology
• ISSN: 0967-3334; 1361-6579
• DOI: 10.1088/0967-3334/28/2/002

Scoring of REM sleep based on polysomnographic recordings is a laborious and time-consuming process. The growing number of ambulatory devices designed for cost-effective home-based diagnostic sleep recordings necessitates the development of a reliable automatic REM sleep detection algorithm that is not based on the traditional electroencephalographic, electrooccolographic and electromyographic recordings trio. This paper presents an automatic REM detection algorithm based on the peripheral arterial tone (PAT) signal and actigraphy which are recorded with an ambulatory wrist-worn device (Watch-PAT100). The PAT signal is a measure of the pulsatile volume changes at the finger tip reflecting sympathetic tone variations. The algorithm was developed using a training set of 30 patients recorded simultaneously with polysomnography and Watch-PAT100. Sleep records were divided into 5 min intervals and two time series were constructed from the PAT amplitudes and PAT-derived inter-pulse periods in each interval. A prediction function based on 16 features extracted from the above time series that determines the likelihood of detecting a REM epoch was developed. The coefficients of the prediction function were determined using a genetic algorithm (GA) optimizing process tuned to maximize a price function depending on the sensitivity, specificity and agreement of the algorithm in comparison with the gold standard of polysomnographic manual scoring. Based on a separate validation set of 30 patients overall sensitivity, specificity and agreement of the automatic algorithm to identify standard 30 s epochs of REM sleep were 78%, 92%, 89%, respectively. Deploying this REM detection algorithm in a wrist worn device could be very useful for unattended ambulatory sleep monitoring. The innovative method of optimization using a genetic algorithm has been proven to yield robust results in the validation set.