Towards a user-friendly sleep staging system for polysomnography part II: Patient-dependent features extraction using the SATUD system

Abstract Manual sleep stages scoring is time-consuming, complex and requires specific medical knowledge. Automatic sleep stages classification, usually based on supervised methods of machine learning, is the object of researchers interest. However, it remains challenging because of the high variability among patients which is not considered with such algorithms. This paper presents a method to extract patient-dependent qualitative features from electrophysiological signals, preceding a supervised machine learning classifier. Instead of using fixed thresholds, the developed method called ”Self-Adaptative Thresholding Using Descriptors” (SATUD), proposes an unsupervised self-adjusting thresholding. Thresholds are automatically adjusted to maximize both the similarity within a same sleep stage and the dissimilarity between different ones. This method is evaluated using manual sleep stages scoring from 60 patients with various pathologies to ensure high variability. The SATUD shows a better adaptation to the patient specificities, compared with two other thresholding methods implemented in this study. Indeed, the number of 30-seconds recording segments respecting all their sleep stage properties increased by more than 80 % with the use of the SATUD, compared to other thresholding techniques. It was also proved robust to noise and sweat artifacts. The SATUD thereby provides patient-dependent qualitative features which can be used for automatic sleep stages scoring using a machine learning method. This last point was presented in the companion paper.