Inter-database validation of a deep learning approach for automatic sleep scoring

In this work we describe a new deep learning approach for automatic sleep staging, and carry out its validation by addressing its generalization capabilities on a wide range of sleep staging databases. Prediction capabilities are evaluated in the context of independent local and external generalization scenarios. Effectively, by comparing both procedures it is possible to better extrapolate the expected performance of the method on the general reference task of sleep staging, regardless of data from a specific database. In addition, we examine the suitability of a novel approach based on the use of an ensemble of individual local models and evaluate its impact on the resulting inter-database generalization performance. Validation results show good general performance, as compared to the expected levels of human expert agreement, as well as state-of-the-art automatic sleep staging approaches. Keywords Sleep Staging, Deep Learning, Inter-database Generalization, Classification ensemble 1. INTRODUCTION Sleep staging is one of the most important tasks during the clinical examination of polysomnographic sleep recordings (PSGs). A PSG records the relevant biomedical signals of a patient in the context of Sleep Medicine studies, representing the basic tool for the diagnosis of many sleep disorders. Sleep staging characterizes the patient's sleep macrostructure leading to the so-called hypnogram. The hypnogram plays also a fundamental role for the interpretation of several other biosignal activities of interest, such as the evaluation of the respiratory function, or the identification of different body and limb movement [1] [2]. Each epoch can be classified into five possible states (wakefulness, stages N1, N2, N3, and R) according to the observed signal pattern activity in the reference PSG interval. Specifically, for sleep staging, neurophysiological activity of interest involves monitoring of different traces of electroencephalographic (EEG), electromyographic (EMG) and electrooculographic (EOG) activity [1]. A typical PSG examination comprises 8 up to 24 hours of continuous signal recording, and its analysis is usually carried out manually by an expert clinician. The scoring process is consequently expensive and highly demanding, due to the involved clinician's time, and the complexity of the analysis itself. Moreover, the demand for PSG investigations is growing in relation with the general public awareness, motivated by clinical findings over the last years uncovering the negative impact that sleep disorders exert over health. This represents a challenge for the already congested sleep centers, with steadily increasing waiting lists.