A Convolutional Network for the Classification of Sleep Stages

Abstract

Sleep stages classification is a crucial task in the context of sleep studies. It involves the simultaneous analysis of multiple signals recorded during sleep. However, it is complex and tedious, and even the trained expert can spend several hours scoring a single night recording. Multiple automatic methods have tried to solve these problems in the past, most of them by classifying a feature vector that is engineered for a specific dataset. In this work, we avoid this bias using a deep learning model that learns relevant features without human intervention. Particularly, we propose an ensemble of 5 convolutional networks that achieves a kappa index of 0.83 when classifying a dataset of 500 sleep recordings.

Addressing database variability in learning from medical data: An ensemble-based approach using convolutional neural networks and a case of study applied to automatic sleep scoring

“…As stated in the previous sections, the learning model was implemented using Convolutional Neural Networks (CNNs). The general used architecture was based on a previous model developed by the authors [18].…”

“…While all these operational blocks maintain the same kernel size, the number of filters in B(k), was set as being two times the number of filters in B(k-1). For B (1), the initial number of filters was fixed to 8 based on previous experiments [18]. The specific type and the size of the convolutional kernel are also left as free parameters in this work.…”

“…Instead it is the focus to analyze the generalization performance of a relatively good general architecture on the multidatabase prediction scenario. In this regard, exhaustive exploration of the hyper-parameter space was not a priority, and many parameterization decisions were based on previous experimentation described elsewhere [18].…”

“…For this purpose we studied the use of different kernel configurations. Based on previous experimentation [18] we took as reference baseline a 1D convolution of size 1x10 (0.1 s) and then explored variations increasing the kernel in the time dimension up to 1x100 (1 s), and the use of 2D convolutions (kernel size of 4x10). In principle the use of 1D convolution would avoid imposing a fixed spatial structure (a priori unknown) between the different input signals.…”

“…Countless attempts have been constantly appearing since then, and up to now (just to mention some, see [4] [5] [6] [7] [8] [9] [10] [11] [12] [13]) evidencing that the task still represents an open challenge, and source of active research activity. More recently, several publications have appeared based on the use of deep learning, claiming advantages over previous realizations, including improved performance, and the possibility to skip handcrafted feature engineering processes [14] [15] [16] [17] [18]. A common drawback, however, has to do with the limited validation procedures presented in these studies with regard to the aforementioned database variability problem.…”