Combination of deep and shallow networks for cyclic alternating patterns detection

Abstract

The cyclic alternating pattern can be seen as an electroencephalogram marker of sleep instability. This pattern consists of alternations between activation and quiescent phases. An automatic cyclic alternating pattern detection method is proposed, having the advantage, over other previously proposed methods, of being featureless. Therefore, there is no need to handcraft features and employ a feature selection procedure. A Deep Auto Encoder is used for automatic feature extraction and classification of the activation phases. A shallow Artificial Neural Network is then employed for cyclic alternating pattern classification using the output of the Deep Auto Encoder. These two-cascaded networks are connected by a memory buffer. Both networks are optimized using a heuristic approach and Kolmogorov's Mapping theorem. A public database with 14 subjects is used to test the methods. For the activation phase classification, a 2 seconds raw EEG is used as an input of the Deep Auto Encoder. For the cyclic alternating pattern classifier, the whole memory buffer is used as input. The accuracy of activation phase detection is 67.2% and the accuracy of cyclic alternating pattern detection is 61.5%.