Improved A-phase detection of cyclic alternating pattern using deep learning

Abstract

In recent years, machine learning algorithms have become increasingly popular for analyzing biomedical signals. This includes the detection of cyclic alternating pattern (CAP) in electroencephalography recordings. Here, we investigate the performance gain of a recurrent neural network (RNN) for CAP scoring in comparison to standard classification methods. We analyzed 15 recordings (n1-n15) from the publicly available CAP Sleep Database on Physionet to evaluate each machine learning method. A long short-term memory (LSTM) network increases the accuracy and F1-score by 0.5-3.5% and 3.5-8%, respectively, compared to commonly used classification algorithms such as linear discriminant analysis, k-nearest neighbour or feed-forward neural network. Our results show that by using a LSTM classifier the quantity of correctly detected CAP events can be increased and the number of wrongly classified periods reduced. RNNs significantly improve the precision in CAP scoring by taking advantage of available information from the past for deciding current classification.