A Portable Wireless Device for Cyclic Alternating Pattern Estimation from an EEG Monopolar Derivation

Abstract

Quality of sleep can be assessed by analyzing the cyclic alternating pattern, a long-lasting periodic activity that is composed of two alternate electroencephalogram patterns, which is considered to be a marker of sleep instability. Experts usually score this pattern through a visual examination of each one-second epoch of an electroencephalogram signal, a repetitive and time-consuming task that is prone to errors. To address these issues, a home monitoring device was developed for automatic scoring of the cyclic alternating pattern by analyzing the signal from one electroencephalogram derivation. Three classifiers, specifically, two recurrent networks (long short-term memory and gated recurrent unit) and one one-dimension convolutional neural network, were developed and tested to determine which was more suitable for the cyclic alternating pattern phase’s classification. It was verified that the network based on the long short-term memory attained the best results with an average accuracy, sensitivity, specificity and area under the receiver operating characteristic curve of, respectively, 76%, 75%, 77% and 0.752. The classified epochs were then fed to a finite state machine to determine the cyclic alternating pattern cycles and the performance metrics were 76%, 71%, 84% and 0.778, respectively. The performance achieved is in the higher bound of the experts’ expected agreement range and considerably higher than the inter-scorer agreement of multiple experts, implying the usability of the device developed for clinical analysis.

Multiple Time Series Fusion Based on LSTM: An Application to CAP A Phase Classification Using EEG

“…Considering this metric, it is possible to conclude that the best state-of-the-art results, which have included sleep disorder patients in the analysis, are in line with the results attained in this work (76%). However, Mendonça et al [ 31 , 63 ] examined patients with sleep-disordered breathing while subjects with NFLE were examined in this work. Sharma et al [ 60 ] also evaluated subjects with NFLE but attained a lower Acc, highlighting how difficult it is to examine subjects with this disorder.…”

“…CAP phases have a strong temporal dependency that can be captured by recurrent neural networks, e.g., LSTM [ 31 ], and the activity can be measured in different EEG channels. Therefore, a novel approach was followed in this work where the information from multiple EEG channels was fused by a proposed deep learning channel fusion methodology, composed of LSTM, concatenation, and fully connected (dense) layers.…”

“…The AUC of the two TFCV cycles was averaged to find the mean AUC considered as the PM for the model under examination. The Adam algorithm [ 48 ] was used for training since it was found to be the most suited for the CAP analysis based on LSTM [ 31 ]. Cost-sensitive learning was employed to deal with the substantial data imbalance (instead of using a balancing operation that can alter the expected distribution of the data) since, for some subjects, more than 80% of the epochs can refer to the “not-A” class.…”

On the use of patterns obtained from LSTM and feature-based methods for time series analysis: application in automatic classification of the CAP A phase subtypes

“…Several classifiers were previously proposed to perform the CAP phase analysis [22][23][24][25], and it was verified that artificial neural networks attained the best performance. Among these, the Long Short-Term Memory (LSTM), a type of the recurrent neural network, was identified as possibly the most suited to perform the classification [25,26] since the CAP phases have a strong temporal dependency [7]. Such temporal correlations frequently occurs in the physiological signals [27] and can be identified by the LSTM, capable of finding both short and long-term correlations in the time series [28].…”

“…Several classifiers were previously proposed to perform the CAP phase analysis [22] [23] [24] [25], and it was verified that artificial neural networks attained the best performance. Among these, the Long Short-Term Memory (LSTM), a type of the recurrent neural network, was identified as possibly the most suited to perform the classification [25] [26] since the CAP phases have a strong temporal dependency [7].…”