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

Abstract

The cyclic alternating pattern is a marker of sleep instability identified in the electroencephalogram signals whose sequence of transient variations compose the A phases. These phases are divided into three subtypes (A1, A2, and A3) according to the presented patterns. The traditional approach of manually scoring the cyclic alternating pattern events for the full night is unpractical, with a high probability of miss classification, due to the large quantity of information that is produced during a full night recording. To address this concern, automatic methodologies were proposed using a long short-term memory to perform the classification of one electroencephalogram monopolar derivation signal. The proposed model is composed of three classifiers, one for each subtype, performing binary classification in a one versus all procedure. Two methodologies were tested: feed the pre-processed electroencephalogram signal to the classifiers; create features from the pre-processed electroencephalogram signal which were fed to the classifiers (feature-based methods). It was verified that the A1 subtype classification performance was similar for both methods and the A2 subtype classification was higher for the feature-based methods. However, the A3 subtype classification was found to be the most challenging to be performed, and for this classification, the feature-based methods were superior. A characterization analysis was also performed using a recurrence quantification analysis to further examine the subtypes characteristics. The average accuracy and area under the receiver operating characteristic curve for the A1, A2, and A3 subtypes of the feature-based methods were respectively: 82% and 0.92; 80% and 0.88; 85% and 0.86.

Heuristic Optimization of Deep and Shallow Classifiers: An Application for Electroencephalogram Cyclic Alternating Pattern Detection

“…The symbolic dynamics transformed the input signal into a sequence of symbols by examining several thresholds for the signal's amplitude, which are multiples of the signal's standard deviation, σ. A total of nine thresholds were used since it was previously identified as a suitable number for the A phase examination [18]. Thus, for each sample point of the epoch (which is composed of 100 sample points), the algorithm evaluated if the point's amplitude is lower than either −5 × σ, −4 × σ, −3 × σ, −2 × σ, −σ, 2 × σ, 3 × σ, 4 × σ, emitting the symbol 1, 2, 3, 4, 5, 6, 7, 8, 9, respectively.…”

“…It was reported by Mendonça et al [18] that the deep learning models have difficulties recognizing the relevant patterns for two of the three subtypes, which compose the A phases, suggesting the need for examining feature-based methods in this work. Specifically, the LSTM was examined since it was identified as a suitable classifier for feature-based analysis with temporal dependencies [19].…”

A Review of Methods for Sleep Arousal Detection Using Polysomnographic Signals

“…Automatic methodologies were proposed by using an LSTM model to perform the classification of one EEG channel signal [108]. The model was composed of three classifiers, one for each subtype, performing binary classification in a one versus all procedure.…”

Nonlinear modeling of hydroturbine dynamic characteristics using LSTM neural network with feedback

“…Since the output of LSTMNN has a strong correlation with the historical input of the model, it is suitable for fitting prediction. LSTMNN is widely used for energy futures prices prediction, 10 water demand of pump prediction, 11 air quality prediction, 12 prediction classification, 8,13 and other aspects. The results show that the overall network prediction accuracy is high when the input data fluctuate gently, but lower when the input data fluctuate greatly.…”