A knowledge discovery methodology from EEG data for cyclic alternating pattern detection

Abstract

BackgroundDetection and quantification of cyclic alternating patterns (CAP) components has the potential to serve as a disease bio-marker. Few methods exist to discriminate all the different CAP components, they do not present appropriate sensitivities, and often they are evaluated based on accuracy (AC) that is not an appropriate measure for imbalanced datasets.MethodsWe describe a knowledge discovery methodology in data (KDD) aiming the development of automatic CAP scoring approaches. Automatic CAP scoring was faced from two perspectives: the binary distinction between A-phases and B-phases, and also for multi-class classification of the different CAP components. The most important KDD stages are: extraction of 55 features, feature ranking/transformation, and classification. Classification is performed by (i) support vector machine (SVM), (ii) k-nearest neighbors (k-NN), and (iii) discriminant analysis. We report the weighted accuracy (WAC) that accounts for class imbalance.ResultsThe study includes 30 subjects from the CAP Sleep Database of Physionet. The best alternative for the discrimination of the different A-phase subtypes involved feature ranking by the minimum redundancy maximum relevance algorithm (mRMR) and classification by SVM, with a WAC of 51%. Concerning the binary discrimination between A-phases and B-phases, k-NN with mRMR ranking achieved the best WAC of 80%.ConclusionsWe describe a KDD that, to the best of our knowledge, was for the first time applied to CAP scoring. In particular, the fully discrimination of the three different A-phases subtypes is a new perspective, since past works tried multi-class approaches but based on grouping of different sub-types. We also considered the weighted accuracy, in addition to simple accuracy, resulting in a more trustworthy performance assessment. Globally, better subtype sensitivities than other published approaches were achieved.

A-phase classification using convolutional neural networks

“…67% and 71% [36,37,38]. On the other hand, it is clear that the benefits of re-training with more than 20% of the data depend on each subject, particularly in the cases of subjects for whom the number of A2-and A3-phases is considerably less or more than half the number of A1-phases, as with Subjects 2, 3 and 4 in this study.…”

“…A similar work by Machado et al uses a set of 55 features to classify the A-phases between two groups: A1 and A2/A3, using different types of classifiers, including quadratic discriminant, k-Nearest Neighbors and Support Vector Machines (SVM). The best result is obtained with an SVM, achieving an accuracy of 71% [36,37] for the classification of A-phases subtypes and 76% for discriminating between A-phases and B-phases. This suggests that distinguishing among the sub-types of A-phases could be a harder problem than distinguishing between A-phases and B-phases.…”

Matrix of Lags: A tool for analysis of multiple dependent time series applied for CAP scoring

“…• EEG band descriptors [25][26][27][28][29][30][31][32][33][34][35] : describes how much the amplitude of the activity, in the selected frequency band, differs from its background; • Differential variance [26,29,[32][33][34][35] : alteration of the variance between the current and the previous epoch; • Detrended fluctuation analysis [36] : characterizes the correlation structure of non-stationary time series; • Hjorth descriptors [26,29,[32][33][34][35] : the activity, mobility, and complexity parameters that are respectively estimated by the variance of the signal, the variance of the slopes that were normalized by the variance of the amplitude distribution and the ratio of the mobility from the first derivative of the signal to the mobility of the signal; • Power spectral density of the band [25,28,[37][38][39][40] : distribution of power into frequency components that compose the signal; • Moving average ratio [41] : activity index determined by the ratio of a short moving average to a long moving average; • Teager energy operator [25][26][27][28]38,40] : nonlinear metric that can be interpreted as an instantaneous measure of energy; • Lempel-Ziv Complexity [25,28,37] : ...…”

“…• Sample entropy [28,37] : metric to assess the complexity of time-series. • Tsallis Entropy [37] : entropy metric that depends on a dimensionless parameter; • Fractal dimension [25,28,37,40] : statistical index for the complexity of the patterns; • Similarity Index [42] : symbolic based method that examines the statistical behavior of local extrema; • Zero-crossing rate [25,28] : measure to analyze the dominant frequency of a signal by counting the number times the baseline was crossed in a fixed time interval; • Phase-amplitude coupling [43] : describes the amplitude of high-frequency activities that are modulated by the phase of low-frequency oscillations.…”

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

“…It was verified that, in general, TEO was the best feature. [36], [35] Empirical mode decomposition Decompose the signal into intrinsic mode functions, and each function denotes an embedded characteristic oscillation on a disjointed time scale [36], [35] Fractal dimension Counts the number of occurrences of a sequence [36], [35], [11], [32], [9] Hjorth activity…”