Detection of microsleep states from the EEG: a comparison of feature reduction methods.

Microsleeps are brief lapses in consciousness with complete suspension of performance. They are the cause of fatal accidents in many transport sectors requiring sustained attention, especially driving. A microsleep-warning device, using wireless EEG electrodes, could be used to rouse a user from an imminent microsleep. High-dimensional datasets, especially in EEG-based classification, present challenges as there are often a large number of potentially useful features for detecting the phenomenon of interest. Thus, it is often important to reduce the dimension of the original data prior to training the classifier. In this study, linear dimensionality reduction methods-principal component analysis (PCA) and probabilistic PCA (PPCA)-were compared with eight non-linear dimensionality reduction methods (kernel PCA, classical multi-dimensional scaling, isometric mapping, nearest neighbour estimation, stochastic neighbourhood embedding, autoencoder, stochastic proximity embedding, and Laplacian eigenmaps) on previously collected behavioural and EEG data from eight healthy non-sleep-deprived volunteers performing a 1D-visuomotor tracking task for 1 h. The effectiveness of the feature reduction algorithms was evaluated by visual inspection of class separation on 3D scatterplots, by trustworthiness scores, and by microsleep detection performance on a stacked-generalisation-based linear discriminant analysis (LDA) system estimating the microsleep/responsive state at 1 Hz based on the reduced features. On trustworthiness, PPCA outperformed PCA, but PCA outperformed all of the non-linear techniques. The trustworthiness score for each feature reduction method also correlated strongly with microsleep-state detection performance, providing strong validation of the ability of trustworthiness to estimate the relative effectiveness of feature reduction approaches, in terms of predicting performance, and ability to do so independently of the gold standard. Graphical abstract Proposed microsleep detection system.

Reservoir computing approaches to microsleep detection.

Objective.The detection of microsleeps in a wide range of professionals working in high-risk occupations is very important to workplace safety. A microsleep classifier is presented that employs a reservoir computing (RC) methodology. Specifically, echo state networks (ESN) are used to enhance previous benchmark performances on microsleep detection.Approach.A clustered design using a novel ESN-based leaky integrator is presented. The effectiveness of this design lies with the simplicity of using a fine-grained architecture, containing up to 8 neurons per cluster, to capture individualized state dynamics and achieve optimal performance. This is the first study to have implemented and evaluated EEG-based microsleep detection using RC models for the detection of microsleeps from the EEG.Main results.Microsleep state detection was achieved using a cascaded ESN classifier with leaky-integrator neurons employing 60 principal components from 544 power spectral features. This resulted in a leave-one-subject-out average detection in performance ofφ = 0.51 ± 0.07 (mean ± SE), AUC- ROC = 0.88 ± 0.03, and AUC- PR = 0.44 ± 0.09.Significance.Although performance of EEG-based microsleep detection systems is still considered modest, this refined method achieved a new benchmark in microsleep detection.

Optimized echo state networks with leaky integrator neurons for EEG-based microsleep detection.

The performance of a microsleep detection system was calculated in terms of its ability to detect the behavioural microsleep state (1-s epochs) from spectral features derived from 16-channel EEG sampled at 256 Hz. Best performance from a single classifier model was achieved using leaky integrator neurons on an echo state network (ESN) classifier with a mean phi correlation (φ) of 0.38 and accuracy of 67.3%. A single classifier model of ESN with sigmoidal inputs achieved φ of 0.20 and accuracy of 48.5% and a single classifier model of linear discriminant analysis (LDA) achieved φ of 0.31 and accuracy of 53.6%. However, combining the output of several single classifier models (ensemble learning) via stacked generalization of the ESN with leaky integrator neurons approach led to a substantial increase in detection performance of φ of 0.51 and accuracy of 81.2%. This is a substantial improvement of our previous best result of φ = 0.39 on this data with LDA and stacked generalization.

EEG-based event detection using optimized echo state networks with leaky integrator neurons.

This study investigates the classification ability of linear and nonlinear classifiers on biological signals using the electroencephalogram (EEG) and examines the impact of architectural changes within the classifier in order to enhance the classification. Consequently, artificial events were used to validate a prototype EEG-based microsleep detection system based around an echo state network (ESN) and a linear discriminant analysis (LDA) classifier. The artificial events comprised infrequent 2-s long bursts of 15 Hz sinusoids superimposed on prerecorded 16-channel EEG data which provided a means of determining and optimizing the accuracy of overall classifier on `gold standard' events. The performance of this system was tested on different signal-to-noise amplitude ratios (SNRs) ranging from 16 down to 0.03. Results from several feature selection/reduction and pattern classification modules indicated that training the classifier using a leaky-integrator neuron ESN structure yielded highest classification accuracy. For datasets with a low SNR of 0.3, training the leaky-neuron ESN using only those features which directly correspond to the underlying event, resulted in a phi correlation of 0.92 compared to 0.37 that employed principal component analysis (PCA). On the same datasets, other classifiers such as LDA and simple ESNs using PCA performed weakly with a correlation of 0.05 and 0 respectively. These results suggest that ESNs with leaky neuron architectures have superior pattern recognition properties. This, in turn, may reflect their superior ability to exploit differences in state dynamics and, hence, provide superior temporal characteristics in learning.