RESUMEN
Mental tasks classification is increasingly recognized as a major challenge in the field of EEG signal processing and analysis. State-of-the-art approaches face the issue of spatially unstable structure of highly noised EEG signals. To address this problem, this paper presents a multi-channel convolutional neural network architecture with adaptively optimized parameters. Our solution outperforms alternative methods in terms of classification accuracy of mental tasks (imagination of hand movements and speech sounds generation) while providing high generalization capability (â¼5%). Classification efficiency was obtained by using a frequency-domain multi-channel neural network feeding scheme by EEG signal frequency sub-bands analysis and architecture supporting feature mapping with two subsequent convolutional layers terminated with a fully connected layer. For dataset V from BCI Competition III, the method achieved an average classification accuracy level of nearly 70%, outperforming alternative methods. The solution presented applies a frequency domain for input data processed by a multi-channel architecture that isolates frequency sub-bands in time windows, which enables multi-class signal classification that is highly generalizable and more accurate (â¼1.2%) than the existing solutions. Such an approach, combined with an appropriate learning strategy and parameters optimization, adapted to signal characteristics, outperforms reference single- or multi-channel networks, such as AlexNet, VGG-16 and Cecotti's multi-channel NN. With the classification accuracy improvement of 1.2%, our solution is a clear advance as compared to the top three state-of-the-art methods, which achieved the result of no more than 0.3%.
Asunto(s)
Electroencefalografía/métodos , Redes Neurales de la Computación , Procesamiento de Señales Asistido por Computador , Algoritmos , Interfaces Cerebro-Computador , Humanos , Curva ROCRESUMEN
Graph theoretic approaches in analyzing spatiotemporal dynamics of brain activities are under-studied but could be very promising directions in developing effective brain-computer interfaces (BCIs). Many existing BCI systems use electroencephalogram (EEG) signals to record and decode human neural activities noninvasively. Often, however, the features extracted from the EEG signals ignore the topological information hidden in the EEG temporal dynamics. Moreover, existing graph theoretic approaches are mostly used to reveal the topological patterns of brain functional networks based on synchronization between signals from distinctive spatial regions, instead of interdependence between states at different timestamps. In this study, we present a robust fold-wise hyperparameter optimization framework utilizing a series of conventional graph-based measurements combined with spectral graph features and investigate its discriminative performance on classification of a designed mental task in 6 participants with amyotrophic lateral sclerosis (ALS). Across all of our participants, we reached an average accuracy of 71.1%±4.5% for mental task classification by combining the global graph-based measurements and the spectral graph features, higher than the conventional non-graph based feature performance (67.1%±7.5%). Compared to using either one of the graphic features (66.3%±6.5% for the eigenvalues and 65.9%±5.2% for the global graph features), our feature combination strategy shows considerable improvement in both accuracy and robustness performance. Our results indicate the feasibility and advantage of the presented fold-wise optimization framework utilizing graph-based features in BCI systems targeted at end-users.
Asunto(s)
Interfaces Cerebro-Computador , Humanos , Encéfalo , Electroencefalografía/métodos , Algoritmos , ImaginaciónRESUMEN
[This corrects the article DOI: 10.3389/fnins.2019.00822.].
RESUMEN
In recent years, electroencephalography (EEG) measured around the ears, called ear-EEG, has been introduced to develop unobtrusive and ambulatory EEG-based applications. When measuring ear-EEGs, the availability of a reference site is restricted due to the miniaturized device structure, and therefore a reference electrode is generally placed near the recording electrodes. As the electrical brain activity recorded at a reference electrode closely placed to recording electrodes may significantly cancel or influence the brain activity recorded by the recording electrodes, an appropriate re-referencing method is often required to mitigate the impact of the reference brain activity. In this study, therefore, we systematically investigated the impact of different re-referencing methods on ear-EEGs spontaneously generated from endogenous paradigms. To this end, we used two ear-EEG datasets recorded behind both ears while subjects performed an alpha modulation task [eyes-closed (EC) and eyes-open (EO)] and two mental tasks [mental arithmetic (MA) and mental singing (MS)]. The measured ear-EEGs were independently re-referenced using five different methods: (i) all-mean, (ii) contralateral-mean, (iii) ipsilateral-mean, (iv) contralateral-bipolar, and (v) ipsilateral-bipolar. We investigated the changes in alpha power during EO and EC tasks, as well as event-related (de) synchronization (ERD/ERS) during MA and MS. To evaluate the effects of re-referencing methods on ear-EEGs, we estimated the signal-to-noise ratios (SNRs) of the two ear-EEG datasets, and assessed the classification performance of the two mental tasks (MA vs. MS). Overall patterns of changes in alpha power and ERD/ERS were similar among the five re-referencing methods, but the contralateral-mean method showed statistically higher SNRs than did the other methods for both ear-EEG datasets, except in the contralateral-bipolar method for the two mental tasks. In concordance with the SNR results, classification performance was also statistically higher for the contralateral-mean method than it was for the other re-referencing methods. The results suggest that employing contralateral mean information can be an efficient way to re-reference spontaneously generated ear-EEGs, thereby maximizing the reliability of ear-EEG-based applications in endogenous paradigms.