Decoding EEG signals of different mental states is a challenging task for brain-computer interfaces (BCIs) due to nonstationarity of perceptual decision processes. This paper presents a novel boosted convolutional neural networks (ConvNets) decoding scheme for motor imagery (MI) EEG signals assisted by the multiwavelet-based time-frequency (TF) causality analysis. Specifically, multiwavelet basis functions are first combined with Geweke spectral measure to obtain high-resolution TF-conditional Granger causality (CGC) representations, where a regularized orthogonal forward regression (ROFR) algorithm is adopted to detect a parsimonious model with good generalization performance. The causality images for network input preserving time, frequency and location information of connectivity are then designed based on the TF-CGC distributions of alpha band multichannel EEG signals. Further constructed boosted ConvNets by using spatio-temporal convolutions as well as advances in deep learning including cropping and boosting methods, to extract discriminative causality features and classify MI tasks. Our proposed approach outperforms the competition winner algorithm with 12.15% increase in average accuracy and 74.02% decrease in associated inter subject standard deviation for the same binary classification on BCI competition-IV dataset-IIa. Experiment results indicate that the boosted ConvNets with causality images works well in decoding MI-EEG signals and provides a promising framework for developing MI-BCI systems.
) is a state-of-the-art technology which establishes a direct communication pathway between human brain and external devices by translating neuronal activities into a series of output commands to accomplish user's intentions [1], and thereby has a wide range of applications from clinic to industry for both patients and normal people [2], such as controlling wheelchair or prosthesis to improve the disabled life quality [3], affecting neural plasticity to facilitate stroke rehabilitation [4], and handling computer games for Manuscript received September 24, 2018. entertainment of healthy users [5]. Despite the impressive advancements in recent years, EEG-BCI technology is still not able to decode complicated human mental state because of the high complexity of cognitive processing procedure in brain and low signal-to-noise ratio in EEG signals. Hence it is necessary to develop an effective EEG decoding scheme for enhancing usability and interpretability of BCI systems.
Analyzing the EEG signals induced by motor imagery (MI) is one of the most popular but challenging paradigm in BCIs [6].
The key step for MI-BCI implementations is to use machine learning techniques to extract information from EEG recordings of brain activities [7]. Among various types of feature representations for MI-EEG decoding, connectivity patterns of multi-channel signals could generate more discriminating features compared with static single-channel derived features [8] such as the well-known common spatial patterns (CSP) [9,10], since the dynamic and oscillatory interactions among different regions in the sensorimotor cortex of brain play a fundamental role in accomplishing movement imaginations [11,12]. Over the latest few years, several approaches have been proposed to analyze connectivity-based MI-BCI systems [13]. For example, Billinger et al. [14] suggested a method to extracting single-trial directed transfer functions (DTF) from vector autoregressive (VAR) models of independent components for MI-BCI classification, where the classification results were similar to band power (BP) features. In the work of Rathee et al. [15], time-domain partial Granger causality (PGC) is used as the connectivity features in a MI-BCI setting, and it turned out that single-trial effective connectivity distribution can enhance discriminability of mental imagery tasks. In general, the connectivity measures mentioned above can produce useful discriminant features for the classification of brain responses evoked during certain tasks. However, these methods, which assume the stationarity of EEG signals, cannot disclose important dynamic temporal information of connectivity, thus fail to provide robust distinction for nonstationary and complex MI-EEGs, and further result in dissatisfactory classification results with commonly used classification algorithms such as support vector machine (SVM).
Compared to conventional DTF and PGC methods, the timevarying Granger causality (GC) analysis [16], which has proven to be effective for detecting dynamic directed interaction patterns from nonstationary EEG signals, provides a new approach to connectivity feature representation. Currently, the most commonly used approaches for dynamic GC analysis can be broadly categorized into three classes: sliding window method [17], adaptive multivariate estimation [18], and parametric modelling approach [19]. In the sliding window approach, the detection performance can be significantly affected by the choice of window size [20]. Most adaptive methods set fixed model structures and estimate model parameters based on recursive least squares (RLS) or Kalman filtering [21]; they cannot track rapid varying causalities because of the slow convergence speed. In contrast, the parametric approach employing basis function expansion scheme can provide better dynamic causal features with high temporal resolution [22]. In such a detection approach, the underlying time-varying models of signals are represented using multiwavelet basis functions with good approximation properties [23,24], and an effective model structure decision algorithm like regularized orthogonal forward regression (ROFR) [25] is applied to reduce and refine the initial model; then both rapid and slow varying causalities between nonstationary signals can be successfully detected [22]. However, the pairwise time-domain GC approach proposed in [22] ignores frequency information and indirect effects caused by mutual sources, which are crucial in MI recognition due to the essential identity of different tasks is the specific regulation of a multi-channel EEG pattern in determined frequency ranges.
The classification method is an another vital part of a connectivity-based MI-BCI system; however, the advantages and potentials of the classifying algorithms for EEG classification have not been fully explored [26]. Recently, a prominent advance in machine learning is the application of deep learning with convolutional neura
This content is AI-processed based on open access ArXiv data.
