Brain Machine Interface (BMI) systems have attracted extensive attention in the past decade, because of their potential in improving human life, especially for those who are affected by motor disabilities. Since gait deficits are commonly associated with spinal cord injuries (SCI), limb loss, and neurodegenerative diseases, there is a need to investigate innovative therapies to restore gait in such patients. Exoskeletons have become prominent tools for the rehabilitation of SCI and stroke patients (Sale et al., 2012; Venkatakrishnan et al., 2014). BMIs have been deployed to infer the user's intent from his/her brain activity to generate output signals to control powered exoskeletons for upper and lower limb rehabilitation (Noda et al., 2012; Contreras-Vidal and Grossman, 2013; Kilicarslan et al., 2013; French, 2014; Venkatakrishnan et al., 2014). In Presacco et al. (2011), Presacco et al. showed decoding of gait kinematics during treadmill walking from EEG of able-bodied subjects with accuracies comparable to that from a similar study in non-human primates with electrodes implanted in their brains (Fitzsimmons et al., 2009). Further, in Kilicarslan et al. (2013), a paraplegic subject's motion intentions were accurately decoded using Locality Fisher Discriminant Analysis and a Gaussian Mixture Model (LFDA-GMM) for two different gait tasks (i.e., repeated walking-turning right-turning left motions and sit-rest-stand motions). The model enabled the closed-loop EEG-based BMI system to control a robotic exoskeleton (NeuroREX) in real-time, resulting in independent walking for the paraplegic user.

To control a device via BMI, different brain activity patterns produced by a user need to be accurately identified by a neural interface system and translated into appropriate commands (Contreras-Vidal et al., 2015). Discrete decoding (neural classification) of intent from EEG signals can be considered as a pattern recognition problem, and advanced machine learning techniques are needed to accurately translate the brain electrical activities to meaningful control commands. Many machine learning methods [e.g., linear discriminant analysis (LDA), support vector machine (SVM), Bayesian classifiers] have been applied for classifying EEG signals in different BMI applications (Kilicarslan et al., 2013; Niazi et al., 2013; Leamy et al., 2014; Lew et al., 2014; Hortal et al., 2015; Jiang et al., 2015). However, most of them serve as a “black box” in that we do not know how the brain activity changes during long-term BMI use nor how the brain regions contribute to the classification process while people perform different tasks. The human brain consists of over 100 billion cells, typically divided into regions by neuroanatomists. Different regions have their specific functionalities while coordinating together to accomplish everyday tasks. Moreover, the specific contributions of brain regions to classification may change due to learning a BMI. Therefore, it is important to identify and track these changes to increase our understanding of brain function, BMI learning and performance. In that context, the hypothesis of this research is that different brain regions contribute differentially to BMI learning and control of robot assisted lower-limb movements—we are interested in learning the importance of these regions for neural classification of gait states.

Kernel learning methods have been effectively applied for many machine learning problems, including feature selection, data regression and classification for EEG signals (Garrett et al., 2003; Lal et al., 2004; Lotte et al., 2007). SVM is one of the most popular kernel methods for pattern recognition. However, a problem with using the standard SVM in BMI applications is that it provides no insight about the importance of distinct features, and thus has little knowledge about the biophysical properties of relevant features used in decoding/classification. Multiple kernel learning (MKL), which makes use of a combination of basis kernels to represent different types of features or data, have been shown to outperform traditional single-kernel machines in different aspects (Sonnenburg et al., 2006; Tian et al., 2012; Samek et al., 2013; Li et al., 2014). The main advantage of using MKL over SVM is that MKL can simultaneously learn the classifier and the optimal weights for basis kernels. In this paper, we investigate and make use of this property to simultaneously decode gait states from multi-channel EEG signals and learn the relative importance of different scalp brain areas. Particularly, we build a composite kernel based on a linear combination of basis kernels, in which each basis kernel can be represented by a group of electrodes corresponding to selected regions of interest (ROIs), and consequently contribute unique biophysical information.

The primary goal of this research is to show the feasibility of simultaneously classifying the pattern of user's internal gait states (e.g., stop, walk, turn left, turn right) from the EEG signals and learning the relative importance of different scalp brain areas. Previous studies have shown that low delta band (0.1–2 Hz) EEG contains intended movement-related information for decoding the kinematics of lower limb or gait states (Presacco et al., 2011, 2012; Jorquera et al., 2013; Kilicarslan et al., 2013; Bulea et al., 2014; Luu et al., 2016). For example, in Presacco et al. (2011), Presacco et al. (2012), and Luu et al. (2016), it was shown that delta band EEG contains information about gait movement kinematics that can be decoded using Wiener or Kalman filters. In Kilicarslan et al. (2013), Jorquera et al. (2013), and Bulea et al. (2014), it was shown that movement-type (e.g., “stop,” “go,” etc.) classifiers can be designed based on delta band EEG signals. Another study (Velu and de Sa, 2013) showed that features corresponding to frequencies less than 2 Hz were the most heavily weighted during single trial classification of walking and pointing direction. Inspired by the above findings, in this study, we utilize delta band (0.1–2 Hz) EEG to build our basis kernels (feature matrices) for neural classification of gait states.

The other goal of the research is to compare the brain regions employed for classifying movement intents from able-bodied subjects and individuals with spinal cord injury (SCI) given differences in neural activity across these populations. Studies have shown that SCI can cause widespread and sustained brain inflammation that leads to progressive loss of brain cells in key brain regions with associated cognitive problems (Wu et al., 2014a,b). Cramer et al. have found that in patients with complete SCI, many features of normal motor system function are preserved, however, the volume and patterns of activation and the modulation of function with change in task are abnormal and absent, respectively, in patients with SCI (Cramer et al., 2005). In this preliminary study, we collected EEG data from a SCI volunteer over multiple sessions to compare the classification results with an able-bodied subject on the important brain regions during learning.

The remainder of the paper is organized as follows. Section 2 introduces our methodology for region importance learning, including experimental protocol, data acquisition, processing and analysis. In particular, we introduce the MKL algorithm and how we apply it to learn the importance of different brain regions. In section 3, we validate the efficacy of the proposed work via experiments using four-class single session and two-class longitudinal EEG data. Section 4 presents our discussions from analyzing the experimental results. Finally, concluding remarks are provided in Section 5.

Previous study (Kilicarslan et al., 2013) has shown the feasibility of classifying user movement intentions using EEG signals in the delta band (0.1–2 Hz) based on a GMM classifier and achieved a high offline evaluation accuracies for the 3-class tasks. In this study, we extend the study of intended motions from three classes to four classes and set the research as a longitudinal study. The experimental results demonstrate that by properly weighting the importance of the features, MKL can be used as an efficient decoder to predict user's movement intentions. For the tested subjects, the overall accuracies reached above 90% for the two-class classification task and above 65% for the more complicated four-class classification task. Compared to some commonly used machine learning approaches (i.e., Bayesian classifiers, LDA, SVM) in BMI, MKL has the following advantages (1) Unlike the LDA and Bayesian classifiers, MKL does not need to make assumptions on the data distribution. MKL is a member of kernel learning methods, which utilizes a linear combination of kernels and transforms the original data into an appropriate (kernel) feature space. Thus, all beneficial properties (e.g., optimality) of linear classifiers are maintained, while MKL is also efficient when the data are non-linear in the input space. (2) MKL is a robust learning method in the high dimensional space (Bach, 2008). In Kilicarslan et al. (2013), a sliding window was used on all 64 channels to extract features, it resulted in a 1,280 dimensional space, and a dimensionality reduction technique was required to reduce the dimension of the data decoding by GMM. Similar to the feature extraction step in Kilicarslan et al. (2013), we applied a 400 ms sliding window to extract the EEG delta band amplitude as input to the classifier. Differently, we divided the channels into different groups and extract features from each group. Thus, the resulting dimension for each group of data is at most 240, which is much lower than 1,280, and a dimensionality reduction method is not necessary for classification by MKL. (3) MKL can also be used to infer the importance of different groups of features, which is not feasible in other machine learning methods. The weight for each group is initialized uniformly at the beginning and optimized during the gradient descent in the MKL algorithm. MKL ranks sets of features corresponding to the meaningful features for solving the classification problem, and the results indicate importance in the representation of movement for different scalp brain areas.

Comparing the results from the SCI subject and the able-bodied subject, we observe the most important brain region changing from the midline fronto-central to the right fronto-central in both tasks. This could be due to the loss of brain cells and degraded cerebral cortex dynamics or lack of afferent input after spinal cord injury. For example, changes in movement-related cortical potentials have been noted after SCI and correlated with the severity of the injury (Boord et al., 2008; Gourab and Schmit, 2010). Moreover, altered spontaneous neuronal activity following SCI has been characterized by a shift in the dominant spectral power peak toward lower frequencies, including in primary and secondary somatosensory cortices (Tran et al., 2004; Sarnthein et al., 2006). Noted that the number of subjects participated in the study was limited, there might be individual variations on the classification results. However, as the classification results were based on standard cross-validation procedure, we believe the proposed approach and model can be generalized to other subjects.

In the longitudinal experiments, we found that the subjects were adapting to the BMI gait task in the first several sessions, so that the brain regions used for neural classification were not stable, which was reflected in the moderate classification performance. After several sessions of training, as the subjects learned to control the exoskeleton for movement, we observed the channels employed for movement classification converged to specific regions—the midline fronto-central areas for the able-bodied subject and the right fronto-central/left centro-parietal areas for the SCI user. In addition, the classification accuracy generally increased along session, and interestingly the weights for the important regions also increased. This demonstrates the cortical plasticity triggered by the BMI use, as the user gradually learns to control the exoskeleton for movement.

In this paper, we presented the feasibility of simultaneously classifying the pattern of user's internal gait states from the EEG signals and learning the relative importance of different scalp brain areas based on the MKL algorithm. The MKL has the advantages of learning the classifier and the optimal kernel weights simultaneously. We investigated these properties and applied the MKL classifier to infer the relative importance of different groups of features (different sources of information) in a BMI application to classify one's motion intention from the EEG signals. The experimental results demonstrated that the frontal/fronto-central regions were the most important regions for classifying gait states of the tested subjects, which is consistent with the brain regions hypothesized to be involved in the control of lower-limb movements. By comparing the results from the SCI subject and the able-bodied subject, the important regions were observed to change, which could be due to the loss of brain cells and degraded cerebral cortex dynamics or lack of afferent input after spinal cord injury. In the longitudinal experiment, while the user learned to control the exoskeleton for movement over multiple sessions, the classification accuracy increased and the weights for important regions stabilized. These findings suggest the cortical plasticity triggered by the BMI use, which will be investigated further in the future study.

YZ developed the learning method, processed and analyzed the data, and wrote the manuscript. SP supervised development of work, helped in data interpretation and manuscript edit. AK helped to acquire and interpret the data. JC supervised development of work, helped in data interpretation and manuscript edit and evaluation.

This work was supported in part by the NIH Award R01 NS075889, Mission Connect—A TIRR Foundation.