Structural and functional correlates of motor imagery BCI performance: Insights from the patterns of fronto-parietal attention network
Introduction
Motor imagery (MI) is an internal mental rehearsal of a special motor action without overt motor output, which reflects high-level aspects of action planning. Although this is an abstract description, numerous studies have demonstrated that MI plays a crucial role in motor skill learning, rehabilitation of motor abilities and prosthesis control (Burianova et al., 2013, Miller et al., 2010). Meanwhile, brain computer interface (BCI) approaches hold promise to provide effective treatment for people with motor impairments, such as spinal cord injury (SCI), stroke, and amyotrophic lateral sclerosis (ALS) (Lebedev and Opris, 2015). Moreover, BCIs can be utilized to enhance normal brain function (i.e., sports skills) (Krauledat et al., 2009). Therefore, many recent efforts to develop MI-based BCI systems to obtain voluntary neural electroencephalography (EEG) signals (i.e., sensorimotor rhythm (SMR) or mu rhythm) for the paralyzed patients or healthy individuals, allowing to control external devices and to better understand cognitive behaviors (Alvarez-Meza et al., 2013, Friedrich et al., 2013, Miller et al., 2010).
In real-word BCI applications, there are large inter-individual differences in MI-BCI performance, not all individuals show satisfactory performance. For example, according to a study conducted by Blankertz et al. (2010b), approximately 15% to 30% of subjects cannot successfully voluntarily control an SMR-BCI, even after several weeks of training. In an earlier study published in 2003, the percentage of subjects who achieved a classification accuracy below 70% was even greater (48.7%) (Guger et al., 2003). Therefore, it is important to better understand the reasons for the individual differences in MI-BCI performance and find reliable biomarkers to predict individual MI-BCI performance (Blankertz et al., 2010a). The development of predictors could identify potentially inefficient SMR-BCI subjects, thereby avoiding frustrating and costly training procedures. Thus, other studies on this topic may be instructive for the establishment of enhanced training strategies for subjects who exhibit poor performance on these tasks (Vidaurre and Blankertz, 2010).
Due to the rapid development of neuroimaging techniques, such as EEG, functional magnetic resonance imaging (fMRI), and structural MRI (sMRI), the understanding of MI mechanisms has been progressively improving. Using the direct neuro-recording, several studies revealed that the primary motor cortex (M1) and posterior parietal cortex (PPC) are rich sources of MI EEG signals that can be used to control BCIs (Aflalo et al., 2015, Hochberg et al., 2006). Meanwhile, MRI technique is also growing attraction and has gain researchers attention. Based on fMRI, MI-BCI performance has been found to be correlated with the activation of the supplement motor area (SMA) (Halder et al., 2011), premotor-parietal network (Hanakawa et al., 2003) and a large fronto-parietal network (Hetu et al., 2013). Moreover, studies based on sMRI revealed that the gray matter volume of the SMA, supplementary somatosensory area and dorsal premotor cortex (Kasahara et al., 2015) and the fractional anisotropy (FA) of the cingulum, corpus callosum and superior fronto-occipital fascicle (Halder et al., 2013) are closely correlated with MI-BCI performance. These related studies were mainly focused on task-related performance and consistently illustrated the important role of the fronto-parietal regions in MI-BCI performance. Although many efforts (Kasahara et al., 2015, Zich et al., 2015) were paid to understand these patterns, the underlying neural mechanism remains unclear.
In recent years, functional connectivity changes in an intrinsic resting-state brain network (i.e., the fronto-parietal attention network (FPAN) and default mode network (DMN)) and neural structural patterns have been increasingly used to investigate cognitive performance (Alavash et al., 2015, Kasahara et al., 2015, Markett et al., 2014). The FPAN is a task-positive network consisting of the areas of the cortex located along the intraparietal sulcus (IPS), dorsolateral prefrontal cortex (DLPFC), inferior parietal lobe (IPL), SMA, and frontal eye field (FEF) (Fox et al., 2005, Ptak, 2012) that is crucially involved in high-level cognitive processes, such as attention and working memory (Markett et al., 2014, Naghavi and Nyberg, 2005, Ptak, 2012, Scolari et al., 2015). Studies have also suggested that MI has a critical functional relationship with these high-level cognitive processes (Ashley Fox, 2013, Madan and Singhal, 2012). Sustained attention and working memory are two crucial factors for healthy subjects to successfully control an MI-based BCI, and general mind wandering or lapses in attention can undermine the user's efforts in performing this task (Friedrich et al., 2013, Lakey et al., 2011). These studies imply that MI-BCI performance relies on the interactions between high-level cognitive and low-level motor functions (Lebedev and Opris, 2015, Moxon and Foffani, 2015). Thus, in the current study we combined EEG and MRI to characterize the relationships between the FPAN and MI-BCI performance.
Functional and structural MRI may offer two complementary sources of information to facilitate an understanding of the relationships between the FPAN and MI-BCI performance. First, the graph tool from network analysis is an important method for capturing the intrinsic functional organization of the brain, which typically reflects the exchange of information (integration and segregation) among brain regions (Sporns, 2013b, Tuladhar et al., 2015). In a functional brain network, the hub plays a critical role in information processing and translation by altering its functional connectivity with other nodes (brain regions) within the network to modulate the various cognitive processes (Cole et al., 2013, Zanto and Gazzaley, 2013). Insults to a hub of a network will result in a disproportionately high impact on behavior or severe cognitive impairment (Osada et al., 2015). A variety of methods allow for the characterization of the importance or āhubnessā of a node in the network, and each measure seems to reflect unique network patterns (Sporns, 2013b, Zuo et al., 2012). Here, the degree centrality (DC) and eigenvector centrality (EC), two common centrality measurements, were selected to assess the network properties of the FPAN (Lohmann et al., 2010, Sato et al., 2015). Second, measurement of the cortical thickness (CT) is other important method to capture the cortical morphology feature of the brain. CT reflects cellular characteristics, such as myelination, cell size, and cell packing density (Lerch et al., 2006, Narr et al., 2007). Several studies have assessed cognitive ability in healthy subjects or in populations with mental disorders using the changes of CT that typically may reflect structural reorganization (local alterations or network-level modulations) (Voss and Zatorre, 2015, Zielinski et al., 2014). Thus, the combination of regional CT and functional network hub (DC and EC) evaluations may provide new insights into the associations between the FPAN and MI-BCI performance.
We hypothesized that determining the specific patterns of FPAN organization, as reflected by the resting-state functional network and the regional morphometric changes in cortical structural, would facilitate our understanding of the individual differences in MI-BCI performance. We also hypothesized that these functional and structural patterns of the FPAN could be used to predict individual MI-BCI performance. Therefore, in the present study, we assessed the structural and functional patterns of the FPAN at the node level. Specifically, we examined one structural measure (CT) and two resting-state functional network node-centrality measures (DC and EC). Based on these functional and structural measurements, we assessed the effects of different patterns of the organization of the FPAN on individual MI-BCI performance using a correlation analysis and a receiver operating characteristic (ROC) analysis. Linear discriminant analysis (LDA) and support vector machine (SVM) classifiers were then used to identify subjects with poor MI-BCI performance.
Section snippets
Subjects
A total of 40 healthy university students were initially recruited, and 26 (9 females and 17 males, aged 22.85Ā Ā±Ā 2.48Ā years, range 19ā26Ā years, 24 right hand-dominant) agreed to complete the EEG and MRI recordings. The subjects did not habitually consume drugs and alcohol, and had no cognitive impairments or neurological disorders. Two subjects had previous experience with MI-based BCI. The experimental protocol was approved by the Institutional Research Ethics Board of the University of Electronic
MI-BCI offline performance
Of the 26 subjects who fully completed the EEG and fMRI recordings in the study, three subjects were excluded due to excessive motion during resting-state fMRI scanning. Table 2 list a detailed overview of all subjects. The MI-BCI performances were used to divide the subjects into low- and high-aptitude groups. Following the criteria proposed in the Halder studies (Halder et al., 2011, Halder et al., 2013), the median performance value (78%) was used to separate the subjects into good and poor
Discussion
In the current study, we applied weighted graph measures to understand how specific patterns of FPAN organization (intrinsic functional connectivity) during rest affect individual MI-BCI performance. We also used structural MRI to quantify a specific regional-dependent cortical structural morphometric feature (cortical thickness) within the FPAN to further understand individual variations in MI-BCI performance. Our findings showed that node-level functional or structural changes in three ācoreā
Conclusions
In this study, we found that the structural and functional patterns of the FPAN are associated with MI-BCI performance. Our findings revealed that the individuals who have efficient FPAN would perform better on MI-BCI. Moreover, combining the structural and functional features (i.e., EC and CT of the left IPL) of the FPAN, we were able to accurately identify individuals as low- or high-aptitude BCI users using the machine learning method. Therefore, our study will be helpful for improving our
Conflict of interest
None.
Acknowledgements
This work was supported in part by grants from the National Natural Science Foundation of China (# 61522105, # 61175117, # 81330032), the program for New Century Excellent Talents in University (# NCET-12-0089), the 863 Project (# 2012AA011601), and the National Science & Technology Pillar Program (# 2012BAI16B02).
References (101)
- et al.
Is functional integration of resting state brain networks an unspecific biomarker for working memory performance?
NeuroImage
(2015) - et al.
Neurophysiological predictor of SMR-based BCI performance
NeuroImage
(2010) - et al.
Neurophysiological predictor of SMR-based BCI performance
NeuroImage
(2010) Centrality and network flow
Soc. Networks
(2005)- et al.
Multimodal functional imaging of motor imagery using a novel paradigm
NeuroImage
(2013) - et al.
The chronnectome: time-varying connectivity networks as the next frontier in fMRI data discovery
Neuron
(2014) - et al.
Cortical surface-based analysis. I. Segmentation and surface reconstruction
NeuroImage
(1999) An introduction to ROC analysis
Pattern Recogn. Lett.
(2006)- et al.
Tracking the mind's image in the brain I: time-resolved fMRI during visuospatial mental imagery
Neuron
(2002) - et al.
Graph analysis of the human connectome: promise, progress, and pitfalls
NeuroImage
(2013)