Machine Learning Classification to Identify the Stage of Brain-Computer Interface Therapy for Stroke Rehabilitation Using Functional Connectivity

Now if we could just have protocols written on this the next researchers could build on the previous research. But that requires a strategy and leadership which we don't have. 
https://www.frontiersin.org/articles/10.3389/fnins.2018.00353/full?

Rosaleena Mohanty^1,2*, Anita M. Sinha^1,3, Alexander B. Remsik^1,4, Keith C. Dodd^1,3, Brittany M. Young^5,6, Tyler Jacobson^1,7, Matthew McMillan^1,3, Jaclyn Thoma^1,6, Hemali Advani¹, Veena A. Nair¹, Theresa J. Kang¹, Kristin Caldera⁸, Dorothy F. Edwards⁴, Justin C. Williams³ and Vivek Prabhakaran^1,5,6,9

• ¹Department of Radiology, University of Wisconsin-Madison, Madison, WI, United States
• ²Department of Electrical Engineering, University of Wisconsin-Madison, Madison, WI, United States
• ³Department of Biomedical Engineering, University of Wisconsin-Madison, Madison, WI, United States
• ⁴Department of Kinesiology, University of Wisconsin-Madison, Madison, WI, United States
• ⁵Medical Scientist Training Program, University of Wisconsin-Madison, Madison, WI, United States
• ⁶Neuroscience Training Program, University of Wisconsin-Madison, Madison, WI, United States
• ⁷Deparment of Psychology, University of Wisconsin-Madison, Madison, WI, United States
• ⁸Department of Orthopedics and Rehabilitation, University of Wisconsin-Madison, Madison, WI, United States
• ⁹Department of Medical Physics, University of Wisconsin-Madison, Madison, WI, United States

Interventional therapy using brain-computer interface (BCI) technology has shown promise in facilitating motor recovery in stroke survivors; however, the impact of this form of intervention on functional networks outside of the motor network specifically is not well-understood. Here, we investigated resting-state functional connectivity (rs-FC) in stroke participants undergoing BCI therapy across stages, namely pre- and post-intervention, to identify discriminative functional changes using a machine learning classifier with the goal of categorizing participants into one of the two therapy stages. Twenty chronic stroke participants with persistent upper-extremity motor impairment received neuromodulatory training using a closed-loop neurofeedback BCI device, and rs-functional MRI (rs-fMRI) scans were collected at four time points: pre-, mid-, post-, and 1 month post-therapy. To evaluate the peak effects of this intervention, rs-FC was analyzed from two specific stages, namely pre- and post-therapy. In total, 236 seeds spanning both motor and non-motor regions of the brain were computed at each stage. A univariate feature selection was applied to reduce the number of features followed by a principal component-based data transformation used by a linear binary support vector machine (SVM) classifier to classify each participant into a therapy stage. The SVM classifier achieved a cross-validation accuracy of 92.5% using a leave-one-out method. Outside of the motor network, seeds from the fronto-parietal task control, default mode, subcortical, and visual networks emerged as important contributors to the classification. Furthermore, a higher number of functional changes were observed to be strengthening from the pre- to post-therapy stage than the ones weakening, both of which involved motor and non-motor regions of the brain. These findings may provide new evidence to support the potential clinical utility of BCI therapy as a form of stroke rehabilitation that not only benefits motor recovery but also facilitates recovery in other brain networks. Moreover, delineation of stronger and weaker changes may inform more optimal designs of BCI interventional therapy so as to facilitate strengthened and suppress weakened changes in the recovery process.

Introduction

Recent advancements in neurotechnology have led to the emergence of the brain-computer interface (BCI), which records neural signals and translates them into signals that can control assistive devices, such as computers or prostheses. To date, BCI-based approaches are being investigated as therapeutic strategies to facilitate recovery for several neurological diseases, including stroke, epilepsy, and Parkinson's Disease. For stroke, the long-term objective of the rehabilitation is to improve impaired brain functions so as to restore autonomy in daily activities for stroke survivors. While conventional approaches such as physical therapy and occupational therapy have proven to be successful in aiding stroke recovery in the acute and sub-acute stages (Bütefisch et al., 1995; Gordon et al., 2004) modern technologies involving robotics (Kwakkel et al., 2008), transcranial magnetic stimulation (Corti et al., 2012), and virtual reality (Lohse et al., 2014) have demonstrated promise in promoting additional motor and cognitive recovery to improve autonomy and overall quality of life for stroke survivors even in the chronic stages. The use of an electroencephalogram (EEG)-based brain-computer-interface (BCI) is an unconventional rehabilitation strategy that has emerged as a potentially effective therapeutic modality for promoting motor recovery in patients with stroke (Silvoni et al., 2011). An EEG-based BCI detects and uses a patient's neural signals as inputs to provide real-time feedback, effectively enabling users to modulate their brain activity (Felton et al., 2009). Additional feedback presented by means of functional electrical stimulation (FES; De Kroon et al., 2002) and tongue stimulation (TS) (Wilson et al., 2012) also provide users with multi-modal feedback as a form of reward for producing certain brain activity patterns while performing tasks. While BCI therapy is often explicitly targeted at restoring motor functions, simultaneous changes in non-motor-related functions in the brain may also result after intervention; to date, neural reorganization of cortical regions outside of the motor network is not well-characterized. Distinction between the overall brain state before and after the therapy could facilitate a more thorough understanding of the mechanisms underlying both the strengthening and/or weakening in motor and non-motor networks in participants. Access to this information could allow us to optimize the design and execution of this therapy for stroke rehabilitation.

While EEG allows for study of real-time brain activity during the BCI therapy with a high temporal resolution, neuroimaging methods have afforded us the ability to study both large-scale and small-scale reorganization of brain networks (Van Den Heuvel and Pol, 2010) at a relatively higher spatial resolution. Resting state functional magnetic resonance imaging (rs-fMRI), specifically, has been demonstrated as a powerful and attractive tool to study changes in brain functions as it is non-invasive, time-efficient, and task-free. Rs-fMRI allows us to measure the temporal correlation of the spontaneous, low-frequency (<0.1 Hz) blood-oxygen-level dependent (BOLD) signals across regions in the resting brain. Oscillations in the BOLD fMRI signals are indicative of cortical dynamic self-organization and have been associated with the neural reorganization underlying cognitive and motor function during stroke recovery (Lee et al., 2013; Bajaj et al., 2015). Previous studies have demonstrated that there are overlapping networks between the rs-fMRI-derived motor network and those observed during motor imagery and motor execution fMRI tasks (Grefkes et al., 2008; Nair et al., 2015). A growing number of studies have utilized neuroimaging methods to study the efficacy of BCI therapy in stroke recovery and found modulating changes in neuroplasticity and improvement in motor functions (Di Bono and Zorzi, 2008; Várkuti et al., 2013; Song et al., 2014; Young et al., 2014b; Nair et al., 2015; Soekadar et al., 2015). In the present study, we aim to use rs-fMRI to examine changes in neuroplasticity in whole-brain networks and to examine interactions between motor and non-motor cortical regions in chronic stroke participants following BCI therapy.

A whole-brain analysis resulting in high-dimensional data calls for the application of machine learning-based approaches which have become increasingly more integrated in neuroimaging analysis as they enable discovery of multivariate relationships beyond those identifiable by traditional univariate analysis. Several studies have underscored the utility of machine learning to not only differentiate among population groups (Dai et al., 2012; Meier et al., 2012; Rehme et al., 2014; Fergus et al., 2016; Khazaee et al., 2016; Ding et al., 2017) but also make predictions about behavioral outcomes using regression models (Dosenbach et al., 2010; Vergun et al., 2013; Mohanty et al., 2017), all of which have advanced our understanding of altered brain functionalities associated with several neurological diseases. In the context of BCI systems, linear and non-linear machine learning classification algorithms (Muller et al., 2003; Lotte et al., 2007) including support vector machines (SVMs; Rakotomamonjy and Guigue, 2008), nearest neighbors (Mason and Birch, 2000), and neural networks (Cecotti and Graser, 2011) have mainly been limited to improvement and optimization of the BCI2000 system from a design perspective to make the system more adaptive and user-friendly (Selim et al., 2008; Danziger et al., 2009; Alomari et al., 2013). Relatively fewer studies have applied machine learning techniques to elucidate the therapeutic impact of BCI interventional therapy in stroke patients based on the dynamics of brain connectivity changes. Specifically, SVM-based classifiers have demonstrated the ability to not only draw a distinction between different classes but also provide insight into underlying features that lead to the separation between them (Dosenbach et al., 2010; Vergun et al., 2013). Given that we aim to extensively investigate whole-brain effects of BCI therapy, a similar classification approach is befitting due to its efficiency in handling high-dimensional rs-fMRI data. Recent developments have brought deep learning approaches into view with applications in the field of medical imaging such as tissue/lesion/tumor segmentation (Birenbaum and Greenspan, 2016; Kamnitsas et al., 2017), image reconstruction/enhancement (Benou et al., 2016; Hoffmann et al., 2016) and population-based classification (Brosch et al., 2013; Payan and Montana, 2015). The efficiency of deep learning algorithms, however, is highly dependent on samples available for training a reliable model. Thus, we adhere to supervised machine learning classifiers given the limited sample size.

With the above considerations in mind, the goal of this study was to identify the stage of therapy using whole brain rs-fMRI data in stroke participants undergoing EEG-based BCI intervention along with additional feedback provided by FES and TS. We analyzed changes in non-motor regions of the brain in addition to the well-studied motor regions following BCI therapy in chronic stroke participants. To this end, we modeled this as a classification problem of discriminating between pre-therapy and post-therapy stages of intervention. Specifically, we illustrated using rs-fMRI that connectivity at the pre-therapy stage can be differentiated from that at post-therapy with reasonable accuracy. A SVM-based machine learning classifier was employed to identify specific functional nodes and connections in the brain between the two stages. The significance of this study is 4-fold: this study suggests that (i) a 10-min task-free rs-fMRI scan could aid in identifying and tracking changes in functional connectivity in the brain over the course of BCI therapy; (ii) SVM-based classification can automate the process of categorizing participants into pre-therapy or post-therapy stages and identify features discriminating between the stages of therapy; (iii) BCI therapy, targeted toward upper-extremity motor restoration, can promote recovery effects related to brain connectivity in both motor and non-motor networks; (iv) identification of specific functional changes that strengthen and weaken between stages of BCI-therapy could inform more tailored designs of BCI systems that facilitate stronger changes and suppress weaker changes to maximize the efficacy of this interventional therapy and improve outcomes for stroke survivors.

Much more at link/