Decoding cortical response during motor tasks using brain connectivity

Abstract

Sensorimotor integration is defined as the capacity of the central nervous system to integrate different sources of stimuli and transform such inputs in motor actions. Traditional approaches to measure sensorimotor dynamics are based on sensorimotor rhythms detected by Electroencephalography (EEG), such as the Event-Related Desynchronization (ERD) and Event-Related Synchronization (ERS). However, it is still not clear what are the underlying cortical dynamics involved in voluntary movements, and there is a lack of understanding of the temporal flow patterns related to a task. Although the models for motor decoding have improved considerably over the past decade, Brain-Machine Interfaces (BMIs), such as those attempting to control upper-limb prostheses, are still far from the goal of reaching naturalistic and dexterous control like our natural limbs. In this thesis, a model using connectivity estimators on EEG signals is proposed, with the aim of mapping cortical dynamics involved in sensorimotor integration while performing motor tasks. Here, special focus is given to wrist movements, since they are extremely important for proper handling of objects and have not been adequately explored in current literature associated with neural and standard control models used for upper-limb prosthesis. After initial screening, Mutual Information (MI) was chosen as the connectivity strategy for the aforementioned task. To estimate the most important channels and connectivity pairs of MI, a preliminary analysis based on the differences between resting and execution was performed. After the selection, MIs were estimated at higher temporal resolution, and separated in alpha and beta bands, from which a set of features was extracted and used as input for a Support Vector Machine (SVM) classifier to estimate the motor tasks (wrist pronation and wrist supination). For validation, we also estimated motor tasks using a conventional method for sensorimotor analysis, extracting significant ERD components and classified the data using SVM. The results showed higher accuracies when using the proposed model in beta band and MI features (89.65%). Alpha band resulted in 73.68% accuracy. On the other hand, using ERD, the accuracy of the classifier was 60.73% and 62.49% for alpha and beta bands, respectively. We conclude that the proposed method using functional connectivity and a proper model for the selection of important pairs over specific frequency bands has better response in identifying wrist movements. This strategy could be potentially applied in BMIs that control prosthetic devices at various levels.