Assessing the utility of Fronto-Parietal and Cingulo-Opercular networks in predicting the trial success of brain-machine interfaces for upper extremity stroke rehabilitation

For stroke participants undergoing motor rehabilitation, brain-machine/computer interfaces (BMI/BCI) can potentially improve the efficacy of robotic or exoskeleton-based therapies by ensuring patient engagement and active participation, through monitoring of motor intent. In such interventions, exploring the network-level understanding of the source space, in terms of various cognitive dimensions such as executive control versus reward processing is fruitful in both improving the existing therapy protocols as well as understanding the subject-level differences. This contrasts to traditional approaches that predominantly investigate rehabilitation from resting state data. Moreover, conventional BMIs used for stroke rehabilitation barely accommodate people suffering from moderate to severe cognitive impairments. In this first-of-the-kind study, we explore the cognitive dimensions of a BMI trial by probing the networks that are core to the BMI performance and propose a network connectivity-based measurement with the potential to characterize the cognitive impairments in patients for closed-loop intervention. Specifically, we tease apart the extent of cognitive evaluation versus executive control aspects of impairments in these patients, by measuring the activation power of a major cognitive evaluation network-the Cingulo-Opercular Network (CON) and a major executive control circuit-the Fronto-Parietal network (FPN), and the connectivity between FPN-CON. We test our hypothesis in a previously collected dataset of electroencephalography (EEG) and structural imaging performed on stroke patients with upper limb impairments, while they underwent an exoskeleton-based BMI intervention for about 12 sessions over 4 weeks. Our logistic regression modeling results suggest that the connectivity between FPN and CON networks and their source powers predict trial failure accurately to about 84.2%. In the future, we aim to integrate these observations into a closed-loop design to adaptively control the cognitive difficulty and passively increase the subjects motivation and attention factor for effective BMI learning.