EEG-based brain-machine interface to a powered exoskeleton for walking and standing: A longitudinal dataset for healthy able-bodied subjects

Brain-machine interfaces (BMI) are systems that translate brain activity into motor commands to control physical or virtual<br> machines such as computers, prostheses, exoskeletons, and avatars. BMI systems have been deployed as assistive devices,<br> such as wheelchairs and robotic systems, for individuals with paralysis due to spinal cord injury; to control artificial limbs or<br> exoskeletons to restore movement after limb amputation or brain injury; or as communication devices (e.g., brain-computer<br> interfaces). However, significant challenges remain, including understanding how BMI training affects brain activity and BMI<br> performance. In this report, we present a multimodal longitudinal dataset containing simultaneously recorded high-density<br> scalp electroencephalography (EEG), electrooculography (EOG), head motion, and the internal states of a powered robotic<br> exoskeleton controlled by a BMI system in a cohort of seven healthy able-bodied subjects. In an initial decoder calibration phase,<br> the participants performed kinesthetic motor imagery of walking and stopping motions according to audible beep instructions in<br> a straight path with the exoskeleton controlled by an operator remotely. This allowed synchronized acquisition of robot motion<br> and EEG data while ensuring user engagement in the tasks. In a second phase, after decoder calibration using Localized<br> Fisher Discriminant Analysis dimensionality reduction and a Gaussian Mixture Model classifier on lower delta band (0.1-2 Hz)<br> EEG signals, the subjects performed closed-loop BMI training sessions in which the BMI output controlled the start and stop of<br> walking motions of the exoskeleton. The BMI performed asynchronous classification (Walk vs. Stop states) which continuously<br> classifies the user’s intention during the motor and stop execution. While both the decoder calibration and closed-loop BMI<br> training phases occurred every session, the decoder parameters were fixed after the fifth session, and the participants’ ability<br> to accurately control the exoskeleton’s walking or stopping state was measured. Overall, the decoder training and BMI training<br> sessions were executed over multiple sessions ranging from 15 to 81 days. In this report, we present the rationale for collecting<br> these longitudinal data, a detailed explanation of the experimental setup and data acquisition procedures, and a brief validation<br> of the data quality of the dataset.