Source: https://www.nature.com/articles/s41587-019-0231-y
The convergence of consumer neurotech, signal processing, and deep learning has enabled an explosion of accessibility in assistive technology and neurorehabilitation. To date, high-fidelity decoding of a user's motor intent requires an invasive surgical procedure to implant microelectrodes in the primary motor cortex (M1). While invasive BCIs offer unparalleled signal-to-noise ratios, the associated costs of surgery, neural tissue damage, and limited economic viability severely limit the technology's reach. Non-invasive sensorimotor rhythm decoders, which read microvolt-scale EEG and fNIRS fluctuations from the scalp, have enabled the creation of assistive software applications and robotic orthoses (e.g., functional electrical stimulation) to help patients control their environment. These non-invasive interfaces decode the user's motor intent to actively control real-time software applications and facilitate neuroplastic motor recovery. The underlying mechanism by which non-invasive motor intent decoders work is the measurement of event-related desynchronization (ERD) and event-related synchronization (ERS) in the sensorimotor rhythm (SMR). When a user performs or imagines a certain motor movement (e.g., flexing a finger), neurons in the associated region of the primary motor cortex fire in synchrony, decreasing overall power in that region of the brain. Modern commodity EEG sensors are able to pick up these subtle fluctuations in electric potential across the scalp. These microvolt-scale signals must be processed using spatial filters (e.g., common spatial patterns) and convolutional deep learning networks to extract relevant features indicative of a motor movement class. After removing eye, muscle, and powerline noise, the resulting signal is then classified by foundation model neural networks into one of several potential motor movement classes. In neurorehabilitation use cases, motor imagery BCIs are not only used as assistive controllers but can actively facilitate neuroplastic motor recovery after a stroke, spinal cord injury, or traumatic brain injury. The human brain exhibits extraordinary plasticity in its ability to form new neural connections when a previously established connection is damaged (e.g., when a stroke destroys a region of the primary motor cortex). This ability can be leveraged by a closed-loop motor imagery BCI to help patients regain voluntary motor control of an afflicted body part. When a stroke patient imagines moving a body part that they are unable to move voluntarily, a motor imagery decoder picks up the associated decrease in alpha and beta band power over the affected region of the primary motor cortex and relays this information to a robotic exoskeleton or functional electrical stimulation device, causing it to move the afflicted body part. The combination of this sensory feedback with the user's imagined movement is sufficient to cause neuroplastic recovery of the afflicted region in many patients. The application of non-invasive BCIs to read out motor intent using consumer EEG equipment presents some additional technical challenges not found in research or clinical settings. Scalp EEG recordings have notoriously poor signal-to-noise ratios and spatial resolution due to the dampening effect of the skull on the electric potential field produced by active neurons. Additionally, the signal characteristics of a user's EEG can vary drastically between sessions due to differences in electrode placement and user discomfort, requiring repeated recalibration between sessions. To address these issues, modern BCI software stacks employ session-adaptive transfer learning to allow a single BCI model to operate across a wide range of users. This combination of commodity non-invasive EEG hardware and modern signal processing and neural decoding techniques has enabled the development of a rich ecosystem of end-user assistive applications built on top of motor imagery decoders running on standard PC hardware. Using motor imagery of motor movements followed by sensory feedback from an assistive controller (e.g., robotic exoskeleton, functional electrical stimulation unit), these BCIs actively participate in the neurorehabilitation process to help patients regain voluntary motor control after a stroke, spinal cord injury, or traumatic brain injury. As the underlying electrode hardware improves and neural decoders become more robust and adaptive, non-invasive motor imagery BCIs will continue to play a leading role in the democratization of neurotech for the masses.
Sources
https://www.sciencedirect.com/science/article/pii/S2324242622000080 https://www.mdpi.com/2306-5354/12/8/820