|Ph.D Student||Furman Daniel|
|Subject||Features of Finger Flexion and Motor Learning for Control|
of Brain-Machine Interfaces
|Department||Department of Medicine||Supervisor||Professor Emeritus Hillel Pratt|
|Full Thesis text|
Brain-machine interfaces create a channel of communication by transforming brain activity into electronic commands. Commands produce actions that are sensed by sensory systems, and evoke perceptions in the brain. Based on these perceptions, the brain generates new activity that leads to new or updated actions and perceptions. This “perception-action” cycle goes on and on as the brain interacts with its environment in a continuous feedback loop.
The present research addresses this multidisciplinary topic of perception-action feedback loops in the human brain. Specifically, the present research focused on the sensorimotor system, and studied the neurobiology of motor learning and motor execution, which correspond to the main branches of this system’s perception-action loop. Furthermore, the research extended its neurobiological findings and insights into the applied science of brain-machine interfacing, with the goal of improving the clinical utility of future brain-machine interfaces.
The study consisted of two experiments. Both used non-invasive scalp-recorded electroencephalography (EEG). The first experiment focused on the perception branch of the sensorimotor loop. Nineteen (19) subjects performed a center-out reaching task in which endpoint feedback was visually perturbed in ~30% of trials, inducing prediction error-related information processing. This experiment emulated a brain-machine interface user perceiving an error in the interface’s inference of their intent. One group of subjects experienced sensory perturbations only (i.e., no explicit reward feedback) while the other group experienced sensory and reward perturbations simultaneously.
We hypothesized that distinct cortical sources were involved in the processing of different types of prediction errors, and found that the frontopolar prefrontal cortex was significantly activated when reward was absent, whereas the secondary sensorimotor cortex was significantly activated when reward was present. These findings carry implications for leading theories of motor learning, as well as future brain-machine interface control strategies that exploit error-related brain activity to continuously calibrate and autocorrect.
Experiment 2 studied the action branch of the sensorimotor loop through an investigation of fine motor control. For the model condition, we chose voluntary finger flexion, both real and imagined. We hypothesized that a deeper understanding of the sources and networks involved in such fine motor control would facilitate the design of more effective signal processing and classification methods. The results showed real and imagined finger flexion to be encoded through significantly different activity in the superior parietal lobe, secondary sensorimotor cortex, and inferior occipital gyrus. Differences were especially distinguished in the left hemisphere.
A new approach to spatiotemporal and spectral feature extraction was then designed to capture key features of finger imagery electrophysiology and to train a classifier to discriminate between different imagined finger movements. This approach obtained a mean population accuracy of 30.86±1.76%, nearly twice (1.85) chance level (16.71±1.68%) for the six-class task of finger imagery evaluated; outperforming equivalent real finger movement classifiers reported in the literature. This suggests that finger imagery may be used to control brain-machine interfaces.
In conclusion, findings from both experiments extended knowledge within the respective fields of motor learning and motor execution, and developed physiologically plausible and technically feasible approaches for better brain-machine interface control.