Brain computer interfaces
Plug-and-play 2D cursor control
In my research at UCSF, I used electrocorticography (ECoG) based BCIs for motor neuroprostheses. ECoG places electrodes directly on the brain tissue and is known to be stable over time without degradation in signal quality. In our first project, to achieve 2D cursor control, we designed a Kalman Filter (KF) that was co-adaptive. Not only did the algorithm update its parameters to the user's brain activity, but it did so with a time-constant that in turn allowed the user to adapt their own brain activity and learn how to use the BCI. Our efforts resulted in a stable "plug-and-play" neural map of control, i.e., a decoder that did not require recalibration. We were able to then stack other control dimensions such as the ability to "point and click" (Silversmith*, Abiri*, Hardy*, Natraj* et. al, 2021).
Demo of the BCI system wherein the user (Bravo1) is controlling a 2D cursor and is able to successfully point and click at targets
high DoF BCI control of a robotic arm and hand
Our next project sought to expand neuroprosthetic control to higher degrees of freedom (hDoF) assistive devices such as the Kinova Jaco robot. Our goal was to allow long-term stable complex BCI control of reach-to-grasp and object manipulation capabilities. This is a high priority for patients with paralysis. Given the complexity of hDoF BCIs, co-adapative decoding algorithms - that involve the formation of a new BCI motor map - might necessitate long learning periods that can hinder real-world adoption. We wondered whether we could lessen the burden on the user, thereby allowing for rapid adoption. To this end, we studied principles underlying the balance between representational stability and plasticity of well-rehearsed, simple imagined movements (such as finger flexion or tongue protrusion). Our impetus was to understand how we could leverage well-rehearsed actions directly as neural commands for hDoF BCIs. Strikingly, we found that when used as discrete neural commands, the statistics of well-rehearsed motor commands (especially variance) could be flexibly adapted for BCI despite drift and without somatotopic changes. Outside the BCI ,the movements maintained a stable representational structure.
Leveraging the flexibility of stable representations and accounting for drift in a novel decoding framework led to proficient and stable BCI control of the Kinova Jaco robotic arm and hand for over 7 months (Natraj et. al., 2023) during complex reach to grasp and object manipulation tasks. We were thus able to achieve long-term hDoF PNP control that can eventually improve real-world functionality in those with paralyses.