Modeling Local Neural Population Responses to Intracortical Microstimulation
Joel Ye, Carnegie Mellon University
Intracortical microstimulation (ICMS) is a primary tool for driving neural population activity, and may provide an avenue to create rich touch percepts when delivered to the somatosensory cortex (S1) in intracortical brain-computer interfaces. However, current stimulation strategies, which are limited to fixed patterns of stimulation pulses, do not provide this rich feedback. These patterns can be coarsely optimized with participant reports about the evoked percepts, but this process is time-consuming and cannot be efficiently scaled to complex stimulation patterns. We thus propose to characterize the evoked neural response as a step towards more sophisticated stimulation.
To do so, we delivered ICMS trains of varying amplitude, timing, and electrode usage in S1 for two participants. During stimulation, we simultaneously recorded full bandwidth data (30 kHz) from these same electrodes using custom headstages (Blackrock Microsystems). We developed two models to characterize the spiking neural response to stimulation. A first deep network artifact estimator is used to extract the spiking response from the voltage recordings (electrically contaminated by stimulation), with spike detection within 1.3 ms of pulse offset. A separate deep network captures the relationship between the commanded stimulation and observed response. Ablations show the evoked response is affected by the ongoing neural activity’s interaction with stimulation parameters.
Next, to understand the exponentially large stimulation parameter space, we measure the response model’s generalization, comparing how performance changes when fit to data from different stimuli. We find some generalization to novel pulse timing and amplitude, and more modest generalization to different electrodes, suggesting a rich neural response space. Finally, we find that multi-session data aggregation is feasible, providing a practical strategy for building a rich ICMS to neural response model.
Authors: Joel Ye1-3, Leila Wehbe1,3,4, Robert Gaunt2,5
1Carnegie Mellon University, USA
2Rehab Neural Engineering Labs, USA
3Neuroscience Institute, Carnegie Mellon University, USA
4Machine Learning Department, Carnegie Mellon University, USA
5Department of Physical Medicine and Rehabilitation, University of Pittsburgh, USA
Abstract Author(s): (see above entries)