Neural Interface System for Virtual High-Density Microelectrode Array Adaptive Neuromodulation

Abstract

The ultimate tool in neuroscience would offer the ability to record from every cell and adaptively stimulate each neuron based on inferred states. Microelectrode arrays (MEAs) have been developed to observe and probe neural populations with subcellular resolution, however, processing the large volumes of generated data streams is a critical bottleneck. Presented here is a microelectrode neural interface system (µNIT) which relaxes the physical electrode density requirement by combining lowdensity MEA technology with the generation of virtual electrodes. This increases the effective electrode density while minimizing the required processing overhead. µNIT integrates hardware accelerators for the real-time analysis of neuron-level activity, and for the adaptive generation of responsive electrical stimuli. Spike clustering is performed using an exponentially decaying memory-based autoencoder (EDM-AE) at a rate of up to 2676 spikes/second across all channels. Inferred states are used to adaptively control programmable waveform generators which create per-electrode stimuli at a 22.3 KHz sample rate. The processing system is demonstrated in vitro with clinically resected human brain tissue using a 60-channel microelectrode array.