Signal processing and deep learning methods for robust and accurate intracortical brain-machine interfaces

Abstract

Brain-machine interfaces (BMIs) have emerged as a promising assistive technology for restoring lost motor functions in individuals with neurological disorders (e.g. spinal cord injury, amyotrophic lateral sclerosis, stroke) by allowing them to control external devices directly with their thoughts. The last two decades have seen impressive progress in research and pilot clinical trials of BMIs using intracortical microelectrode arrays. Most high-performance BMIs use well-isolated neuronal action potentials (spikes) also known as single unit activity (SUA) as the input signal. These SUA-based BMIs have been shown to suffer from signal instability and longevity, therefore hampering their translation into widespread clinical use.

To mitigate this problem, numerous studies have proposed two alternative neural signal inputs: multiunit activity (MUA) and local field potential (LFP). MUA represents the aggregate spikes from an ensemble of neurons, whereas LFP is thought to mainly reflect summed synaptic activity from a local population of neurons around the recording electrode. Despite offering better signal stability and longevity, MUA- and LFP- based BMIs have been shown by several studies to yield lower decoding accuracy than that of SUA-based BMIs. Thus, developing robust and accurate BMIs remains a major challenge.

This thesis addresses the aforementioned challenge from neural signal processing and decoding points of view and is composed of three parts. The first part focuses on our efforts to push the performance of MUA-based BMIs forward. Specifically, we propose an adaptive firing rate estimation method as feature extraction and evaluate its impact on the decoding performance. The second part describes our attempts to improve the performance of LFP-based BMIs further. Particularly, we develop decoding algorithms based on two variants of deep learning methods, namely long short-term memory (LSTM) and temporal convolutional network (TCN), and assess their decoding performance. In the last part of this thesis, we explore another alternative neural signal referred to as entire spiking activity (ESA) and different deep learning architectures to improve both the robustness and accuracy of decoding. We then compare the decoding performance of ESA-based BMIs with that of SUA-, MUA-, and LFP-based BMIs described in the previous parts. Furthermore, we investigate the relationship between ESA and LFP by examining whether ESA can be solely inferred from LFP with high accuracy. Overall, this thesis offers a new approach to advancing BMI research and gives insights into signal processing and decoding design consideration towards clinically viable translation of BMIs.