Toward High Performance, Power Efficient Brain-Machine Interfaces

Abstract

Brain-machine interfaces (BMIs) are a promising solution for restoring mobility and communication to people who suffer from sensorimotor impairments, including spinal cord injury, stroke, and neurodegenerative diseases. Intracortical BMIs consist of an electrode array implanted in the brain, a signal processing pipeline to amplify and extract neural features, and a decoding algorithm to predict the user’s intentions. The aim of this work is to develop high performance, real-time decoding algorithms that reduce the power consumption of the entire BMI, facilitating clinical translation toward fully implantable, wireless systems.

One method for improving decoding accuracy is to increase the number of recording electrodes within the cortex. Wireless, “mote” electrodes may enable this scaling but are significantly limited in their power consumption due to thermal restrictions within the body. In the first study we investigate and optimize the use of a low power, low data-rate wireless communication scheme that allows for thousands of wireless mote electrodes to transmit neural data. We show decoding performance can be maintained despite a large reduction in data-rate and simulate a wireless receiver architecture that can process the received data while staying within power limits.

In the second study, we move one step down the processing pipeline to investigate the use of artificial neural network decoder architectures for real-time movement prediction. We find that recurrent neural networks (RNNs) can outperform other feedforward architectures for finger movement decoding, where performance can approach that of an able-bodied user when the task is simplified. We also find that the training data and task has an influence on the decoder’s internal dynamics, where the RNN can act both like a classifier and a continuous decoder for different output dimensions.

While RNNs can achieve fast and accurate closed-loop control, one may want the ability to control how precisely they replicate learned movement patterns versus being able to produce a full repertoire of movements. In the third study we develop a training regularization that can increase or decrease the strength of the decoder’s internal dynamics, to help control the decoder’s degree of task memorization. We find that decoders can have similar accuracy with varied internal processing methods, and that decoders with strong dynamics may have poor closed-loop control.

Improvements in BMI performance have been achieved here and in other recent works using higher channel counts and using computationally heavy neural networks with greater accuracy. However, both changes increase the power consumption of the BMI, making it challenging to translate to a fully implantable system with a long battery life. In the final study, we aim to reduce this power consumption by reducing the number of active channels and by compressing the decoder. We find that over half of the channels may be turned off without performance loss, and that neural network decoders can be compressed by over 100x for lower power consumption and the ability to run the decoder on the wireless device.

The work presented here aims to accelerate the widespread use of high performing, implantable BMIs with the hope of restoring mobility and function to those who suffer from debilitating neurological impairments. Future work will include testing these algorithms in humans and with potentially larger electrode arrays, as well as testing their ability to generalize to more complex tasks.