Paralysis of the upper extremity is a devastating outcome of many neurological diseases and disorders. Brain-machine interfaces (BMIs) attempt to bypass the disability by recording information directly from the subject’s brain and predicting the user’s intentions to control a prosthetic device. Modern brain-machine interfaces have made limited translation to clinical use, where studies have not expanded far beyond controlled laboratory environments. Two of the primary hindrances to their widespread clinical translation is their dependence on stacks of power-hungry computers and performance compared to the able-bodied hand. The aim of this work is to establish low-power brain-machine interface technologies that restore fine control to paralyzed hands.

The first study presents the 300-1,000Hz spiking band power (SBP), which is a low power neural spiking feature that requires 90% less data than the standard threshold crossing rate (TCR) neural feature. In simulation, we found that SBP can extract accurate neural spiking patterns at lower signal-to-noise ratios and with greater unit specificity than TCR. Because of this, closed-loop decoders which used SBP performed as well or better than decoders using TCR in two rhesus macaques.

In the second study, we investigated whether BMIs could be implemented on embedded devices fit for implantation. We used three off-the-shelf low-power amplifiers controlled by a 32-bit microcontroller to perform SBP recording, feature extraction, and decoding. The device could achieve equivalent performance to our high-powered BMI when closed-loop predicting one-finger movements and comparable performance when predicting two-finger movements in a nonhuman primate. To do so, the device required 58.4mW (equivalent to 11.3hr usage time with a standard 200mAh implantable battery), which we could compress to 12.5mW (52.8hr usage time) with an optimized processing pipeline implemented on an integrated circuit.

The third study showed, for the first time, that BMIs can control the simultaneous and independent movements of two finger groups in real-time with nonhuman primates. With the BMI, the primate could acquire targets at a rate of nearly 2 per second. Additionally, we found that cortical activity for independent finger movements and combined finger movements were similar. This allowed linear models to predict behaviors that were not used for training with a correlation coefficient at least 90% as high as a linear decoder trained on all behaviors.

In the fourth study, we investigated how well continuous finger movements could be restored with a brain-controlled functional electrical stimulation (BCFES) system. Following temporary paralysis delivered via nerve block, a nonhuman primate improved success rates to 89% with a 1.4s median target acquisition time in a one-finger task by using the BCFES system, up from 2.6% and a 9.5s median target acquisition time (near chance) when using his paralyzed native hand. Additionally, we allowed the monkey to use the BMI (no stimulation) to complete the two-finger version of the task following paralysis, and performance could be recovered by performing recalibrated feedback-intention training one time following paralysis, despite the absence of sensory feedback.

The results of this work demonstrate that low-power BMIs can restore substantial function in cases of upper extremity paralysis. All of the work presented here uses low-power technology that can simply be implemented on implantable devices. Next steps for this work will require validation of the hand control results in humans and development of completely implantable neuroprostheses to restore native hand functions to people with paralysis.

Date: Monday, February 14, 2022
Time: 9:30 AM EST
Location: NCRC Building 10 Research Auditorium
Zoom: https://umich.zoom.us/j/99401963138?pwd=N1JoMnQzNmNFa3FYa0EwSnVFZk9kZz09 Passcode 842683
Chair: Dr. Cindy Chestek