Summary: Researchers have cut the power needs of neural interfaces while improving accuracy by focusing on a specific band of brain activity.
Source: University of Michigan
University of Michigan researchers discovered that tuning neural interfaces to a narrow band of brain activity dramatically lowers power consumption and increases decoding accuracy. This advance could help enable long-lasting, implantable brain–machine interfaces that both treat neurological disorders and allow precise control of prosthetics and external devices.
The research team, led by Cynthia Chestek, associate professor of biomedical engineering and core faculty at the Robotics Institute, estimates their method reduces power consumption by roughly 90% compared with current approaches.
“Today, translating brain signals into user intentions often requires massive computers and substantial electrical power—comparable to several car batteries,” said Samuel Nason, the study’s first author and a Ph.D. candidate in Chestek’s Cortical Neural Prosthetics Laboratory. “Cutting power needs by an order of magnitude is a critical step toward practical, at-home brain–machine interfaces.”
Neurons communicate through noisy electrical activity. Recording systems and electrodes are effectively listening to a noisy radio signal, and they must distinguish meaningful neural signals from background activity. That massive data stream drives high power and processing demands, which limits the safety and practicality of fully implanted devices.
In laboratory settings, researchers can use transcutaneous electrodes—wires that pass through the skin—to achieve high-performance decoding of complex behaviors, such as grasping. These setups often rely on about 100 electrodes capturing roughly 20,000 samples per second, enabling capabilities like restoring movement to paralyzed limbs or providing sensory feedback for prosthetic hands. However, transcutaneous systems are impractical for everyday use outside the lab and carry infection risks.
Some wireless implant designs use custom, application-specific integrated circuits to approach the performance of transcutaneous systems and can transmit on the order of 16,000 signals per second. Although promising, these bespoke chips face challenges in consistent operation and regulatory approval compared with industry-standard components.
To reduce both data and power demands, researchers traditionally compress neural signals using techniques such as threshold crossing rate (TCR). TCR detects when neural activity crosses a preset threshold, dramatically reducing the amount of data that must be processed. But TCR depends on continuously monitoring a wide bandwidth to detect crossings, and thresholds can vary across individuals and over time, requiring recalibration and additional hardware and power.
Chestek’s lab pursued a different compression strategy by targeting a specific feature of neural recordings they call spiking-band power (SBP). SBP isolates an integrated band of frequencies between 300 and 1,000 Hz that captures spikes from nearby single neurons. By listening only to this targeted frequency range—effectively sampling from a narrow straw rather than the full hose of brain activity—the researchers achieved accurate behavior prediction with far less data.
When compared with transcutaneous systems, SBP matched decoding accuracy while collecting just one-tenth the data: approximately 2,000 samples per second versus 20,000. Relative to threshold-based methods, SBP not only reduces raw data volume but also shows greater robustness to noise and does not require ongoing threshold tuning to maintain performance.
SBP also addresses a common limitation of neural implants: electrode degradation over time. Because SBP remains effective even when signals are weaker—permitting useful decoding at signal amplitudes that would be too low for other methods—implants could retain functionality for longer periods without invasive adjustments.
Beyond new implant designs optimized for SBP, the approach can extend the capabilities of many existing devices by lowering their technical requirements for translating neural activity into intentions. Devices that previously lacked the bandwidth or power for advanced decoding may now be viable for brain–machine interface applications.
“Many devices have been selling themselves short,” Nason said. “Using the same bandwidth and power, existing circuits can now support a broader set of brain–machine interfaces.”
The study, titled “A low-power band of neuronal spiking activity dominated by local single units improves the performance of brain–machine interfaces,” appears in Nature Biomedical Engineering.
About this neuroscience research article
Source:
University of Michigan
Media contacts:
Press Office – University of Michigan
Image source:
The image is in the public domain.
Original research:
“A low-power band of neuronal spiking activity dominated by local single units improves the performance of brain–machine interfaces.” Authors: Samuel R. Nason, Alex K. Vaskov, Matthew S. Willsey, Elissa J. Welle, Hyochan An, Philip P. Vu, Autumn J. Bullard, Chrono S. Nu, Jonathan C. Kao, Krishna V. Shenoy, Taekwang Jang, Hun-Seok Kim, David Blaauw, Parag G. Patil & Cynthia A. Chestek. Published in Nature Biomedical Engineering.
Abstract
A low-power band of neuronal spiking activity dominated by local single units improves the performance of brain–machine interfaces
High power requirements remain a major obstacle to clinical adoption of brain–machine interfaces. This study shows that the downsampled magnitude of the 300–1,000 Hz spiking band—referred to as spiking-band power (SBP)—can predict movement in basic behavioral tasks with performance comparable to the threshold crossing rate (TCR) sampled at 30 kilo-samples per second. Using simulations of neural recordings, the authors demonstrate that SBP is dominated by local single-unit spikes and offers spatial specificity similar to or better than TCR. SBP also correlates more strongly with the firing rates of lower signal-to-noise units than TCR. Experiments in non-human primates, including an online one-dimensional finger-group movement decoding task and an offline two-dimensional cursor-control task, show that SBP performs as well as or better than TCR. By enabling comparable decoding performance with substantially reduced bandwidth and power, SBP may facilitate more practical, long-lived neural implants for medical and assistive applications.