Published on 25 Aug 2020

A Smarter Way to Decode Brain Signals

Detecting and decoding brain signals is a crucial step in neuroscience research and neural prosthetics to help people with brain injuries and diseases. Researchers from the Nanyang Technological University (NTU) have found a way to improve the process.

Scientists from the NTU's School of Electrical and Electronic Engineering (EEE), working with other researchers, have developed a processing chip that uses less power to better detect brain signals and identify their neuron sources, a process called spike sorting. Their real-time processing, multi-channel chip is believed to be the first of its kind, and could lead to more compact neural implants.

Traditionally, for neuroscience research and neural implants, a recording chip is implanted into the brain and the detected neural signals are transmitted to a nearby computer for sorting. This approach has limitations due to the high data rates and power consumption.

On-chip spike sorters are more power-efficient and have shorter lag times, which is essential for real-time, multi-channel neural signal processing.

The EEE-led scientists created an algorithm for their on-chip sorter that strengthens the brain signals and filters out noisy data to better detect the relevant neural spikes.

Their invention also uses a feature extractor and dimensionality reduction to cut down on the information needed to classify the neural spikes. This means that the chip can use a smaller and less power-hungry SRAM memory component.

“We can make the chip very compact because we’ve used the new algorithm to significantly reduce the power needed and the amount of data to be processed or sent outside of the brain,” said Professor Tony Tae-Hyoung Kim from EEE.

The EEE-led scientists also developed another algorithm that makes their chip “smarter” than its commercial counterparts. Traditionally, scientists have to input the mean values of neural clusters, with each cluster representing one neuron source, into a spike sorter’s training engine component.

This means that the sorters can classify only pre-identified and programmed neural clusters.

The EEE-invented algorithm removes such limitations. It forms a new cluster for new data when the weighted distance of the new cluster to the existing clusters is larger than that between the existing ones.

The scientists fabricated a 0.414 millimetre square prototype of their chip, and verified its functionality using datasets. Prof Kim said: “To our knowledge, this is the first real-time, multi-channel spike-sorting chip that includes spike detection and alignment, feature extraction, dimensionality reduction and clustering. The power efficiency is improved more than twenty times when compared to other state-of-the-art spike-sorting chips.”