A neural network for online spike classification that improves decoding accuracy

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Separating neural signals from noise is crucial for improving brain-computer interface (BCI) performance and stability.
• Existing spike-sorting algorithms often lack real-time applicability and show variable effects on decoding accuracy.

Purpose of the Study:

• To automate the process of distinguishing neural spikes from noise in real time.
• To develop a tunable artificial neural network (ANN) for precise spike classification, enhancing online decoding.

Main Methods:

• Trained an ANN with a hidden layer on hand-labeled neural waveforms (spikes vs. noise).
• Utilized a likelihood metric and tunable stringency threshold for waveform classification.
• Applied the ANN to decode remembered target locations in a memory-guided saccade task using prefrontal cortex recordings from rhesus macaques.

Main Results:

• The ANN classified neural waveforms in real time, producing results comparable to human spike-sorters.
• Excluding low-likelihood waveforms significantly improved decoding performance compared to traditional threshold crossing methods.
• The benefits of the ANN classifier increased over time since electrode array implantation.

Conclusions:

• The developed ANN classifier is a feasible, low-risk preprocessing step for both offline and online neural data analysis.
• This automated approach enhances the reliability and performance of brain-computer interfaces by effectively separating neural signals from noise.