SpikeDeeptector: a deep-learning based method for detection of neural spiking activity.

Area of Science:

• Electrophysiology
• Neuroscience
• Machine Learning

Background:

• Microelectrodes are crucial for recording neural data, but channels often contain noise or no signal.
• Automatic identification of channels with neural data is significant for spike sorting and brain-computer interface (BCI) applications.
• Existing methods lack universal applicability across different recording technologies, subjects, and brain areas.

Purpose of the Study:

• To develop a universal algorithm for automatic identification and tracking of neural data channels.
• To overcome limitations of current methods in handling diverse recording conditions and data types.

Main Methods:

• Proposed SpikeDeeptector, a novel algorithm using deep learning for feature vector extraction.
• SpikeDeeptector constructs feature vectors from waveform batches for contextual learning.
• The deep learning model identifies patterns to classify channels containing neural spike data versus noise.

Main Results:

• Trained on data from one tetraplegic patient, SpikeDeeptector was evaluated on six epileptic patients and other tetraplegic patients.
• Achieved a cumulative evaluation accuracy of 97.20% on 1.56 million hand-labeled test inputs.
• Demonstrated generalization across different brain areas, subjects, and electrode types not used during training.

Conclusions:

• SpikeDeeptector successfully and universally identifies channels with neural data.
• The algorithm's generalization capabilities across diverse datasets signify a breakthrough in automated neural channel selection.
• This method holds significant potential for advancing online and offline spike sorting and BCI applications.