Deep learning based decoding of single local field potential events.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Traditional analysis averages neural signals, losing crucial single-trial information.
• The brain functions as a single-trial processor, necessitating new analytical methods.
• Extracting meaningful patterns from electrophysiological recordings at the single-trial level remains a challenge.

Purpose of the Study:

• To develop and validate an unsupervised machine learning approach for analyzing single-trial electrophysiological recordings.
• To extract interpretable clusters of neural activity patterns from local field potential (LFP) events.
• To determine the direction of information flux in the cerebral cortex using LFP signal shapes.

Main Methods:

• Utilized an auto-encoder network for dimensionality reduction of single local field potential (LFP) events.
• Clustered LFP events to identify distinct neural activity patterns.
• Applied the method to rodent extracellular neural recordings and human intracranial EEG data.

Main Results:

• Demonstrated that unsupervised machine learning can extract meaningful information from single-trial electrophysiological data.
• Identified that specific LFP shapes correlate with latency differences across recording channels, indicating information flux direction.
• Showed that spontaneous LFP event shapes mirror those observed during stimulus-evoked activity.

Conclusions:

• Unsupervised machine learning offers a powerful tool for analyzing single-trial neural data, overcoming limitations of traditional averaging methods.
• LFP event shapes provide insights into the directionality of information processing in the cerebral cortex.
• Single-channel LFP event shapes during spontaneous activity are representative of stimulus-evoked patterns, extending previous findings.