Multichannel electrophysiological spike sorting via joint dictionary learning and mixture modeling.

Area of Science:

• Computational Neuroscience
• Machine Learning for Electrophysiology

Background:

• Accurate analysis of multichannel extracellular electrophysiological data is crucial for understanding neural function.
• Existing methods for action potential detection and classification (sorting) face challenges with data variability and sparsely firing neurons.

Purpose of the Study:

• To develop an advanced methodology for joint feature learning and clustering of electrophysiological data.
• To improve action potential detection and classification across multiple recording periods, especially for chronic experiments.

Main Methods:

• A novel "focused mixture model" (FMM) for joint feature learning and clustering.
• Bayesian methodology to handle missing data and share information across channels.
• Direct modeling of spike rate for improved detection of sparsely firing neurons.

Main Results:

• Enhanced ability to distinguish single-unit spikes from artifacts by sharing information across channels.
• Effective handling of neural units appearing, disappearing, or reappearing over multiple recording days.
• State-of-the-art performance achieved without manual hyperparameter tuning on public and experimental datasets.

Conclusions:

• The proposed methodology offers significant improvements in analyzing complex electrophysiological data.
• The approach is robust for chronic experiments and addresses limitations of previous two-stage learning methods.
• This work advances the state-of-the-art in neural spike sorting and analysis.