A Bayesian clustering method for tracking neural signals over successive intervals.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Signal Processing

Background:

• Accurate sorting and tracking of individual neuron action potentials are crucial for understanding neural circuits.
• Existing methods often struggle with signal nonstationarity and changing neuron populations in multiunit extracellular recordings.

Purpose of the Study:

• To introduce a new unsupervised method for sorting and tracking neuronal action potentials.
• To improve the consistency and accuracy of neural signal classification in complex recordings.

Main Methods:

• Developed an unsupervised method extending traditional mixture models using Bayesian inference for sequential data.
• Incorporated clustering results from preceding intervals to handle signal nonstationarity and dynamic neuron numbers.
• Utilized prior data for seeding clustering algorithms and model selection.
• Applied the method in a principal components space, adaptable to any feature space with Gaussian spike distributions.

Main Results:

• The proposed method demonstrated significantly more consistent clustering and tracking of neurons compared to expectation-maximization-based mixture models.
• Successfully applied to recordings from macaque parietal cortex, validating its effectiveness.
• The technique naturally facilitates matching signal clusters over time for neuron tracking.

Conclusions:

• The new unsupervised method offers a more robust and consistent approach to sorting and tracking neuronal action potentials.
• This improved tracking capability is vital for advanced neuroscience research applications.
• The algorithm's flexibility allows application across various feature spaces and recording conditions.