Bypassing spike sorting: Density-based decoding using spike localization from dense multielectrode probes.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Neural decoding is crucial for brain-computer interfaces (BCI) and understanding neural activity-behavior links.
• Spike sorting, assigning action potentials to neurons, is a key but often inaccurate step.
• Current methods discard valuable information by not modeling spike assignment uncertainty.

Purpose of the Study:

• To develop a spike sorting-free decoding method for improved BCI performance.
• To directly model neural activity using spike features and account for assignment uncertainty.
• To enable dynamic decoding that adapts to behavioral changes.

Main Methods:

• Proposed a novel decoding method using a mixture of Gaussians (MoG) to model spike feature distributions.
• Integrated time-varying mixing proportions in the MoG to capture behavioral responses.
• Employed variational inference for model fitting and decoding, bypassing explicit spike clustering.
• Utilized advanced computational methods and high-density probes (e.g., Neuropixels) for feature extraction.

Main Results:

• The spike sorting-free decoder consistently outperformed thresholding (multi-unit activity) and traditional spike sorting methods.
• Demonstrated robust performance across diverse animal recordings and probe configurations.
• Successfully decoded behavioral correlates directly from spike features with uncertainty modeling.

Conclusions:

• The proposed spike sorting-free decoding method offers a more effective alternative to current BCI approaches.
• Directly modeling spike features and their uncertainties enhances neural decoding accuracy.
• This approach unlocks the potential of rich spike features from high-density probes for BCI applications.