Accurate, Very Low Computational Complexity Spike Sorting Using Unsupervised Matched Subspace Learning.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Computational Neuroscience

Background:

• Spike sorting is crucial for analyzing neural signals from implantable devices.
• Existing methods often face challenges with computational complexity and power consumption.
• Adaptable feature extraction is needed for real-time processing within strict device constraints.

Purpose of the Study:

• To develop an adaptable, dictionary-based feature extraction approach for spike sorting.
• To ensure high accuracy and low computational complexity for implantable applications.
• To create a system compatible with limited chip area and power budgets.

Main Methods:

• Employed matched unsupervised subspace filtering for feature extraction from evolving subspaces.
• Utilized a dictionary with {-1, 0, 1} arrays, requiring only addition and subtraction.
• Developed a neural signal simulator for performance evaluation against existing systems.

Main Results:

• All three dictionary-based feature extractors demonstrated robust performance with <8% average classification error.
• Successfully classified up to 6 clusters reliably, a first for implantable applications.
• An ASIC implementation occupied 0.09 mm² and consumed 10.48 µW at 1V, 30 kHz.

Conclusions:

• The proposed adaptive feature extractors offer a viable solution for accurate spike sorting in resource-constrained implantable devices.
• The method achieves high classification accuracy and low power consumption.
• This technology enables advanced neural signal processing for future implantable systems.