Noise-robust unsupervised spike sorting based on discriminative subspace learning with outlier handling

Area of Science:

• Neuroscience
• Computational Neuroscience
• Signal Processing

Background:

• Spike sorting is crucial for analyzing neural activity from spike trains.
• Conventional methods often fail to find the most discriminative features, leading to poor accuracy, especially in noisy conditions.
• Existing techniques like PCA may not effectively separate overlapping neuronal clusters.

Purpose of the Study:

• To develop a noise-robust, unsupervised spike sorting algorithm.
• To improve the accuracy and reliability of neuronal cluster detection.
• To facilitate detailed analysis of single- and multi-unit activities.

Main Methods:

• Utilizes discriminative subspace learning to extract low-dimensional features from spike waveforms.
• Employs iterative subspace selection with Linear Discriminant Analysis (LDA).
• Performs clustering using Gaussian Mixture Models (GMM) with outlier detection and an automated cluster number detection method.

Main Results:

• Demonstrates substantially improved cluster distinction on simulated and real in vivo datasets.
• Achieves higher sorting accuracy compared to conventional methods like PCA and wavelets.
• Successfully detects highly overlapping and previously undetectable clusters.

Conclusions:

• The proposed algorithm offers high robustness to neural noise and outliers.
• Enables more accurate analysis of neuronal activity for neuroscience and brain-machine interface applications.
• Provides a more reliable tool for understanding neural coding and brain function.