ELVISort: encoding latent variables for instant sorting, an artificial intelligence-based end-to-end solution.

Area of Science:

• Computational Neuroscience
• Machine Learning for Neuroscience
• Neural Signal Processing

Background:

• Increasing channel counts in silicon-based neural probes necessitate automated signal processing for complex neural dynamics.
• Existing spike sorting tools face challenges in handling large datasets efficiently for real-time analysis.

Purpose of the Study:

• To develop a highly automated spike sorting model for high-performance online analysis.
• To achieve performance comparable to offline solutions while maintaining high efficiency.

Main Methods:

• Introduced ELVISort, a deep learning method for end-to-end detection and clustering of action potentials.
• Utilized deep learning frameworks to leverage GPU parallel processing capabilities.

Main Results:

• ELVISort demonstrated comparable performance to manual/semi-manual methods, with average F1 scores of 0.96, 0.82, and 0.81 on independent datasets.
• Achieved real-time processing, with computation time being only 1/15.71 of the actual data duration.
• Outperformed existing automatic spike sorting methods.

Conclusions:

• ELVISort offers a powerful, efficient, and real-time solution for spike sorting.
• Its end-to-end nature and parallel processing potential pave the way for integrated, portable, and high-performance neural data analysis systems.