CUSP: Complex spike sorting from multi-electrode array recordings with U-net sequence-to-sequence prediction.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning in Biology

Background:

• Complex spikes (CSs) in Purkinje cells are crucial but difficult to detect due to waveform variability and low frequency.
• Automated detection of CSs is challenged by recording artifacts like electrode drift and inherent signal complexity.

Purpose of the Study:

• To introduce CUSP (CS sorting via U-net Sequence Prediction), a novel deep learning framework for automated and accurate CS detection.
• To enable reconstruction of complete Purkinje cell activity by integrating CS and Simple Spike (SS) data.

Main Methods:

• CUSP utilizes a U-Net architecture with self-attention inception blocks for sequence-to-sequence prediction of CS events.
• The framework integrates local field potential and action potential signals for enhanced detection accuracy.
• Detected CS events are clustered and paired with concurrently detected SSs.

Main Results:

• CUSP achieves human-expert performance (F1 = 0.83 ± 0.03) in detecting CSs in macaque cerebellar recordings.
• The method successfully identifies CS events missed during manual annotation.
• CUSP demonstrates robustness against waveform variability, spikelet composition, and electrode drift, outperforming existing algorithms.

Conclusions:

• CUSP offers a scalable and robust solution for analyzing complex spike patterns in large-scale cerebellar and other neural datasets.
• The framework's generalizability extends to neural systems like hippocampal pyramidal cells where complex bursts are computationally significant.
• By automating expert-level accuracy, CUSP facilitates the study of information coding across neural circuits.