Application of deep reinforcement learning for spike sorting under multi-class imbalance

Area of Science:

• Computational Neuroscience
• Machine Learning
• Signal Processing

Background:

• Spike sorting is crucial for analyzing neural activity from high-density microelectrode arrays.
• Simultaneous recording leads to multi-class imbalance issues like overlapping spikes and varying firing rates.

Purpose of the Study:

• To develop a deep reinforcement learning (DRL) method, ImbSorter, for effective spike sorting under multi-class imbalance.
• To enhance sensitivity to minor neuronal classes using a dynamic reward function (DRF).

Main Methods:

• Spike sorting framed as a Markov decision process.
• A dynamic reward function (DRF) incorporating inter-class imbalance ratios guides the DRL agent.
• Evaluation on Wave_Clus and macaque datasets with overlapping spikes and multi-scale imbalance.

Main Results:

• ImbSorter demonstrated improved Macro_F1 scores compared to classical DRL, traditional ML, and advanced techniques.
• The method showed robustness against overlapping spikes and noise interference.
• High stability and promising performance on datasets with skewed neuronal firing distributions.

Conclusions:

• ImbSorter offers a promising solution for spike sorting in the presence of significant class imbalance.
• The DRL approach with DRF effectively handles complex neural data challenges.
• This method advances the analysis of neural firing patterns in neuroscience research.