Validation of neural spike sorting algorithms without ground-truth information

Area of Science:

• Neuroscience
• Computational Neuroscience
• Data Science

Background:

• High-throughput electrophysiological recordings generate vast amounts of neural data.
• Numerous spike sorting algorithms exist to extract neural firing events.
• Standardized, automated quality evaluation of spike sorting is urgently needed.

Purpose of the Study:

• Introduce novel validation metrics for automatic spike sorting algorithms.
• Assess the credibility and reliability of spike sorting outputs.
• Provide a standardized method for evaluating algorithm performance.

Main Methods:

• Develop a suite of validation metrics based on algorithm stability.
• Rerun spike sorting algorithms multiple times to measure stability under perturbations.
• Make minimal assumptions about algorithm internals or noise characteristics.

Main Results:

• Demonstrate new metrics on various spike sorting algorithms and datasets (in vivo, ex vivo, overlapping spikes).
• Compare new metrics against existing quality measures and ground-truth accuracy.
• Provide a software implementation for the developed metrics.

Conclusions:

• Stability is crucial for the reproducibility of electrophysiological recording results.
• New metrics can significantly reduce manual validation effort in large-scale automated spike sorting.
• These metrics should be integral to systematic benchmarking of spike sorting algorithms.