KIASORT: Knowledge-Integrated Automated Spike Sorting for Geometry-Free Neuron Tracking
Identifying single units from extracellularly recorded neural signals is critical for understanding brain circuit dynamics. With the advancements of large-scale recordings, efficient and precise automated spike sorting methods have become essential. Existing approaches face challenges with channel quality variability, neuron-specific waveform drifts, and nonlinear changes in spike shapes that depend on neuronal morphology and electrode proximity. We introduce KIASORT (Knowledge-Integrated Automated Spike Sorting), which integrates knowledge from channel-specific classifiers trained on clustered spike waveforms to sort the data. KIASORT evaluates channel quality, automatically excludes noisy recordings, and identifies spike classes using a hybrid dimensionality reduction approach combining linear with nonlinear embeddings. To validate our approach against existing methods, we developed biophysical simulations demonstrating that conventional one-dimensional drift correction methods cannot address heterogeneous neuron-specific drift and nonlinear waveform changes. KIASORT's geometry-free, per-neuron tracking approach overcomes these limitations without assumptions about cluster shape or temporal stability. Specifically, KIASORT identified significantly greater number of high-quality units than Kilosort4 in ground-truth simulations with neuron-specific drift, while maintaining real-time processing capability. Complementing these advances, KIASORT also includes a unified graphical interface for data inspection, sorting, and curation, which is freely available online through link https://github.com/banaiek/KIASORT.