BCIwiki (bciwiki.com) — An algorithm called SpikeCleaner can label the quality of neural units with 97% accuracy, reducing the need for manual curation. The algorithm was developed by researchers at the University of Michigan, Ann Arbor, and uploaded to the bioRxiv preprint server on July 30, 2026; it has not yet been peer reviewed.
SpikeCleaner aims to standardize the labeling of unit quality after automated spike sorting. It combines spike rate, spike timing metrics (from autocorrelogram), and waveform-based features (such as peak amplitude, slopes, half-width, and inter-channel correlation) to classify units as good, noise, or multi-unit activity. These labels can be imported directly into Phy, serving as curation aids rather than absolute classifications.
The researchers benchmarked the algorithm against labels curated by two expert users across three recordings. SpikeCleaner achieved an average of 97% accuracy and 92% F1 score in classifying single units; 97% accuracy and 92% F1 score in full-category agreement (single unit, multi-unit activity, noise); and 97% accuracy and 95% F1 score in distinguishing neuronal versus non-neuronal units.
The algorithm addresses the challenge of high-density probes like Neuropixels, where an hour of data can exceed 80 GB, making manual curation no longer scalable. By automating standard criteria, SpikeCleaner aims to speed data curation and ensure dataset quality.