EMUsort

EMUsort sorts motor-unit action potentials (MUAPs) from high-density intramuscular arrays such as Myomatrix. It is a fork of Kilosort4, adapted to what makes muscle recordings hard: the same unit reaches channels at different times (conduction along the fibres), waveforms are complex and polyphasic, and units overlap heavily.

In dsp-kitchen it lives in the dspk-sorter-kilosort crate (feature emusort). It is written from the paper and the published defaults, with upstream’s code read to match its behaviour and no code copied, and is GPL-3.0 like upstream. Because it is a fork, it reuses every Kilosort4 stage with its own settings and adds two stages.

What EMUsort changes in Kilosort4

Change	Why (paper)
Channel-delay removal before detection	a MUAP reaches channels up to a few ms apart
9 temporal PCs and 9 universal templates (Kilosort4: 6 and 6)	more complex waveforms
Several clip thresholds [6, 9, 12, 15] (Kilosort4: one, 6)	later-recruited, larger units were missed by one threshold
HDBSCAN outlier removal before the templates are clustered	movement artifacts contaminate the templates
No common average reference	a MUAP spans many channels; subtracting the average removes signal
Template-learning stride nskip 2 (Kilosort4: 25)	more MUAP waveforms for the templates
Clips scaled by one common factor (Kilosort4: each to unit norm)	keeps relative amplitudes, so the templates separate MUAP sizes

What is implemented

Stage	Status
Whole-recording run (Emusort::run, dspk sort), Kilosort4’s runner with EMUsort’s plan	implemented
Channel-delay estimation and removal	implemented (device)
Universal templates learned with HDBSCAN outlier removal and several thresholds	implemented (device)
Universal-template detection and features	implemented (device, shared with Kilosort4)
Clustering, learned templates, learned-template matching, clustering of the matches	implemented (device, shared with Kilosort4)
Refractory (CCG) criteria, global merges, duplicate-spike removal	implemented (shared with Kilosort4)
EMUsort’s unit score (cluster_score_threshold)	not yet

EMUsort learns its universal templates from every recording; it publishes no weight or template files.

Citation

O’Connell S., Michaels J. A., Wang R., Mamidipaka S., Venkatesh M., Aresh N., Pachitariu M., Pruszynski J. A., Sober S. J., Pandarinath C. High performance sorting of motor unit action potentials with EMUsort. eLife 15:RP110417 (2026), reviewed preprint, CC-BY 4.0. doi:10.7554/eLife.110417.1. Preprint: bioRxiv, doi:10.64898/2026.01.06.697952.

@article{oconnell2026emusort,
  title   = {High performance sorting of motor unit action potentials with {EMUsort}},
  author  = {O'Connell, Sean and Michaels, Jonathan A and Wang, Runming and Mamidipaka, Sahit and
             Venkatesh, Manikandan and Aresh, Nevin and Pachitariu, Marius and Pruszynski, J Andrew
             and Sober, Samuel J and Pandarinath, Chethan},
  journal = {eLife},
  year    = {2026},
  volume  = {15},
  pages   = {RP110417},
  doi     = {10.7554/eLife.110417.1}
}

Code	snel-repo/EMUsort (GPL-3.0, commit a06bb60) and its Kilosort4 fork snel-repo/Kilosort4 (GPL-3.0, compared against Kilosort4 v4.0.18)
Arrays	Myomatrix: Chung et al., Myomatrix arrays for high-definition muscle recording, eLife (2023), doi:10.7554/eLife.88551

Run it

dspk sort emusort recording.bin -o sorted --format phy

Every setting and its default: EMUsort settings. How it works: pipeline. What to change: tuning.