Motor Unit Mode Network (MMNet)

This repository provides an implementation of the Motor-unit Mode Network (MMNet) for extracting low-dimensional motor-unit modes from motor-unit discharge-rate signals.

MMNet models coordinated activity across multiple motor units using a variational autoencoder architecture to identify latent patterns of neural drive underlying muscle activation.

---

Repository Contents

• mmnet.py – main script for training the MMNet model
• mmnet_demo.ipynb – simple notebook demonstrating how to run the model and inspect latent modes

---

Input Data

MMNet operates on motor unit discharge-rate signals obtained from high-density surface EMG (HD-sEMG).

Typical preprocessing pipeline:

1. Record HD-sEMG signals
2. Decompose EMG signals to obtain motor unit spike trains
3. Convert spike trains into continuous discharge-rate signals using smoothing
4. Provide the discharge-rate signals as input to MMNet

The model expects a MATLAB .mat file containing a matrix of motor unit discharge-rate signals (default variable name: concatenated_data) with shape: [n_mus, n_samples]

where:

• n_mus = number of motor units
• n_samples = number of time samples

Each row corresponds to one motor unit discharge-rate signal.

---

Example Command

python scripts/mmnet.py \
--mat_path All_MUs_S01s35bef.mat \
--latent_dim 4 \
--out_dir outputs_S01 \
--save_full_recon \
--save_full_latent \
--save_config

Citation

If you use MMNet in your research, please cite:

Kamankesh, A. (2026).
Motor-Unit Mode Network (MMNet).
Zenodo. https://doi.org/10.5281/zenodo.18895984

BibTeX:

@software{kamankesh_mmnet_2026,
  author  = {Kamankesh, Alireza},
  title   = {Motor-Unit Mode Network (MMNet)},
  year    = {2026},
  doi     = {10.5281/zenodo.18895984},
  url     = {https://github.com/AKamankesh96/Motor-Unit-Network---MMNet}
}