MSEMG

This repository provides the source code for the implementation of "MSEMG: Surface Electromyography Denoising with a Mamba-based Efficient Network".

We developed this repo with python=3.10 and pytorch=1.12. You can reproduce our experiment by following the steps:

• Construct The Environment
• Prepare The Dataset
• Run Training
• Run Testing

Construct The Environment

Clone the repository

Clone our repository by running the following command and entering the directory. This will be your working directory.

git clone https://github.com/yt-tony-liu/MSEMG.git
cd MSEMG

Install the Python packages

pip install -r requirements.txt

Prepare The Dataset

Download the sEMG signals

We use the surface electromyography signals from Non-Invasive Adaptive Prosthetics (NINAPro) DB2. The website doesn't provide an all-in-one compressed file of the database, so you might need to download data points from every subjects separately and unpack them into 1 folder shown as follows:

[EMG_corpus_dir]
|-- DB2_s1
|   |-- S1_E1_A1.mat
|   |-- S1_E2_A1.mat
|   `-- S1_E3_A1.mat
|-- DB2_s2
|   |-- S2_E1_A1.mat
|   |-- S2_E2_A1.mat
|   `-- S2_E3_A1.mat
...

Download the ECG signals

The ECG signals for simulating interference can be downloaded from MIT-BIH Normal Sinus Rhythm Database. You should specify the path to the folder containing all signals for the later implementation.

[ECG_corpus_dir]
|-- 16265.atr
|-- 16265.dat
|-- 16265.hea
|-- 16265.xws
|-- 16272.atr
|-- 16272.dat
|-- 16272.hea
|-- 16272.xws
...

Prepare the config file

You can specify the paths to the corresponding directories in the data_cfg.yaml in /cfg. Note that the ECG_storage_dir is the directory you would like to store the process ECG signals, the sEMG_dataset_dir is the directory to the dataset for later use, and the result_dir is the directory for your experiment results.

EMG_corpus_dir: [directory to downloaded NINAPro DB2]
ECG_corpus_dir: [directory to downloaded MIT-BIH NSRD]
ECG_storage_dir: [directory to store ECG for simulating interference]
sEMG_dataset_dir: [sEMG dataset directory]
result_dir: [your result directory]

Data preprocessing

To preprocess all downloaded data and prepare the dataset, you need to run the following after you have specified all paths in the local_cfg.yaml.

python preprocess.py

Run Training

You need to train SDEMG before you perform denoising. You can train SDEMG on a single GPU or multiple GPUs. Note that you can specify the number of workers in data_cfg.yaml, the recommended number is half of your CPU cores on your device. Also, adjust the experiment setting and the hyperparameters (e.g. batch size) in cfg/default.yaml

Single GPU Training

python main.py --train

Multiple GPU Training

SDEMG supports multiple GPU training with Hugging Face Accelerate. You need to install accelerate by running pip install accelerate and adjust the configuration for your device by running accelerate config.

accelerate launch main.py --train

Run Testing

You can run testing and find the denoise result in your project directory.

python main.py --test

Inference

You can fill in the path to the files you would like to run inference in line 72 of main.py.

72      file_paths = ['demo file paths']

SDEMG will run a single inference on the files.

python main.py --sample

Citing our work

Y. -T. Liu, K. -C. Wang, R. Chao, S. M. Siniscalchi, P. -C. Yeh and Y. Tsao, "MSEMG: Surface Electromyography Denoising with a Mamba-based Efficient Network," ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Hyderabad, India, 2025, pp. 1-5, doi: 10.1109/ICASSP49660.2025.10887547.