MyoGestic

Real-time biosignal experiment GUI builder. A compact Python framework that turns a short script into a live experiment - signal viewers, recording, training, prediction - without classes, registries, or config files.

Built at the n-squared lab at Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) for the myocontrol research community. v2 is a ground-up rewrite of MyoGestic v1 focused on small, composable API surfaces and live extensibility.

Provides: live LSL ingest, on-disk recording (Zarr → .session.zip), ML pipeline lifecycle (train/predict on their own threads), Dear ImGui widgets, output filters, gRPC + LSL dual-plane integration with the Virtual Hand Interface.

Does not provide: DSP, ML models, feature extraction. You bring scipy, MyoVerse, CatBoost, PyTorch - whatever fits.

Try it in your browser

A live MyoGestic app runs entirely in your browser via Pyodide at https://nsquaredlab.github.io/MyoGestic/playground/. Synthetic EMG, in-memory recording, sklearn LDA training, live prediction. No install.

Install

uv sync                      # core dependencies only
uv sync --extra examples     # + catboost, myoverse, torch, scikit-learn (to run the demos)
uv sync --extra dev          # + pytest, ruff, the examples extras above

Optional extras: [brainflow] [bdi] [serial] [grpc] [zarrs] [docs]. [grpc] pulls in grpcio + grpcio-tools + protobuf for the VHI control plane.

Quick start

from myogestic import App, Stream
from myogestic.sources import LSLSource
from myogestic.widgets import recording_controls, signal_viewer

app = App("Hello EMG")
app.streams(Stream("emg", source=LSLSource("EMG"), window_seconds=1.0))

@app.ui
def ui(ctx):
    signal_viewer(ctx, "emg")
    recording_controls(ctx, ["Rest", "Fist"],
                       on_record=app.start_recording,
                       on_stop=app.stop_recording)

app.run()

That's the whole loop. Add a Pipeline, decorate extract / train / predict, and you have a closed-loop experiment.

Six runnable end-to-end demos live in examples/synthetic/:

• emg_classification.py - the canonical first read (CatBoost binary)
• emg_classification_grpc.py - adds the VHI gRPC control plane
• emg_regression.py - continuous 5-DoF regression
• emg_regression_raulnet.py - same flow with a PyTorch Lightning CNN
• emg_32ch_multi_model.py - selectable classifier + Save/Load
• emg_popout_layout.py - the same flow in a dockable tear-off layout

Documentation

The full docs live as a ProperDocs site under docs/ - tutorials, how-to guides, concept explanations, an auto-generated API reference, and the in-browser playground.

Build and serve locally:

uv sync --extra docs --extra grpc --extra serial
uv run properdocs serve

Then open http://127.0.0.1:8000/MyoGestic/.

Quick links into the source:

• Getting Started - install + run the synthetic-EMG demo.
• Tutorials - emg-classification, emg-regression-with-vhi.
• How-to guides - recipes (custom source, custom widget, custom model, integrate the Virtual Hand, install VHI, the contrib feature set, ...).
• Concepts - architecture, streams, pipeline, threading, recording, the Px/Fr grid, the EdgeTrigger pattern.
• API reference - auto-generated from docstrings.
• API cheatsheet - every public symbol on one page.
• Playground - the in-browser demo (no install).

The docs are deployed to GitHub Pages via .github/workflows/docs.yml on every push to main.

Development

uv sync --extra dev
uv run pytest -q
uv run ruff check .

Some integration tests need a live LSL outlet (port-bind sensitive in sandboxed CI). A handful of test files lag recent refactors and are scheduled for cleanup.

How to cite

If you use MyoGestic in your research, please cite our Science Advances paper:

@article{Simpetru2025,
    author  = {Raul C. S{\^i}mpetru and Dominik I. Braun and Arndt U. Simon
               and Michael M{\"a}rz and Vlad Cnejevici
               and Daniela Souza de Oliveira and Nico Weber and Jonas Walter
               and J{\"o}rg Franke and Daniel H{\"o}glinger and Cosima Prahm
               and Matthias Ponfick and Alessandro Del Vecchio},
    title   = {MyoGestic: EMG interfacing framework for decoding multiple
               spared motor dimensions in individuals with neural lesions},
    journal = {Science Advances},
    volume  = {11},
    number  = {15},
    pages   = {eads9150},
    year    = {2025},
    doi     = {10.1126/sciadv.ads9150},
    url     = {https://www.science.org/doi/abs/10.1126/sciadv.ads9150},
}

License

MyoGestic is licensed under the GNU General Public License v3.0 (GPL-3.0), matching the v1 release. Derivative work must remain open under the same license.