myojam

Real-time EMG gesture classification for assistive human-computer interaction - with an open education platform built on top of the research.

---

What it does

myojam reads surface electromyographic (EMG) signals from forearm muscles and classifies hand gestures in real time, enabling people with motor impairments to control a computer through muscle signals alone.

6 gesture classes: index flex · middle flex · ring flex · pinky flex · thumb flex · fist

Each gesture maps to a computer action: cursor left/right, scroll up/down, left click, spacebar.

Performance

Metric	Value
Cross-subject accuracy	84.85%
Training subjects	10 (Ninapro DB5)
EMG channels	16 surface electrodes
Feature vector	64-dimensional (MAV, RMS, ZC, WL x 16 ch)
Classifier	Random Forest (500 trees)
Inference latency	< 5 ms

Evaluated under leave-one-subject-out (LOSO) cross-validation.

Architecture

EMG Electrodes -> MyoWare 2.0 -> Arduino Uno -> USB Serial
                                                    |
                              FastAPI Backend (signal processing + inference)
                                                    |
                        React Web App  <->  macOS Desktop App

Stack

Layer	Technology
Signal processing	Python, SciPy, NumPy
ML model	scikit-learn Random Forest
Backend API	FastAPI (Render)
Web frontend	React, Vite, Three.js, OGL (WebGL)
Desktop app	PyQt6, Quartz (macOS)
Hardware	MyoWare 2.0, Arduino Uno R3

Interactive demos

All demos run in the browser - no hardware or install required.

Demo	Description
Live Signal	Real-time dataset playback with 3D hand model, confidence bars, and action log
Signal Guesser	5-round game: identify a gesture from real EMG waveforms before the AI does
Signal Playground	Draw a waveform with your mouse and watch MAV, RMS, ZC, WL update live
Gesture Game	Match target gestures against the clock, three difficulty levels
Confusion Explorer	Interactive confusion matrix from the 10-subject test set
Frequency Analyzer	Explore the spectral properties of EMG signals
Pipeline Explorer	Walk through each stage of the signal processing pipeline with real data
MyoCode	A Scratch-like block coding environment triggered by EMG gestures

Education platform

11 articles, 3 structured lesson plans (NGSS/AP/IB aligned), and an interactive tool suite for classrooms. No hardware required to teach the concepts.

• EMG Explainer - from motor cortex to gesture prediction
• Why EMG is Hard - noise, placement variability, individual anatomy
• Build Your Own - $68 hardware setup, full wiring guide
• Educator Hub - lesson plans, rubrics, curriculum alignment

Research

Four peer-structured papers published at myojam.com/research:

1. Main paper - classifier design, dataset, evaluation methodology
2. Classifier Analysis - Random Forest vs SVM vs KNN cross-subject comparison
3. Variability Review - sources of cross-subject accuracy loss
4. Windowing Analysis - window size vs accuracy vs latency trade-offs

Download

macOS desktop app available from Releases.

Requirements: macOS 12+, Python 3.13, cliclick (brew install cliclick)

pip3 install PyQt6 scipy scikit-learn numpy pyserial
chmod +x run.sh && ./run.sh

Grant Accessibility permission when prompted (System Settings -> Privacy & Security -> Accessibility).

Hardware setup

1. Place bipolar EMG electrodes on the forearm flexor muscle belly
2. Connect MyoWare 2.0 via Link Shield + TRS cable to Arduino Shield
3. Connect Arduino Uno to Mac via USB (laptop on battery power)
4. Upload arduino_sketch/emg_stream.ino via Arduino IDE
5. Click "Connect sensor" in the app

Train your own model

python3 collect_data.py    # record your gestures - 60 samples each
python3 train_my_model.py  # trains and saves a personalized model

Per-user calibration improves accuracy significantly. Cross-subject baseline is ~85%; within-session can exceed 95%.

Project structure

myojam/
|- frontend/                   # React web app (Vite + React Router)
|  |- src/
|  |  |- articles/             # long-form education articles
|  |  |- components/           # reusable UI components
|  |  \- educators/            # lesson plan pages
|  \- public/
|
|- backend/                    # FastAPI inference server (Render)
|  |- main.py
|  \- requirements.txt
|
|- desktop-app/                # PyQt6 macOS application
|  |- myojam.py                # main app
|  |- keywatcher.py            # global keyboard listener (Quartz)
|  |- collect_data.py          # personal EMG data collection
|  |- train_my_model.py        # model training pipeline
|  |- emg_classifier.py        # classifier logic
|  |- arduino_sketch/          # Arduino firmware (emg_stream.ino)
|  \- model/                   # trained model files (.pkl)
|
|- data/                       # Ninapro DB5 dataset (10 subjects)
|  \- DB5_s*/
|
|- model/                      # shared model files used by backend
|- scripts/                    # development & validation scripts
|- docs/                       # brand assets
\- vercel.json                 # SPA routing config for Vercel

Team

Built by Jaden W. with contributions from Matthew T. and Darren N.

License

MIT - see LICENSE for details.

---

© 2025 myojam™