EMG Gesture Classification Project

Multi-dataset EMG analysis for hand gesture recognition using GRABMyo (32ch) and NinaPro DB2 (12ch).

Datasets

GRABMyo: 32 channels, 43 subjects, 17 gestures, 3 sessions NinaPro DB2: 12 channels, 40 subjects, 49 movements

Data Acquisition

Please run the following shell scripts to download all (or parts) of the data. For the GRABMyo, you need to download at least 6 participants from Session 1.

# Download GRABMyo (interactive prompts for sessions/participants)
./data-installers/install-GRABMyo.sh

# Download NinaPro DB2 (interactive prompts for subjects)
./data-installers/install-ninapro-db2.sh

Data Exploration & Cleaning

• Signal quality validation (range, NaN/Inf checks)
• Channel variability analysis across subjects
• Cross-dataset gesture separability testing (diagnostic-analysis.ipynb)

Preprocessing Pipeline

1. Per-channel normalization: (signal - mean) / std
2. Bandpass filter: 20-450 Hz, 4th order Butterworth
3. Sliding windows: 409 samples, 50% overlap
4. Feature extraction (8 features/channel):
   • Time-domain: Mean, STD, RMS, MAV
   • Frequency-domain: MNF, MDF, PSD Mean, Peak Frequency

Labeling

• GRABMyo: 17 gesture classes (1-17)
• NinaPro DB2: 49 movement classes + rest
• Trial-level labels for Leave-One-Trial-Out validation

Baseline Results

Method	Channels	Accuracy	Validation
Random Forest	32 (GRABMyo)	58.0% ± 3.0%	LOTO CV
Random Forest	12 (reduced)	51.8% ± 3.6%	LOTO CV

Channel Reduction: 12 selected channels retain 89% of baseline accuracy with 62.5% fewer features. Selected channels: 2, 3, 5, 6, 10, 13, 16, 23, 28, 29, 30, 31

See channel-reduction-analysis.ipynb for methodology and diagnostic-analysis.ipynb for feature validation.

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt