Cross-User Zero-Shot EMG Gesture Classification on EPN612

End-to-end benchmark for user-invariant EMG gesture classification. Models train on a population of users and are evaluated zero-shot on held-out subjects with no per-user calibration. This is the precondition for any EMG interface that ships beyond the lab, since per-user calibration breaks under electrode shift, sweat, fatigue, and time.

Dataset

EPN-612: 612 subjects, 5 hand gestures plus rest, 8-channel Myo armband at 200 Hz.

Methods

• Architectures: LDA, MLP, CNN, LSTM, Transformer
• Input representations: raw windows, segmented windows, hand-crafted feature sets
• Cross-user training with embedding-based variants (metric learning, gradient reversal)
• Within-user baselines and fine-tuning ablations (standard, segmented, raw)
• Incremental user adaptation experiments

Results

Best cross-user zero-shot accuracy 80% on held-out subjects with no calibration. Within-user upper bound 85%.

Repo layout

• process_epn612.py, EPN612.py: dataset ingestion and windowing
• models.py, utils.py: shared model and training utilities
• cross_*.py: cross-user training per architecture
• within_*.py: within-user baselines and fine-tuning ablations
• increment_cnn_raw.py, inc_rank_cnn_raw.py: incremental adaptation
• Analysis.ipynb: aggregated results and figures

Author

Amir Hariri, Institute of Biomedical Engineering, University of New Brunswick.