Cross-Subject EMG Gesture Recognition for Precision Pinch Grasps
Banks II, Jason James
Banks II, Jason James
Citations
Altmetric:
Author
Supervisor
Department
Computer Vision
Embargo End Date
Type
Thesis
Date
2026
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Surface electromyography provides a non-invasive means of decoding motor intent from the forearm musculature, with applications spanning prosthetic control, surgical assis tance, and rehabilitation. While within-subject gesture classification is effective, gesture and contact detection performance drops when applied to cross-subject data. Inter-subject variability in anatomy, subcutaneous tissue, motor unit recruitment, and electrode place ment is a common factor that hinders performance. This cross-subject generalization gap is the central obstacle to deploying calibration-free EMG interfaces in clinical and surgical settings, where per-session individual calibration is often impractical. This thesis benchmarks 10 architectures and 3 normalization strategies under Leave One-Subject-Out evaluation across 12 subjects performing 6 precision pinch gestures across 3 wrist orientations. Three subjects (S09, S11, S12) had essential tremors, one (S08) had Erb’s palsy, and one (S02) had a distal radius fracture from a car accident that required open reduction and internal fixation with a titanium surgical plate. A 1D CNN with per channel normalization reached 50.6% accuracy and 78.4% macro area-under-curve (AUC) without subject-specific calibration, a 13.7 percentage point improvement over the MyoGestic CatBoost baseline reproduced on the same data. Normalization strategy affected deep learning cross-subject performance, while classical ML was relatively unaffected. Essential tremor subjects clustered at the bottom of the cross-subject distribution, while subjects with stable neuromuscular conditions generalized comparably to healthy individuals. These findings are applied to a real-world surgical context, where cross-subject contact detection and gesture prediction are evaluated during simulated coronary intervention on an unseen test subject. Binary contact detection reached 70.32% accuracy on a held-out surgical subject using a model trained on dynamometer recordings, showing the feasibility of calibration-free instrument contact monitoring during catheter-based procedures.
Citation
Banks II, Jason James, "Cross-Subject EMG Gesture Recognition for Precision Pinch Grasps," M.S. Thesis, Computer Vision, MBZUAI, 2026.
Source
Conference
Keywords
Hand Gesture Recognition, Surface Electromyography, Cross-Subject, Dynamometer, Hand Contact Detection, Signal Processing