• Citer

Linear non-conservative unsupervised domain adaptation for cross-subject EMG gesture recognition

par Colot, Martin;Simar, Cédric;Cebolla Alvarez, Ana Maria ;Bontempi, Gianluca
Référence Biomedical Signal Processing and Control, 111, 108283
Publication Publié, 2026-01-01

Article révisé par les pairs

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

Electromyography (EMG) offers a natural and non-invasive interface for human–computer interaction, and machine learning (ML) is increasingly used to recognize gestures from recorded EMG signals. However, the statistical distributions of EMG show high inter-person variability, making cross-subject interpretation unreliable. Calibration phases with guided exercises are currently used in commercial devices to address this issue. Modern solutions rely on unsupervised domain adaptation (UDA) methods which align the sample distributions of different subjects without requiring tedious calibration exercises. However, when a real concept shift occurs (e.g. a similar signal pattern corresponds to different gestures in different subjects), these alignment methods are inefficient. This paper presents a novel linear and shallow UDA method based on Linear Discriminant Analysis and K-Means to address cross-subject calibration. To better handle real concept shifts, this parsimonious and easy-to-use method is non-conservative, meaning that it relies only on target samples and initial pseudo-labels rather than performing domain alignment. We perform in-depth evaluation against state-of-the-art deep-learning methods across multiple datasets and feature extraction pipelines and emulate a realistic system using streamed EMG to show the significant potential of our method. Our findings show that our approach significantly improves cross-subject accuracy compared to existing methods, effectively closing the accuracy gap between intra-subject and cross-subject classification. It requires only a few gesture repetitions to converge to accurate decision boundaries and remains robust to variations in data characteristics.