mMyoHMI: Real-Time EMG-Based Pattern Recognition with On-Device Learning Adaptation

With the advancements in sensor technology and learning algorithm, applying Electromyography (EMG) sensing systems for human-machine interface (HMI) applications has gained tremendous attentions in the past decades. However, most of the existing EMG pattern recognition systems have limitations in terms of reliability, and accessibility. To address these issues, we introduce mMyoHMI, a flexible, low-cost, and adaptive mobile EMG-based pattern recognition system for HMI applications. The mMyoHMI interface is compatible with the commercial Myo armband and possesses the capability to adapt to the inherent time-varying biological heterogeneity present in EMG signals through on-device learning. Our system provides a choice between two learning models: a Deep Learning (DL) model using Convolutional Neural Networks (CNN) for robust feature learning from a diverse user cohort, and a conventional Machine Learning (ML) model employing Linear Discriminant Analysis (LDA) for rapid, efficient adaptation in resource-constrained mobile devices. The CNN model, pretrained on a large-scale EMG dataset, enables itself to acquire generalized EMG feature knowledge from a diverse user cohort, and thus establishes a robust foundation for on-device learning. While the LDA model offers a lightweight user-specific adaptation optimized for mobile devices with limited computational resources. Empirical evaluation shows that the proposed MyoHMI system successfully meets stringent real-time requirements for mobile devices while achieving an accuracy of 99 % and 98.8 % with the LDA and CNN models, respectively, upon EMG adaptation. Therefore, mMyoHMI has the potential to empower a broader user base to access EMG pattern recognition, offering the flexibility and scalability to accommodate diverse user-specific biological heterogeneities for lifelong usage. We released all the source code at: https://github.com/MIC-Laboratory/mMyoHMI