The Role of Functional Muscle Networks in Improving Hand Gesture Perception for Human-Machine Interfaces

TL;DR

This work addresses hand gesture perception for neurorobotics by shifting from single-muscle activation to coherence-based functional muscle networks derived from surface EMG. It defines and exploits Magnitude-Squared Coherence ($MSC$) across 12 muscles to form functional networks, then uses the frequency-averaged MSC as features for a shallow polynomial SVM, achieving $85.1\%$ average accuracy on the Ninapro DB2 Exercise B dataset (40 subjects, 17 gestures) with substantially reduced computational demands. The approach demonstrates that muscular coordination patterns captured by $MSC$ can encode essential gesture information more efficiently than many deep learning models trained on the same data. This has practical implications for real-time, energy-efficient neurorobotic control and interactive systems, with potential extensions to broader gesture sets and real-world deployment.

Abstract

Developing accurate hand gesture perception models is critical for various robotic applications, enabling effective communication between humans and machines and directly impacting neurorobotics and interactive robots. Recently, surface electromyography (sEMG) has been explored for its rich informational context and accessibility when combined with advanced machine learning approaches and wearable systems. The literature presents numerous approaches to boost performance while ensuring robustness for neurorobots using sEMG, often resulting in models requiring high processing power, large datasets, and less scalable solutions. This paper addresses this challenge by proposing the decoding of muscle synchronization rather than individual muscle activation. We study coherence-based functional muscle networks as the core of our perception model, proposing that functional synchronization between muscles and the graph-based network of muscle connectivity encode contextual information about intended hand gestures. This can be decoded using shallow machine learning approaches without the need for deep temporal networks. Our technique could impact myoelectric control of neurorobots by reducing computational burdens and enhancing efficiency. The approach is benchmarked on the Ninapro database, which contains 12 EMG signals from 40 subjects performing 17 hand gestures. It achieves an accuracy of 85.1%, demonstrating improved performance compared to existing methods while requiring much less computational power. The results support the hypothesis that a coherence-based functional muscle network encodes critical information related to gesture execution, significantly enhancing hand gesture perception with potential applications for neurorobotic systems and interactive machines.

Figures (5)

• Figure 1: A schematic of the presented research. The NinaPro DB2 (Exercise B) dataset provides sEMG recordings of $17$ movements performed by $40$ healthy subjects. After minor pre-processing, the recordings are used to obtain muscular coherence values, which are then used as features for a shallow support vector machine gesture classifier.
• Figure 2: The $17$ gestures performed in the Ninapro dataset DB2 (exercise set B).
• Figure 3: Coherence matrices obtained for each gesture from Subject 19. Each matrix is $12 \times 12 = 144$ squares, representing the corresponding signal combination's median coherence values across all repetitions. The color is proportional to the pairwise coherence. Since interconnectivity is uni-directed, the matrices are all symmetric. The diagonal terms, corresponding to self-pairing signals, were manually set to $0$ in order to avoid overshadowing other connectivity patterns with low MSC values.
• Figure 4: Coherence muscle networks from Subject 1 obtained for each hand gesture. Each node in the network represents a sensor, while the color of the edges between the nodes denotes the strength of the coherence between them, where blue corresponds to $MSC = 0$ (no coherence) and red corresponds to $MSC = 1$ (maximum coherence).
• Figure 5: Average confusion matrix across all subjects. On average, $28.9$ samples were correctly predicted out of the total $34$ samples ($2$ repetitions for $17$ gestures). It can be noted that classes presenting similar MSC values, such as Class 15 and Class 9, are the most likely confused.