Beyond Subject-Specific Models in Dynamical Human-Machine Interaction: Benchmarking and Optimization Strategies

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Surface electromyography (EMG) offers intuitive control for human-machine interfaces.
• Accurate EMG-based finger position estimation requires robust temporal modeling and user adaptation.
• Existing methods often struggle with user variability and real-time performance.

Purpose of the Study:

• To benchmark various deep learning architectures for continuous finger position estimation from EMG.
• To investigate adaptive learning strategies for improved cross-subject generalization.
• To introduce and evaluate neural ordinary differential equations (NODEs) for EMG-based regression.

Main Methods:

• Benchmarking recurrent neural networks, temporal convolutional networks (TCNs), Transformers, and neural ordinary differential equations (NODEs).
• Systematic tuning of model receptive fields based on EMG autocorrelation.
• Implementation of adaptive learning methods including multitask, transfer, and meta-learning, with lightweight fine-tuning (LoRA, adapter layers).

Main Results:

• Temporal Convolutional Networks (TCNs) achieved state-of-the-art performance on the Ninapro DB8 dataset.
• Mean absolute errors (MAEs) below 5.4 were achieved with multitask and transfer learning.
• A mean absolute error (MAE) of 6.47 was obtained with two-shot meta-learning.

Conclusions:

• TCNs provide a highly effective architecture for EMG-to-kinematics regression.
• Adaptive learning strategies significantly enhance cross-subject generalization for EMG-based control.
• These advancements offer practical solutions for personalized, real-time control in prosthetics, VR, and teleoperation.