Continuous finger position estimation from surface electromyography (EMG) enables smoother, more intuitive control in human-machine interfaces than discrete gesture classification. Accurate regression, however, requires effective temporal modeling and adaptation to user variability. We benchmark recurrent neural networks, temporal convolutional networks (TCNs), Transformers, and, for the first time in this context, neural ordinary differential equations (NODEs) on the Ninapro database. Each model's receptive field (RF) is systematically tuned from EMG autocorrelation to ensure fair comparison. We further introduce a nonautonomous NODE variant with external inputs to represent finger movements as dynamical systems. To address cross-subject generalization, we explore adaptive learning paradigms, multitask, transfer, and first-order meta-learning, and adapt lightweight fine-tuning methods such as LoRA and adapter layers for temporal biosignal regression. On Ninapro DB8, the TCN achieves state-of-the-art mean absolute errors (MAEs) below 5.4 for multitask and transfer learning, and 6.47 for two-shot meta-learning. These findings advance EMG-to-kinematics regression and offer practical solutions for personalized, real-time control in prosthetics, virtual reality (VR), and teleoperation.
Manneschi et al. (Thu,) studied this question.