Impact of Kernel Dimensionality on the Generalizability and Efficiency of Convolutional Neural Networks to Decode Neural Drive from High-density Electromyography Signal

Abstract

Convolutional neural networks (CNNs) have been widely used to estimate neural drive from high-density surface electromyography (HD-sEMG) signals in neural machine interfaces owing to their real-time capability. Depending on kernel dimensionality (1D, 2D, or 3D), CNNs can extract temporal, spatial, or spatiotemporal features. Given that motor unit action potentials propagate across both space and time, architectures that exploit spatial features may offer advantages for neural drive estimation. Despite the potential importance of kernel dimensionality, its influence on neural drive estimation remains poorly understood. Existing studies have mainly evaluated CNN generalizability across participants, contraction intensities, or muscles within the same HD-sEMG dataset, while computational efficiency has seldom been considered. As a result, it remains unclear whether different kernel dimensionalities affect cross-dataset generalizability and computational efficiency. In this study, we implemented three CNN architectures--differing only in kernel dimensionality-- to investigate whether exploiting the spatial and spatiotemporal features of motor unit action potentials improves the generalizability and computational efficiency of neural drive estimation from HD-sEMG recorded during lower-limb isometric contractions. We trained the CNNs on one HD-sEMG dataset and evaluated them, without retraining, on two independent, unseen datasets recorded from different participants, sessions, and protocols--one spanning three contraction intensities and the other three muscles. All three architectures are generalized to both unseen datasets. The 2D and 3D CNNs marginally outperformed the 1D CNN with a 0.2% increase in R, while the 3D CNN showed no advantage over the 2D CNN. Computational efficiency depended on kernel dimensionality in a platform-dependent manner. On the CPU, the 3D CNN showed the slowest inference time, which was 2x slower than the 2D and 1D CNN, owing to the higher arithmetic cost of its spatiotemporal convolutions. On the GPU, all three architectures achieved similar inference times of about 1.36 ms/sample. These findings indicate that increased architectural complexity of CNN does not improve generalizability for neural drive estimation, and that a 2D CNN offers the best balance of accuracy and efficiency for a reliable, deployable CNN-based neural drive estimator--particularly on CPU-only or resource-constrained platforms.