Sonomyography-Based Decoding of Attempted Hand Movements in Individuals with Spinal Cord Injury

Restoring hand function is critical for individuals with spinal cord injury (SCI). This study develops and evaluates a sonomyography-based human-machine interface (HMI) for decoding attempted hand movements across individuals with different levels and severities of cervical SCI, using a sparse-ultrasound processing pipeline designed for 8 channel A-mode ultrasound systems.

A custom convolutional neural network, SonoSCINet, was developed using sparse B-mode ultrasound data and evaluated in non-disabled individuals and individuals with SCI, and compared with traditional machine learning classifiers. Feature selection and two-dimensional UMAP were used to examine handcrafted and learned representations. The same network architecture and processing pipeline were retrained and evaluated on an 8-channel A-mode ultrasound system in individuals with cervical SCI spanning different neurological injury levels and functional impairments.

SonoSCINet outperformed traditional models for gesture classification using sparse B-mode data (non-disabled: $91.3 \pm 12.79\%$; SCI: $82.03 \pm 23.48\%$). With the 8-channel A-mode system, SonoSCINet achieved a mean accuracy of $94.17 \pm 4.75\%$, numerically higher than LDA-based methods ($86.69 \pm 6.27\%$; not statistically significant), with improved robustness and gesture separability in the learned latent space. A real-time feasibility study in individuals with SCI further supports online implementation of the proposed pipeline for HMI applications.

SonoSCINet, trained with a subject-specific pipeline, enables accurate decoding of hand gestures across sparse B-mode and A-mode ultrasound systems and across a range of SCI impairment levels, supporting the feasibility of a unified processing framework across modalities.

These results support sonomyography combined with deep learning as a promising direction for portable A-mode ultrasound HMIs in individuals with SCI, with real-time feasibility of the underlying acquisition pipeline demonstrated using a lightweight classifier.