The peripheral nervous system plays a crucial role in facilitating communication between biological systems. However, decoding neural signals from peripheral nerve recordings remains a challenge due to their complex spatiotemporal patterns. In this study, we propose a graph-based learning approach to more effectively capture temporal and spatial information for classifying neural signal patterns. Unlike previous work, our method incorporates the physical geometry of the nerve cuff, addressing the underrepresented relationships between electrodes. We used a publicly available dataset consisting of neural recordings from eight Long-Evans rats, obtained using a 56-channel nerve cuff electrode. We constructed graphs where each node represents the time series recorded from an electrode, and edges correspond to the distances between electrodes along the surface of the nerve cuff (e.g., geodesic distance). We employed a leave-one-out strategy to evaluate the generalizability of the approach. We further evaluated the within-rat performance of the model by training on two folds and testing on the remaining fold of each rat’s data. In generalizability evaluation, we achieved a mean F1 score of 65.03%, representing a 17.74% improvement over the previous study, and in within-rat testing, we achieved a mean F1 score of 77.50%, representing a 3.14% increase. These findings highlight the value of incorporating the recording geometry into model design, particularly in this small dataset setting, where explicit spatial priors help compensate for limited training examples and improve decoding performance.
Ji et al. (Fri,) studied this question.