Abstract
Neurostimulation and neural recording are crucial to develop neuroprostheses that can restore function to individuals living with disabilities. While neurostimulation has been successfully translated into clinical use for several applications, it remains challenging to robustly collect and interpret neural recordings, especially for chronic applications. Nerve cuff electrodes offer a viable option for recording nerve signals, with long-term implantation success. However, nerve cuff electrodes' signals have low signal-to-noise ratios, resulting in reduced selectivity between neural pathways. The objective of this study was to determine whether deep learning techniques, specifically networks tailored for time series applications, can increase the recording selectivity achievable using multi-contact nerve cuff electrodes. We compared several neural network architectures, the impact and trade-off of window length on classification performance, and the benefit of data augmentation. Evaluation was carried out using a previously collected dataset of 56-channel nerve cuff recordings from the sciatic nerve of Long-Evans rats, which included afferent signals evoked using three types of mechanical stimuli. Through this study, the best model achieved an accuracy of 0.936 ± 0.084 and an F1-score of 0.917 ± 0.103, using 50 ms windows of data and an augmented training set. These results demonstrate the effectiveness of applying CNNs designed for time-series data to peripheral nerve recordings, and provide insights into the relationship between window duration and classification performance in this application.