Artificial Neural Networks-based Real-time Classification of ENG Signals for Implanted Nerve Interfaces

TL;DR

The paper tackles real-time classification of ENG signals for implanted ND&S systems by introducing a MIMO ENG signal model that captures multi-axon, cuff-electrode measurements and channel coupling. Four ANN architectures (CNN, Inception Time, ENGNet, LSTM) are evaluated on a public rat ENG dataset, with pre-processing and latency analysis to ensure feasibility within the human $\sim300$ ms response window. ENGNet emerges as the most suitable for real-time deployment, achieving >$90\%$ F1-score with $T_w$ as small as $100$ ms while maintaining low computational load, and the MIMO model provides a principled basis for kernel sizing and feature extraction. The work demonstrates that real-time neural decoding and forward stimulation is feasible with carefully designed architectures and a physics-informed ENG model, suggesting practical ND&S pathways and highlighting the importance of latency-aware design in implanted devices.

Abstract

Neuropathies are gaining higher relevance in clinical settings, as they risk permanently jeopardizing a person's life. To support the recovery of patients, the use of fully implanted devices is emerging as one of the most promising solutions. However, these devices, even if becoming an integral part of a fully complex neural nanonetwork system, pose numerous challenges. In this article, we address one of them, which consists of the classification of motor/sensory stimuli. The task is performed by exploring four different types of artificial neural networks (ANNs) to extract various sensory stimuli from the electroneurographic (ENG) signal measured in the sciatic nerve of rats. Different sizes of the data sets are considered to analyze the feasibility of the investigated ANNs for real-time classification through a comparison of their performance in terms of accuracy, F1-score, and prediction time. The design of the ANNs takes advantage of the modelling of the ENG signal as a multiple-input multiple-output (MIMO) system to describe the measures taken by state-of-the-art implanted nerve interfaces. These are based on the use of multi-contact cuff electrodes to achieve nanoscale spatial discrimination of the nerve activity. The MIMO ENG signal model is another contribution of this paper. Our results show that some ANNs are more suitable for real-time applications, being capable of achieving accuracies over $90\%$ for signal windows of $100$ and $200\,$ms with a low enough processing time to be effective for pathology recovery.

Figures (19)

• Figure 1: (a) Representation of a nerve injury and related issues. (b) Steps necessary to re-establish the connection [6]Elisa and their possible implementation within the body [7]Federica.
• Figure 2: The electromagnetic nanonetwork of the nerve/axon-cuff electrode interface. From the ENG measurement to the data processing.
• Figure 3: Sampled signal along the contacts of the cuff electrode.
• Figure 4: Detail of the all operations implemented in a ND&S system to implement a functional neural bypass.
• Figure 5: Section of a peripheral nerve showing the $n$th axon disposed inside the nerve generating a signal $x_n(t)$.