Low-Complexity CNN-Based Classification of Electroneurographic Signals
1 st Arek Berç Gökdağ, 2 nd Given Name Surname, 3 rd Given Name Surname, 4 th Given Name Surname, 5 th Given Name Surname, 6 th Given Name Surname, Arek Berç Gökdağ, Michele Zhu, Silvia Mura, Maurizio Magarini, Antonio Coviello, Umberto Spagnolini · 2025
Read Low-Complexity CNN-Based Classification of Electroneurographic Signals by 1 st Arek Berç Gökdağ, 2 nd Given Name Surname, 3 rd Given Name Surname, 4 th Given Name Surname, 5 th Given Name Surname, 6 th Given Name Surname, Arek Berç Gökdağ, Michele Zhu, Silvia Mura, Maurizio Magarini, Antonio Coviello, Umberto Spagnolini (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-06241.txt.
Introduction
Arek Berç Gökdağ and co-authors shrink a nerve-signal classifier until it fits inside an implant in a 2025 paper. Peripheral nerve interfaces record the faint electrical chatter of nerves to drive prosthetics and therapies. Reading that chatter live, inside the body, leaves no room for heavy computation or slow answers.
Their network, MobilESCAPE-Net, reworks a leading architecture for thrift. It matches the original's accuracy and F1 score while cutting trainable parameters by nearly all of them and arithmetic operations by over nine tenths. Fewer operations mean faster verdicts and less heat, the two currencies implants live by. The authors position the design for real-time classification in devices that must sip power for years. The study shows accuracy and frugality need not trade off as steeply as assumed.
Catalog record
| CREATOR | 1 st Arek Berç Gökdağ, 2 nd Given Name Surname, 3 rd Given Name Surname, 4 th Given Name Surname, 5 th Given Name Surname, 6 th Given Name Surname, Arek Berç Gökdağ, Michele Zhu, Silvia Mura, Maurizio Magarini, Antonio Coviello, Umberto Spagnolini |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 25 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-06241
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Low-Complexity CNN-Based Classification of Electroneurographic Signals 1 st Arek Berç Gökdağ, 2 nd Given Name Surname, 3 rd Given Name Surname, 4 th Given Name Surname, 5 th Given Name Surname, 6 th Given Name Surname, Arek Berç Gökdağ, Michele Zhu, Silvia Mura, Maurizio Magarini, Antonio Coviello, Umberto Spagnolini Abstract Peripheral nerve interfaces (PNIs) facilitate neural recording and stimulation for treating nerve injuries, but real-time classification of electroneurographic (ENG) signals remains challenging due to constraints on complexity and latency, particularly in implantable devices. This study introduces MobilESCAPE-Net, a lightweight architecture that reduces computational cost while maintaining and slightly improving classification performance. Compared to the state-of-the-art ESCAPE-Net, MobilESCAPE-Net achieves comparable accuracy and F1-score with significantly lower complexity, reducing trainable parameters by 99.9% and floating point operations per second by 92.47%, enabling faster inference and real-time processing. Its efficiency makes it well-suited for low-complexity ENG signal classification in resource-constrained environments such as implantable devices. I Introduction The peripheral nervous system (PNS) transmits sensory and motor information between the central nervous system and the body. Damage to the PNS can cause peripheral neuropathy (PN), leading to sensory, motor, and autonomic impairments with long-term functional and socio-economic consequences. Mechanical injuries are a major cause of PN, contributing to disabilities that affect individuals and healthcare systems worldwide [ 1 ] . Although the PNS can regenerate, nerve injury recovery is often incomplete, resulting in lasting impairments. Traditional treatments have limited effectiveness, increasing interest in bioelectronic medicine, particularly implantable peripheral nerve interfaces (PNIs) [ 2 ] . These devices record electroneurographic (ENG) signals and enable neural decoding and stimulation (ND&S), showing promising therapeutic potential [ 3 ] . A major challenge in ND&S systems is the real-time classification of ENG signals, as the complexity of nerve signal propagation makes it difficult to capture meaningful activity with low latency. Classification is essential for interpreting neural intent or physiological events and delivering appropriate stimulation. Minimizing processing delays is particularly critical for restoring sensory and motor functions, where responses must occur within the 300\, ms threshold, beyond which patients begin to perceive a delay [ 4 ] . Overcoming these challenges is key to improve neural interfacing and advancing bioelectronic medicine. Various classification methods have been explored for electroencephalography (EEG) and electromyographic (EMG) signals, which share similarities with ENG signals. These approaches include statistical models [ 5 ] , machine learning (ML) methods [ 6 ] , and deep learning (DL) frameworks [ 7 ] . However, statistical models and ML methods often struggle with the significant noise and distortion in ENG, EEG, and EMG signals, leading to lower classification performance compared to DL methods [ 6 ] , which reduces their feasibility for use in ND&S systems. Deep learning techniques, particularly convolutional neural networks (CNNs), have been applied to ENG signal classification, including recordings from the sciatic nerve of Long-Evans rats [ 7 , 8 ] . The study in [ 7 ] compared several models, including ENGNet (adapted from EEGNet [ 9 ] ), Long Short-Term Memory (LSTM), and inception time models, with ENGNet achieving high classification performance while maintaining a simple design. Similarly, [ 8 ] introduced ESCAPE-Net, where raw ENG signals are preprocessed into spatiotemporal signatures (detailed in Sec. III ) for classification. While DL techniques generally offer superior classification accuracy compared to statistical and traditional ML methods [ 10 ] , their high computational demands [ 7 , 8 ] pose challenges for real-time applications. This limitation underscores the need for more efficient models capable of reducing computational overhead while maintaining classification performance. To address this issue, this work introduces MobilESCAPE-Net, a novel architecture designed to reduce computational cost while preserving robust classification performance significantly. MobilESCAPE-Net is validated against ESCAPE-Net, which serves as a baseline using the same dataset from [ 11 ] . Numerical results demonstrate that a key advantage of MobilESCAPE-Net is its substantial reduction in model size and computational complexity, achieving a 99.92\,\% decrease in parameters, requiring only 68\, k, and a 92.47\,\% reduction in floating point operations per second (FLOPs), requiring just 81.8\, MFLOPs. Despite this significant decrease in model size and computational overhead, MobilESCAPE-Net not only matches but slightly outperforms ESCAPE-Net in both accuracy and F1-score. This efficiency makes MobilESCAPE-Net particularly well-suited for low-latency, resource-constrained environments, demonstrating the effectiveness of its optimized architecture. The rest of the paper is structured as follows. Section II describes the system model under consideration, while Sec. III details the preprocessing of the ENG signal. Section IV presents the classification methods, including both state-of-the-art and the proposed approach. Numerical results are provided in Sec. V , and conclusions are drawn in Sec. VI . Notation Lower-case letters \mathbf{x} refer to vectors and the notation x_{l,k} refers to the (l,k) th element. The symbols ||\cdot|| , \mathbb{E} and \mathbb{R} refer to the Euclidean norm operator, expectation operator, and to the real number set, respectively. II System Model Nerves consist of tightly packed axon bundles that facilitate the transmission of electrical ENG signals between the central nervous system and peripheral tissues, playing a crucial role in maintaining both signal conduction and structural stability. These ENG signals travel in an afferent direction to convey sensory input to the central nervous system or in an efferent direction to deliver motor commands to muscles [ 1 ] . To capture the aggregated neural activity generated by individual axons, we consider the cuff electrode illustrated in Fig. 1 , which is modeled as a cylindrical structure composed of N rings, each containing M electrodes to enhance the redundancy of the measured data. Consequently, the total number of electrodes in the cuff electrode is given by L \,=\, N \,\times\, M . Assuming the presence of K uncorrelated ENG sources, represented by \mathbf{s}(t)\in\mathbb{R}^{K\times 1} , the signal {y}_{l} received by the l th electrode can be modeled as [ 7 ] \displaystyle{y}_{l}(t)=\sum_{k=1}^{K}{h}_{l,k}\,\,{s}_{k}(t-\tau_{l,k})+{v}_{l}(t)+{u}(t), (1) where {u}(t) represents the i.i.d. additive white Gaussian noise common to all electrodes and {v}_{l}(t) denotes the interference signal. This interference arises from the cumulative effect of EMG signals generated by surrounding muscles during contractions and common to all the electrodes, along with electrode-specific artifacts [ 7 ] . Therefore, the interference signals exhibit a nonzero correlation, \mathbb{E}[{v}_{l+1}(t)\,\,{v}_{l}(t)]\neq 0,\forall l . The source signal \mathbf{s}(t) combines efferent and afferent neural activity and is modeled as the sum of multiple compound action potentials (CAP). CAPs are extracellular spikes observed in nerve cuff recordings, resulting from the synchronous activation of small groups of nerve fibers [ 8 ] . The parameter \tau_{l,k} denotes the propagation delay associated with the k th source as perceived by the l th electrode. The lead field h_{l,k} is defined as in [ 7 ] \displaystyle h_{l,k}=-\frac{1}{4\pi\sigma||\mathbf{q}_{l}-\mathbf{q}_{k}||^{2}} (2) with \sigma representing the conductivity and \mathbf{q}_{l} and \mathbf{q}_{k} representing the position of the l th electrode and the position of the k th source , respectively, with respect to an arbitrary reference system. Fig. 1 : Cuff electrode model with example ENG signals recorded for dorsiflexion, plantarflexion, and pricking. III ENG Signal Preprocessing This section describes the preprocessing steps designed to enhance the quality of noisy ENG signals, which require a robust processing pipeline. A key aspect of this process is the detection of CAPs, which