Real-time gesture classification via spike trains and multi-scale temporal features from sEMG signals

Abstract Classical gesture classification based on surface electromyographic (sEMG) signals often suffers from the cross-talk problem. To solve this, we propose a lightweight deep learning framework that integrates multi-scale temporal features from spike trains to improve the accuracy, stability, and real-time applicability. First, a hybrid offline-online decomposition framework based on Convolution Kernel Compensation (CKC) is adopted. In the offline phase, active signal segments are isolated with iteratively optimized filters to extract individual Motor Units (MUs). In the online phase, these pre-optimized filters are reused to enable real-time MU tracking. Second, features are extracted using Spike Count (SC) and Temporal Difference (TD) within a 200 ms sliding window with a 50 ms step, followed by gesture classification using a lightweight two-layer Multi-Layer Perceptron (MLP). The method was validated on three high-density sEMG (HD-sEMG) datasets: a 65-Gesture dataset (20 subjects), CSL-HDEMG (5 subjects), and CapgMyo DBa (10 subjects). Specifically for the 65-Gesture dataset (subset validation), the offline decomposition yielded an average of 1153.7 ± 75.2 high-quality MUs per subject, with an average Pulse-to-Noise Ratio (PNR) of 27.28 ± 2.4 dB. The system demonstrated high efficiency, with a single-window decomposition latency of less than 5 ms and a total end-to-end latency below 60 ms. In terms of classification performance, the proposed method achieved an accuracy of 91.02% ± 2.68% on this dataset. Additionally, it outperformed traditional time-domain features such as Root Mean Square on standard benchmarks, achieving 92.06% ± 11.17% on CSL-HDEMG and 86.63% ± 7.76% on CapgMyo DBa. The proposed framework achieves high precision and low latency in decoding MU discharges across diverse gesture types. By demonstrating the efficacy of multi-scale temporal features derived from decomposed motor unit spike trains and lightweight deep learning, the study provides a robust foundation for real-time neural decoding in advanced prosthetic control and rehabilitation interfaces.