Análise Multivariada de Sinais sEMG para Identificação de Unidades Motoras e Classificação de Fibras Musculares

Abstract

Surface electromyography (sEMG), a non-invasive technique, captures signals that allow for the assessment of neuromuscular activity, proving useful in both patient diagnosis and sports science. Accurate analysis of these signals presents several challenges, including external electromagnetic interference and electrical potentials from other physiological processes, which may cause signal overlap. This study focuses on processing discrete sEMG signals with the objective of estimating a virtual number of Motor Units (MUs) and muscle fibers involved in muscular activity (Type I - slow-twitch, Type IIa - fast-twitch oxidative, and Type IIx/IIb - fast-twitch glycolytic), based on the ionic currents detected by the electrodes, which are currently obtained, with due precision, only through invasive methods. The methodology begins with signal pre-processing, including Notch filtering, fourth-order Butterworth band-pass filtering for noise removal, and smoothing via a moving average to eliminate outliers. Subsequently, feature extraction is performed, encompassing variance, standard deviation, root mean square (RMS), skewness, and kurtosis, allowing for the assessment of the statistical and stochastic behavior of the signal. For time-domain and frequency-domain analyses, the Discrete Wavelet Transform (DWT) is applied. This output serves as the input for the K-Means clustering algorithm, which segments activation patterns and identifies dominant frequencies. The analyses were conducted using experimental data obtained under controlled and standardized conditions, considering voluntary muscle contractions in upper limb regions (Upper Limb Protocol) and lower limb regions (Lower Limb Protocol – Anterior and Posterior Views) of individuals with a load of 50 kg in both protocols. Experimental results demonstrate the algorithm’s effectiveness in distinguishing muscle activation levels and estimating clinically relevant parameters, such as neural firing rates, while differentiating the types of muscle fibers involved. This provides a virtual estimate of both activated motor units and fiber types. Further clinical validation is required to confirm the robustness of these findings and support their medical application. Where data are available, it may also be possible to map virtual motor units and fibers to actual anatomical structures. The proposed method is safe and suitable for therapeutic applications and for the prevention of muscle mass loss.