Classification Model for Lumbar Muscle Fatigue Evaluation Based on Machine Learning

Abstract

In this study, a classification model for lumbar muscle fatigue evaluation was constructed using surface electromyographic (sEMG) signals combined with subjective fatigue assessment scales. Twelve healthy volunteers were selected, and sEMG signals of their lumbar erector spinae muscles were collected; the signals were preprocessed through infinite impulse response (IIR) notch filtering, Butterworth high-pass filtering, and 4-level soft threshold denoising with db4 wavelets, then segmented using a sliding window with a window length of 1000 sampling points and a step size of 200 sampling points, followed by the extraction of 14 features in total (including those from the time domain, frequency domain, time-frequency domain, and nonlinear domain). Feature dimensionality reduction was conducted via recursive feature elimination (RFE) and principal component analysis (PCA); based on 3092 samples categorized into three classes, the performance of six model combinations was compared using 10-fold cross-validation. The results showed that the combination of PCA and random forest (RF) achieved the optimal performance, with an average accuracy of 90.42% and a Kappa coefficient of 0.8056, and both the RF and support vector machine (SVM) models exhibited a recognition rate of over 83% for all three fatigue states. This study demonstrates that the proposed model can realize non-invasive evaluation of lumbar muscle fatigue, thereby providing technical support for related fields.