In order to address the limitation that surface electromyography (sEMG) signals from a single muscle group provide insufficient reference for rehabilitation training assessment, the multi-muscle sEMG feature fusion and selection method was proposed to provide a high-precision quantitative evaluation for patients with chronic pain and motor dysfunction. sEMG signals of the biceps brachii, triceps brachii, and brachioradialis were synchronously collected from 20 healthy subjects along with their ratings of perceived exertion (RPE). Wavelet denoising and feature extraction were performed to obtain time-domain (root mean square, RMS) and frequency-domain (median frequency, MF) features, and their correlations with RPE were further analyzed. Significant features were selected based on Pearson correlation coefficients, and multiple regression models—including linear regression (LR), support vector machine regression (SVMR), Lasso regression, and stepwise regression (SWR)—were constructed and compared. Results show that the SVMR model achieved the best performance when fusing RMS and MF features of the three muscle groups, with the RMS of the brachioradialis exhibiting the highest correlation with RPE (r = 0.841). The mean square error (MSE) is 0.077, representing a 78% reduction compared with single-muscle models, while the coefficient of determination (R²) reached 0.950, corresponding to a 35% improvement in evaluation accuracy. The proposed multi-muscle sEMG feature processing and modeling method can effectively enhance the accuracy of rehabilitation training assessment and demonstrates promising potential for clinical application and translation.