<p>For surface electromyography (sEMG) based human–machine interaction systems, accurately recognizing the users’ gesture intent is crucial. However, due to the existence of subject-specific components in sEMG signals, subject-specific models may deteriorate when applied to new users. In this study, we hypothesize that in addition to subject-specific components, sEMG signals also contain pattern-specific components, which is independent of individuals and solely related to gesture patterns. Based on this hypothesis, we disentangled these two components from sEMG signals with an auto-encoder and applied the pattern-specific components to establish a general gesture recognition model in cross-subject scenarios. Furthermore, we compared the characteristics of the pattern-specific information contained in three categories of EMG measures: signal waveform, time-domain features, and frequency-domain features. Our hypothesis was validated on an open source database. Ultimately, the combination of time- and frequency-domain features achieved the best performance in gesture classification tasks, with a maximum accuracy of 84.3%. For individual feature, frequency-domain features performed the best and were proved most suitable for separating the two components. Additionally, we intuitively visualized the heatmaps of pattern-specific components based on the topological position of electrode arrays and explored their physiological interpretability by examining the correspondence between the heatmaps and muscle activation areas.</p>