Currently, electromyography pattern-recognition (EMG-PR) based myoelectric prosthesis is widely used in many laboratories worldwide. In the EMG-PR based method, EMG features would be extracted from the EMG signals and used to predict the user's motion intent. However, in clinical use, many interferences such as muscle fatigue, electrode shift and so on, were usually introduced to degrade the feature quality, which would decay the performance of a trained EMG-PR classifier in identifying motion intentions. In this study, a novel preprocessing strategy, feature filtering, was proposed to improve the performance of EMG-PR based classifier in motion classification. Three feature filtering methods of mean filter (MF), Median filter (MDF), and Weighted Average filter (WAF) were designed to investigate the effectiveness of this strategy. By analyzing the results of six able-bodied subjects, it demonstrated that the motion classification performance could be improved by using the feature filtering strategy, achieving the increments of 4.4%, 2.8%, and 3.5% for MF, MDF and WAF, respectively. These preliminary results suggest that using the feature filtering strategy may enhance the robustness of EMG-based myoelectric control.