Electromyograph (EMG) has the properties of large variations and nonstationarity. There are two issues in the classification of EMG signals. One is the feature selection, and the other is the classifier design. Subject to the first issue, we propose a supervised feature mining (SFM) method, which is an intelligent approach based on genetic algorithms (GAs), fuzzy measure, and domain knowledge on pattern recognition. The SFM can find the optimal EMG feature subset automatically and remove the redundant from a large amount of feature candidates without taking trial-and-error. In the experiments, all feature candidates and optimal feature subset are conducted to demonstrate the validity of the proposed SFM. Moreover, experimental results show that the optimal EMG feature subset contained from SFM can obtain higher classification rates compared with using all feature candidates by K-NN method.