The applications of wearable sensors in daily life are increasing day by day. The wearable sensor technologies provide a wide range of knowledge that can be utilized by human gesture/action systems. Although the electromyography (EMG) sensors, which are one of the most simple and effective wearable sensors, are usually employed in especially medical diagnosis, it holds a great amount of information (e.g., activity descriptors) regarding human activities. This study aims to evaluate the applicability of EMG signals in action recognition by identifying the relevant features of EMG signals alongside their commonly used properties. These (relevant) features are categorized into general basic statistical variables in time/frequency/aLBP/uniform aLBP domains, special statistical features in time and frequency domains, AR coefficients, and features in the wavelet domain. In order to evaluate and validate the applicability of EMG signal, a set of features is extracted from a wide range of datasets established by the authors of this paper in a way of describing distinct human actions (total of 132 distinct features). The study determines the dominant features for human action recognition systems by proposing a new wearable sensor-based methodology that exploits feature selection algorithms. Hence, the determination of the relevancy order of the features is one of the significant outcomes of this study. The results figure out that the most relevant ones are the metrics that detect fluctuations, i.e., a higher amount of entropy, belonging to the signal.