Accurate classification of hand gestures from electromyography (EMG) data depends heavily on effective feature extraction. However, selecting and validating feature extraction methods (FEMs) remains computationally intensive and often dataset-specific. In this work, we evaluate several common FEMs using a fixed Convolutional Neural Network (CNN) architecture across multiple publicly available EMG datasets to ensure fair and consistent comparison. We also introduce Class Similarity Ratio (CSR), a novel heuristic for rapidly estimating the effectiveness of FEMs prior to full model training, significantly reducing computational overhead. In addition, we propose Target Activation Projection (TAP), a new FEM designed to improve robustness across datasets and segmentation strategies by abstracting temporal gesture features. Our findings show that the mean absolute value of unidirectional normalized EMG signals (mav(L2A)) achieves an average of 68% classification accuracy when full gestures are available. In contrast, TAP maintains stronger generalization with a consistent 57% average accuracy across both full and partial gestures. Lastly, results from CSR suggest that several FEMs could perform better with alternative classifiers or CNN configurations, highlighting the need for continued evaluation of the relationship between FEMs and classifiers.