In this paper, we decided to study the effect of extracted audio features, using the analysis tool Essentia, on the quality of constructed music emotion detection classifiers. The research process included constructing training data, feature extraction, feature selection, and building classifiers. We selected features and found sets of features that were the most useful for detecting individual emotions. We examined the effect of low-level, rhythm and tonal features on the accuracy of the constructed classifiers. We built classifiers for different combinations of feature sets, which enabled distinguishing the most useful feature sets for individual emotions.