This manuscript explores the development of a technique for detecting the affective states of Virtual Reality (VR) users in real-time. The technique was tested with data from an experiment where 18 participants observed 16 videos with emotional content inside a VR home theater, while their electroencephalography (EEG) signals were recorded. Participants evaluated their affective response toward the videos in terms of a three-dimensional model of affect. Two variants of the technique were analyzed. The difference between both variants was the method used for feature selection. In the first variant, features extracted from the EEG signals were selected using Linear Mixed-Effects (LME) models. In the second variant, features were selected using Recursive Feature Elimination with Cross Validation (RFECV). Random forest was used in both variants to build the classification models. Accuracy, precision, recall and F1 scores were obtained by cross-validation. An ANOVA was conducted to compare the accuracy of the models built in each variant. The results indicate that the feature selection method does not have a significant effect on the accuracy of the classification models. Therefore, both variations (LME and RFECV) seem equally reliable for detecting affective states of VR users. The mean accuracy of the classification models was between 87% and 93%.