This paper proposes an empirical framework that aims to classify difficulty according to the player's physiological response. As part of the experimental protocol, a simple puzzle-based Virtual Reality (VR) videogame with three levels of difficulty was developed, each targeting a distinct region of the valence-arousal space. A study involving 32 participants was conducted, during which physiological responses (EDA, ECG, Respiration), were measured alongside emotional ratings, which were self-assessed using the Self-Assessment Manikin (SAM) during gameplay. Statistical analysis of the self-reports verified the effectiveness of the three levels in eliciting different emotions. Furthermore, classification using a Support Vector Machine (SVM) was performed to predict difficulty considering the physiological responses associated with each level. Results report an overall F1-score of 74.05% in detecting the three levels of difficulty, which validates the adopted methodology and encourages further research with a larger dataset.