Recently, Crisis Management Serious Games (CMSG) have proved their potential for teaching both technical and soft skills related to managing crisis in a safe environment while reducing training costs. In order to improve learning outcomes insured by CMSGs, many works focus on their evaluation. Despite its great interest, the learner emotional state is often neglected in the evaluation process. Indeed, negative emotions such as boredom or frustration degrade the learning quality since they frequently conduct to giving up the game. This research addresses this gap by combining gaming and affect aspects under an Educational Data Mining (EDM) approach to improve learning outcomes. Therefore, we propose an EDM-based multimodal method for assessing learners' affective states by classifying data communicated in text messaging and facial expressions. This method is applied to assess learners' engagement during a game-based collaborative evacuation scenario. The obtained assessment results will be useful for adapting the game to the different players' emotions.