Cost-sensitive learning (CSL) emerges as a promising approach for machine learning (ML) model performance evaluation. CSL weighs the severity of the prediction errors/ misclassifications made in the form of costs. Varying consequences of misclassification costs are often overlooked in existing research that uses standard metrics such as prediction accuracy leading to suboptimal performances in real-world settings. This paper emphasizes the need of CSL for emotion recognition models that use Electroencephalography (EEG) data. It proposes a novel cost sensitive framework that addresses the challenge of reducing the overall misclassification cost of existing EEG based emotion recognition models and thereby, creating a tradeoff with prediction accuracy. This proposed cost-sensitive framework is based on ensemble ML. Random Forest, Support Vector Machine, and k-Nearest Neighbor algorithms were heterogeneously ensembled to create this framework. It was compared with individual ML models built from same algorithms, which are considered cost insensitive as they neglect the varying misclassification costs. EEG data that classifies students as confused and non-confused during learning were used as the dataset. Results indicated that the designed ensemble model achieved a misclassification cost of 1811.5, which is a reduction of 38.20% compared to the highest misclassification cost incurred by a cost-insensitive individual model. Moreover, the ensemble model reached a prediction accuracy of 70.01%, whereas the individual models did not exceed 70% accuracy. This demonstrates the effectiveness of the proposed cost-sensitive model, as the cost-insensitive models could not achieve this level of tradeoff between misclassification cost and prediction accuracy.