What happened to disruptive transformation of education?

Technology has become an essential part of our everyday life, and its use in educational environments keeps growing. In addition, games are one of the most popular activities across cultures and ages, and there is ample evidence that supports the benefits of using games for assessment. This field is commonly known as game-based assessment (GBA), which refers to the use of games to assess learners' competencies, skills, or knowledge. This paper analyzes the current status of the GBA field by performing the first systematic literature review on empirical GBA studies. It is based on 65 research papers that used digital GBAs to determine: (1) the context where the study has been applied; (2) the primary purpose; (3) the domain of the game used; (4) game/tool availability; (5) the size of the data sample; (6) the computational methods and algorithms applied; (7) the targeted stakeholders of the study; and (8) what limitations and challenges are reported by authors. Based on the categories established and our analysis, the findings suggest that GBAs are mainly used in K-16 education and for assessment purposes, and that most GBAs focus on assessing STEM content, and cognitive and soft skills. Furthermore, the current limitations indicate that future GBA research would benefit from the use of bigger data samples and more specialized algorithms. Based on our results, we discuss current trends in the field and open challenges (including replication and validation problems), providing recommendations for the future research agenda of the GBA field.

In the twentyfirst-century society, several soft skills are fundamental, such as stress management, which is considered one of the key ones due to its strong relationship with health and well-being. However, this skill is hard to measure and master without external support. This paper tackles stress detection through artificial intelligence (AI) models and heart rate, analyzing in WESAD and SWELL-KW datasets five supervised models and five unsupervised anomaly detection models that had not been tested before for stress detection. Also, we analyzed the transfer learning capabilities of the AI models since it is an open issue in the stress detection field. The models with the highest performance on test data were the anomaly detection Local Outlier Factor (LOF) with F1-scores of 88.89% in WESAD and 77.17% in SWELL-KW, and the supervised Multi-layer Perceptron (MLP) with F1-scores of 99.03% in WESAD and 82.75% in SWELL-KW. However, when evaluating the transfer learning capabilities of these AI models, MLP performed much worse on the other dataset, decreasing the F1-score to 28.41% in SWELL-KW and 57.28% in WESAD. In contrast, LOF reported better transfer learning performance achieving F1-scores of 70.66% in SWELL-KW and 85.00% in WESAD. Finally, we found that training AI models with both datasets (i.e., with data from different contexts) improved the average performance of the models and their generalization; with this setup, LOF achieved F1-scores of 87.92% and 85.51% in WESAD, and 78.03% and 82.16% in SWELL-KW; whereas MLP obtained 78.36% and 81.33% in WESAD, and 79.37% and 80.68% in SWELL-KW. Therefore, we suggest as a promising direction the use of anomaly detection models or multi-contextual training to improve the transfer learning capabilities in this field, which is a novelty in the literature. We believe that these AI models combined with the use of non-invasive wearables can enable a new generation of stress management mobile applications.