To Örebro University

Early identification of mathematical difficulties of young learners is critical for their individual development, yet traditional assessments such as standardized tests or diagnostic interviews face challenges, e.g., in scalability and practicality. To address these challenges, we developed an AI-assisted, pedagogical domain-informed digital tool, DIDUNAS, to evaluate first graders’ individual skill profiles and identify their need for support. As our research question we ask, How accurately can AI identify first graders' need for support in mathematics?

Our study involved 290 students from 8 European schools (in Germany & Cyprus). The participants included multilingual students, students with disabilities, and diverse socioeconomic backgrounds. We used the ZAREKI-K standardized test (von Aster et al., 2009) as the ground truth of the students’ mathematical skill profiles. The test results indicated whether a student needed support or not. In our study, 71 students were identified as being in need of support in mathematics.

For developing the digital tool, we first let 55 students engage with 12 digitally presented mathematical activities (e.g., number line, quantity comparison, etc.) designed by pedagogical experts, and take individual standardized ZAREKI-K tests. We assessed both results with machine learning techniques to have an insight about which tasks contributed most to accurate identification of support needs based on students' errors in each task. The analysis allowed us to refine the assessment to just 10 tasks. The AI-assisted DIDUNAS app with the refined task set was validated with 235 students, and it was able to identify students in need of support with 85% accuracy with a much less time consumption compared to one-by-one diagnostic interviews.

Our results demonstrate that AI can enhance scalability for accurately identifying first graders’ need for support in mathematics. We use the results from the digital app DIDUNAS, which runs on typical tablet computers and PCs, to support teachers with overviews over their students’ skill profiles and needs for support. Future work will explore real-time adaptive feedback to improve user experience, longitudinal studies to assess long-term learning outcomes and applications across additional grade levels.