In an AI-based early warning system, the predictive importance of CAM dimensions shifted over an instructional unit, with affect/metacognition self-reports leading early and cognitive artifacts gaining accuracy toward the end¶

Claim · All claims
Evidence · 2 studies (3 entries) · 1 review r1, 1 associational r2 · q2
Currency · newest study 2026 · checked against its sources 2026-10-22 (automated, 3 of 3 entries) · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Self-reports on affect and metacognition were more accurate in predicting learning productivity at the beginning of the instructional unit, while the predictive accuracy of students' cognitive artifacts increased toward the end of the unit. → Joni Lämsä 2026 q2 i? As this commentary summarises Kubsch et al. (2025), among the machine learning methods that study applied, boosting yielded the highest overall predictive accuracy and showed the least bias in predictive accuracy by gender, though the decision tree model outperformed boosting when accuracy was examined by parental education background. → Joni Lämsä 2026 (2)

q2 i? Models using only affective and/or metacognitive data showed flat accuracy trends over time and outperformed cognitive-only models at P1 and P2, while cognitive-only models yielded higher accuracy at P4 and P5. → Kubsch 2025

Evidence¶

Joni Lämsä 2026¶

Joni Lämsä, Susanne de Mooij, Martine Baars, Inge Molenaar, Sanna Järvelä. (2026). Self-Regulated Learning, Multimodal Data, and Analysis Grid: Where Are We Now and Where Are We Going? Educational Psychology Review. https://doi.org/10.1007/s10648-025-10113-4

q2 · i? · theoretical · r2

The commentary summarizes Kubsch et al.'s (2025) empirical study of an AI-based early warning system in inquiry-based science classrooms, combining workbook log data (cognitive processes) with self-reports (metacognitive and affective processes) across an instructional unit spanning 2–3 weeks. The authors found that "the importance of different CAM dimensions" varied over time.

"While self-reports on affect and metacognition were more accurate in predicting learning productivity at the beginning of the instructional unit (based on students' scores across all unit tasks), the predictive accuracy of students' cognitive artifacts increased toward the end of the unit."

Joni Lämsä 2026 (2)¶

Joni Lämsä, Susanne de Mooij, Martine Baars, Inge Molenaar, Sanna Järvelä. (2026). Self-Regulated Learning, Multimodal Data, and Analysis Grid: Where Are We Now and Where Are We Going? Educational Psychology Review. https://doi.org/10.1007/s10648-025-10113-4

q2 · i? · review · r1

The commentary reports, from Kubsch et al.'s (2025) study (not the commentary authors' own work), that across the range of machine learning methods Kubsch et al. applied, "boosting yielded the highest overall predictive accuracy" and showed the least gender bias in predictive accuracy. However, when predictive accuracy was examined by parental education background, the decision tree model outperformed boosting.

"By applying various machine learn - ing methods, the authors found that boosting yielded the highest overall predictive accuracy. Boosting also showed the least bias regarding predictive accuracy by gen- der."

Kubsch 2025¶

Kubsch, M., Strauß, S., Grimm, A., Gombert, S., Drachsler, H., Neumann, K., & Rummel, N. (2025). Self-regulated Learning in the Digitally Enhanced Science Classroom: Toward an Early Warning System. Educational Psychology Review. https://doi.org/10.1007/s10648-025-10011-9

q2 · i? · associational · r2

Cross-validation results across five timepoints (Fig. 3) comparing unimodal data-source combinations. Affective/metacognitive-only models showed "rather flat time-trends" while models including cognitive data showed increasing accuracy over time.

"Interestingly, the models that used only cognitive data had a lower accuracy at timepoints P1 and P2 compared to the models that use only affective and/or metacognitive data. At time point P3, models that use only cognitive and models that used only affective, and/or metacognitive data, were similar in accuracy."

Discussion¶

Related Claims¶

• Confounding may explain the large retention effects reported for the Course Signals early warning system — related
• An agentic AI instructor sustained and deepened student engagement across successive course modules in a field deployment — related
• Within a single LLM session, students' help-seeking shifts toward answer- and verification-seeking while explicit self-regulation modestly increases — related
• Warning-only LA feedback can increase mental load, whereas pairing warnings with encouragement leads to higher emotional and cognitive engagement — related
• Machine-learning analysis of multimodal classroom indicators can predict student productivity, and prior knowledge becomes relevant later in the regulation process — possibly the same claim (merge candidate)
• Boosting models (XGBoost) achieved the highest predictive accuracy among the compared ML models, while logistic regression showed the lowest — related
• Combining cognitive, affective, and metacognitive data from digital workbooks predicts middle-school students' learning productivity with accuracy rising from about 60% to 100% as an inquiry unit progresses — possibly the same claim (merge candidate)
• Regular and explicit teacher addressing of metacognitive strategies is linked to students applying more metacognitive strategies over time — related
• Cognitive artifact scoring in the digital workbook achieved high intercoder reliability, with average Spearman's rho of 0.97 overall — related