A deep learning approach to estimating interaction levels in face-to-face lessons

Interaction is a key component of effective learning, fostering active participation and deeper understanding. Real-time insights into classroom interaction help instructors adapt their teaching strategies, while longitudinal data inform the design of improved learning activities. However, many educators remain unaware of the interaction levels during lessons. This study proposes an AI-powered algorithm capable of estimating classroom interaction levels every few seconds during face-to-face instruction. The method relies on a fine-tuned YOLOv8 model for detecting nonverbal interaction cues, followed by postprocessing and data fusion to compute interaction proportions relative to the number of students, ensuring full anonymity. The algorithm classifies behaviors into four interaction categories—student–professor, student–student, student–object, and no-interaction—and outputs a composite interaction score. The detection model achieved high accuracy (average precision, AP@0.5> 92 % across all categories and mean average precision mAP@0.5 = 96 %). Validation with ten experienced professors who rated 100 classroom images revealed poor intraclass correlation (ICC = 0.209), underscoring the subjective nature of “interaction.” Statistical comparisons showed no significant differences between the algorithm’s estimates and expert ratings, though equivalence testing (TOST) did not confirm statistical equivalence for any comparison. The closest alignment occurred between the system’s student–student ratio and the experts’ median evaluations. Two case studies further illustrated the algorithm’s sensitivity to pedagogical context, capturing higher interaction levels during teamwork and workshop sessions. These findings demonstrate the potential of data-driven analytics to support reflective teaching and adaptive learning design.