EduGage: Methods and Dataset for Sensor-Based Momentary Assessment of Engagement in Self-Guided Video Learning

Engagement is critical for learning, especially in self-guided online video environments where learners struggle with self-regulation. Researchers Zikang Leng, Edan Eyal, Yingtian Shi, Jiaman He, Yaqi Liu, and Thomas Plötz from Georgia Tech developed EduGage to measure engagement at a fine-grained, momentary level using wearable and camera-based sensors. They conducted a user study with 16 participants watching instructional videos and periodically reporting their engagement via brief in-situ probes. The sensing suite captured physiological signals (PPG, ECG, EDA, EEG), motion (IMU), heart rate, temperature, and eye-tracking, providing a rich multimodal stream synchronized with each engagement label.

The team trained models to predict engagement from these signals, achieving a mean absolute error of 0.81 and 83.75% within-1 accuracy on a 1–7 scale, with binary accuracy of 73.93% and Macro-F1 of 68.45%. This performance exceeded sensor-free baselines, statistical models, deep temporal networks, foundation models, and LLM-based approaches. A key finding is that full multimodal instrumentation isn't necessary—lightweight combinations of behavioral and physiological sensors work nearly as well, reducing system complexity. The EduGage dataset, including all synchronized sensor signals, probe labels, video metadata, quizzes, and study materials, is released publicly to support reproducible research on real-time engagement sensing in self-guided learning.