Introduction
Online learning environments often suffer from high dropout rates because digital systems lack the ability to "sense" when a student is drifting off or becoming frustrated. While physiological sensors like EEG headsets provide accurate measures of cognitive state, they are intrusive and impractical for widespread home use. This research explores the feasibility of using standard webcams to extract interpretable physical features—such as facial landmarks and head movements—to predict student attention levels, validated against EEG "ground truth" data.
Methodology
The proposed system utilizes a multi-stage pipeline to transform raw video into attention predictions. First, 51 facial landmarks and head pose parameters (roll, pitch, yaw) are extracted from each frame to define the student's physical orientation.
From these landmarks, the researchers engineered several interpretable features:
- Eye Aspect Ratio (EAR): Calculated from eyelid distances to detect blinks and eye-closure frequency.
- Head Dynamics: Using "nose length" as a proxy for the student's distance from the screen, indicating whether they are leaning in (focus) or back (distraction).
- Mouth Activity: Measuring lip dimensions to detect vocalization or emotional expressions.
The study compared a Random Forest regressor against a linear regressor and a baseline model (mean attention value). Random Forest was selected for its capacity to model the complex, non-linear relationships between physical cues and internal attention.
Experimental Setup
The dataset comprised 56 problem-solving sessions (176 minutes total) from seven middle school students (ages 15–16). Ground truth attention values were obtained using commercial EEG headsets. To ensure generalizability, the models were evaluated using "leave-one-student-out" and "leave-one-question-group-out" cross-validation techniques.
Results and Discussion
The Random Forest model significantly outperformed both the baseline and linear regression across all validation methods, demonstrating the effectiveness of the engineered features.
Key findings include:
- Error Reduction: The proposed approach achieved a 29% reduction in Root Mean Squared Error (RMSE) compared to the baseline (12.66 vs. 17.82).
- Interpretability: By using specific features like EAR and head trajectory rather than "black box" deep learning representations, the model provides actionable insights for educators (e.g., identifying distraction through increased blinking and leaning back).
- Non-Linearity: The superior performance of Random Forest suggests that engagement manifests through a complex interplay of physical signals rather than simple linear correlations.
Conclusion and Future Work
While limited by a small sample size and reliance on proprietary EEG algorithms, this study demonstrates that non-intrusive visual monitoring can effectively proxy cognitive attention. Future research will focus on expanding the dataset and exploring "feature fusion" by integrating video data with keystroke dynamics or other modalities to enhance prediction robustness.
Summary updated