A Study on Real-time Early Warning and Adaptive Intervention for Online Learning Burnout Using Multimodal Temporal Feature Fusion

Abstract

 INTRODUCTION: The lack of physical presence in online learning makes it difficult for teachers to perceive students' cognitive and emotional states in real time, with learning burnout being a particularly prominent issue. Existing research primarily relies on single-modality data or lagged learning analysis, making it difficult to achieve precise and timely burnout early warning and intervention.

OBJECTIVES: This paper proposes an online learning burnout early warning model based on multi-modal temporal feature fusion and designs a hierarchical adaptive intervention mechanism.

METHODS: First, utilizing lightweight convolutional neural networks and temporal encoders, facial expression features and body posture features are extracted in real-time from the video stream, respectively. Second, a multi-modal temporal fusion module based on attention mechanisms is designed to model the coordinated temporal evolution of facial expressions and body postures, enabling precise identification of fatigue states such as "confusion", "fatigued" and "bored".  Finally, a hybrid decision model combining rule-based and reinforcement learning approaches is developed. This model dynamically triggers personalized intervention strategies—such as content adjustment, suggested breaks, and interactive Q&A sessions—based on the type, intensity, and duration of fatigue.

RESULTS: Experiments conducted on a self-built authentic online learning dataset demonstrate that this system achieves an F1 score of 0.91 in burnout state recognition, which significantly outperforms unimodal approaches. Furthermore, user studies validate the effectiveness and high acceptance of its intervention mechanism.

CONCLUSION: This study offers a viable technical and practical approach for enabling precise, timely detection and adaptive intervention of burnout in online learning environments.