A multimodal deep learning framework for real-time burnout detection and personalized intervention in foreign language teachers

Abstract

Teacher burnout poses significant challenges to educational quality, with foreign language instructors facing distinctive stressors including cross-cultural mediation demands and language anxiety transmission. This study proposes a multimodal deep learning framework for real-time burnout detection and personalized psychological intervention targeting foreign language teachers. The framework integrates facial expression sequences, vocal prosodic features, and heart rate variability signals through modality-specific feature extraction networks, including a modified ResNet-50 for visual processing, bidirectional LSTM for audio analysis, and one-dimensional convolutional networks for physiological signals. A Transformer-based cross-modal attention fusion module dynamically weights modality contributions based on signal reliability and contextual relevance. Experimental validation on a dataset comprising 156 foreign language instructors demonstrates that the proposed model achieves 85.6% classification accuracy with an AUC of 0.912 (averaged over five independent runs), outperforming unimodal baselines and conventional fusion approaches, with the margin over the strongest baseline confirmed as statistically significant. Furthermore, a tiered intervention framework matching cognitive behavioral therapy, mindfulness training, and social support enhancement to individual burnout profiles reduces the composite burnout index by roughly a third to a half depending on the matched modality—from 33% for the social-support pathway to 51% for the combined program—across eight-week follow-up assessments relative to a minimally changed control group. The findings establish an integrated approach bridging computational detection with evidence-based intervention for occupational health management in foreign language education.