Spatiotemporal Self-Attentive Graph-TCN Framework for Enhanced Teen Stress Detection

Abstract

Timely and precise detection of adolescent stress is critical for early intervention, yet current CNN- or Transformer-based models fail to jointly capture both structural facial dependencies and temporal emotion dynamics. To address this gap, we propose a novel spatiotemporal self-attentive graph-TCN (STG-TCN) framework that integrates graph-based spatial reasoning with temporal convolutional modeling. Unlike conventional approaches that concatenate spatial and temporal features, our method employs a cross-attention fusion mechanism to align and enrich these modalities. The proposed model was evaluated on the dataset for affective states in e-environments (DAiSEE) under strict cross-subject protocols, achieving 90.2% classification accuracy, a 12% improvement over graft attention network (GAT)-only baselines, and a 9% improvement over temporal convolutional networks (TCN)-only baselines. Furthermore, STGTCN reduced false negatives by 20%, demonstrating high sensitivity to subtle stress cues such as micro-expressions and transient muscle contractions. These findings establish STG-TCN as the first framework specifically tailored for adolescent stress recognition in real-world video settings, thereby advancing the field of affective computing with both methodological innovation and practical implications for unobtrusive mental health monitoring.