Scholarlink

<p>Mental stress is a dangerous health issue with heart diseases, depression, and low productivity, and it should be identified as soon as possible. Conventional unimodal methods are insufficient to understand the complexity of stress that is expressed both physiologically and behaviorally. To do this, we present H-C<sup>3</sup>AT-G (Hierarchical Cross-Modal Contrastive Attention Transformer with Graph Fusion), a new end-to-end multimodal stress detection framework. The model uses contrastive pretraining to align the modalities, uses graph neural networks to fuse adaptive features, and uses a hierarchical cross-attention transformer to learn multi-scale dependencies. Moreover, Monte Carlo dropout improves predictive accuracy, whereas knowledge distillation reduces the model size to be used on a wearable. Tests on benchmark data demonstrate that H-C<sup>3</sup>AT-G performs better than CNN-BiLSTM, multimodal transformers, or GAT-based fusion, with an accuracy of 94.8%, an AUROC of 0.96, and a much lower Expected Calibration Error. The model gets close to real-time inference (8.6 ms) using only 4.2&#xa0;M parameters, making stress monitoring practical.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hierarchical Cross-Modal Contrastive Attention Transformer with Graph Fusion for Multimodal Stress Detection

  • 期刊论文

  • 发表日期：

    2026年2月24日

查看全文
  • MD Jaffar Sadiq,
  • Padmavathi Vurubindi,
  • Sayyada Hajera Begum,
  • Bhanu Chander Balusa,
  • Giasuddin,
  • MD Riyazuddin

摘要

Mental stress is a dangerous health issue with heart diseases, depression, and low productivity, and it should be identified as soon as possible. Conventional unimodal methods are insufficient to understand the complexity of stress that is expressed both physiologically and behaviorally. To do this, we present H-C3AT-G (Hierarchical Cross-Modal Contrastive Attention Transformer with Graph Fusion), a new end-to-end multimodal stress detection framework. The model uses contrastive pretraining to align the modalities, uses graph neural networks to fuse adaptive features, and uses a hierarchical cross-attention transformer to learn multi-scale dependencies. Moreover, Monte Carlo dropout improves predictive accuracy, whereas knowledge distillation reduces the model size to be used on a wearable. Tests on benchmark data demonstrate that H-C3AT-G performs better than CNN-BiLSTM, multimodal transformers, or GAT-based fusion, with an accuracy of 94.8%, an AUROC of 0.96, and a much lower Expected Calibration Error. The model gets close to real-time inference (8.6 ms) using only 4.2 M parameters, making stress monitoring practical.