Stress Detection Using Hybrid Deep Learning Models with Attention Mechanisms: A Comparative Study of CNN-LSTM, CNN-GRU, and Ensemble Approaches

Accurate and reliable stress detection remains a critical challenge in health monitoring due to the multifaceted nature of stress and the difficulty in capturing its temporal and spatial characteristics from physiological data. Existing methods often lack the ability to effectively model these dependencies, leading to suboptimal performance and limited interpretability, which hinder their application in real-world scenarios such as wearable devices and mobile health systems. This study addresses these limitations by investigating hybrid deep learning models with attention mechanisms, specifically focusing on CNN-LSTM, CNN-GRU, and CNN-BiLSTM architectures and their ensemble. Leveraging the complementary strengths of convolutional and recurrent layers, these models aim to capture both spatial and temporal dependencies in stress-related data, while attention layers enhance interpretability by prioritizing relevant features. Experimental results reveal that the CNN-LSTM with Attention model achieved the best performance, with the lowest Mean Squared Error (MSE) and Mean Absolute Error (MAE), demonstrating its suitability for complex stress prediction tasks. The CNN-GRU model also performed well, offering a balance between computational efficiency and accuracy, while the CNN-BiLSTM model showed limitations, suggesting that additional model complexity may lead to overfitting. The ensemble model, combining predictions from all three architectures, delivered stable performance across metrics, underscoring the value of ensemble approaches in improving robustness and mitigating model-specific biases. These findings have significant implications for practical applications, such as wearable devices and mobile health systems, where accurate, interpretable, and reliable stress monitoring is essential for timely interventions. Future work should focus on optimizing these models for real-time deployment, exploring adaptive learning for personalized stress detection, and validating across diverse datasets to enhance generalizability. This research highlights the importance of hybrid architectures and attention mechanisms in addressing the challenges of stress detection, paving the way for responsive and user-centered health monitoring systems.