Attention based hybrid deep learning model for wearable based stress recognition

Abstract

Stress recognition is the process of identifying and assessing an individual's physiological and psychological responses to stressors, which has significant implications for human well-being. Artificial intelligence plays a pivotal role in stress diagnosis by leveraging advanced algorithms to analyse complex physiological data and uncover subtle stress indicators. Previous studies have predominantly employed conventional methods for stress recognition, such as handcrafted features and machine learning algorithms, but these approaches may lack precision and robustness. The proposed method in this research addresses these limitations through a hybrid deep learning model with an attention mechanism, enabling comprehensive feature extraction and dynamic information prioritization. This novel approach enhances stress recognition by accurately fusing multiple physiological modalities and capturing both short-term and long-term patterns associated with stress. The input to the proposed model are physiological signals such as electrocardiogram (ECG) and electrodermal activity (EDA), which are normalized and preprocessed. Afterwards, hybrid CNN-LSTM model leverages the strengths of both convolutional and recurrent networks, enabling it to extract features from the preprocessed data. Then the inclusion of an attention mechanism layer further enhances the model's capability to dynamically weigh and prioritize features from different modalities, enhancing the model's ability to capture key stress-related patterns. Finally, the proposed model involves utilizing the trained model to categorize stress based on the fused features and attention-weighted inputs, achieving improved performance. Experimental findings showcase the effectiveness of the proposed model, achieving an accuracy of 92.70% and a weighted F1-score of 90%. These results outperform the CNN, CNN-LSTM, and decision-based fusion methods.