A Mental Stress Classification Method Based on Feature Fusion Using Physiological Signals

Abstract

Mental stress can cause a range of mental health issues, which makes it challenging to develop a stress classification method especially based on physiological signals. Although cutting-edge deep learning models are currently popular for recognizing mental stress, most frameworks rely solely on deep features, which may not provide a comprehensive understanding of physiological signals. In response to this concern, we propose a mental stress classification method that uses feature fusion. We integrate the squeeze-excitation attention mechanism and voting classifier technique to learn detailed and typical information about mental stress. To be more precise, the feature fusion segment consists of two steps: we first extract shallow statistic features and deep features separately from raw signal recordings, and the deep features are dimensionally reduced using principal component analysis to enable better integration with the shallow features. We then flatten both kinds of features and concatenate them by column to create a combined set that contains more salient information about physiological signals. Our experiments show that the attention mechanism and voting classifier technique improve the accuracy of stress classification. Furthermore, our proposed model based on feature fusion achieves remarkable performance compared to state-of-the-art methods.