A Robust Transformer-based Fully Connected Neural Network Approach for Multilevel Stress Classification

Early identification of mental disorders, including stress, is critical as these conditions rank among the leading causes of death globally. Timely and efficient medical intervention is essential yet traditional machine-learning approaches for classifying stress levels are often complex and require further development. To confront this issue, we propose a novel two-stage hybrid machine learning approach for multi-level stress detection, utilizing an aggregate feature selection technique. Our dataset consists of raw ECG signals from ten healthy subjects collected under relaxed conditions and two stress levels, which are pre-processed by applying a combined filter and subjected to three widely recognized feature selection methods: least absolute shrinkage selection operator (LAS), information gain system (IGS), and minimum-redundancy-maximal-relevance (MRMR). The proposed model aggregates the strengths to choose the optimal feature subset. By analyzing the correlation and normality of the selected features, we propose a novel two-stage machine-learning method that leverages the capabilities of base classifiers. In the meta stage, a transformer-based classifier demonstrates a promising and robust performance, achieving a test accuracy of 94.12%. This outcome demonstrates the effectiveness of combining the advantages of base classifiers. This study will inspire more AI researchers to look into different stress detection methods.