STRESS AND ANXIETY CLASSIFICATION BASED ON PHYSIOLOGICAL SIGNALS USING MACHINE LEARNING

Abstract

Stress and anxiety significantly affect cognitive, emotional, and physiological functions. Physiological signals such as Electroencephalogram (EEG), Electrocardiogram (ECG), Electrodermal Activity (EDA), Blood Volume Pulse (BVP), and Respiration (RESP) provide objective, real-time insights compared to subjective assessments. Machine learning-based systems have recently gained attention for classifying stress and anxiety from high-dimensional physiological data, enabling real-time interventions through wearable devices. However, challenges such as redundant features, class imbalance, and high computational complexity continue to limit their scalability and practical use. To address these challenges, this thesis focuses on the following key research gaps: (1) lack of optimized channel selection in EEG-based stress classification frameworks, (2) absence of systematically collected multimodal datasets reflecting progressive stress in academic environments, (3) need for scalable and interpretable feature selection algorithms, and (4) limited exploration of reinforcement learning strategies for stress detection using physiological signals. Firstly, this thesis presents a comprehensive literature review on automated stress and anxiety classification using physiological signals. It covers stress and anxiety theory, physiological signals, and their relationship with mental states, preprocessing methods, domain-specific feature extraction, feature selection techniques, and machine learning models. Secondly, the thesis proposes an ensemble-based EEG stress classification framework for EEG signal-based wearable applications. It introduces the KRAFS- ANet framework, which stacks Bagging K-Nearest Neighbor, Bagging Support Vector Machine, and Bagging Random Forest classifiers, with an Artificial Neural Network (ANN) as the meta-classifier. The approach incorporates optimized channel selection vi and ensemble stacking to improve accuracy while reducing computational time. The thesis validates the framework on three benchmark datasets: MAT, SAM40, and DASPS. It achieves accuracies of 98.63%, 97.25%, and 94.92%, respectively, along with consistently high F1-scores. However, this work does not address multimodal signal fusion, class imbalance, or feature interdependencies. Thirdly, this thesis introduces the Academic Stress Dataset (ASD), where physiological signals such as Interbeat Interval (IBI), BVP, and EDA are recorded during the Montreal Imaging Stress Task (MIST) to induce progressive mental arithmetic stress in engineering students. This work applies a hybrid feature selection to select the most informative features and further utilizes Bayesian optimization for hyperparameter tuning. The Gradient Boosting model achieves accuracy of 98.28% and 97.02% for 2-level and 3-level classification, respectively. Using only EDA and HRV features provides comparable accuracy, and SHAP-based Explainable AI (XAI) analysis further confirms them as the most informative features. Finally, this thesis proposes and compares two efficient and effective feature selec- tion algorithms. The first, CorLMI-FSA, combines Correlation, Logistic Regression (LR), and Mutual Information (MI) to reduce redundancy in stress classification using EDA and HRV features from the self collected ASD dataset. It achieves highest accuracy of 96.82% for binary and 95.84% for three-level classification. The second, ST-CIRL (SMOTETomek-Correlated Interactive Reinforcement Learning), addresses class imbalance and optimizes feature selection through interactive reinforcement learning. ST-CIRL utilizes ECG, EDA, and RESP features from the Spider-phobic anxiety dataset, applies Optuna optimization, and achieves the highest accuracy of 95.35% and an F1-score of 95.49% using LightGBM, outperforming existing methods. Cross-dataset evaluation shows that CorLMI performs better in binary classification with lower runtime, while ST-CIRL achieves higher accuracy in multi- class classification. This thesis advances intelligent, wearable stress monitoring systems and lays the foundation for real-time health assessment, stress-aware learning environments, and Human-Computer Interaction applications. It outlines future directions to enhance stress and anxiety detection by improving model generalizability, enabling real-time implementation, and integrating personalized interventions.