KRAFS-ANet: A novel framework for EEG-based stress classification using channel selection and optimized ensemble stacking

Mental stress poses a widespread societal challenge, impacting daily routines and contributing to severe health problems. The earlier studies have utilized Electroencephalograms (EEG) for stress classification; however, the computational demands of processing data from numerous channels often hinder the translation of these models to wearable devices. This paper proposes KRAFS-ANet, a novel framework designed for enhanced stress classification using EEG data on wearable devices. KRAFS-ANet framework incorporates two major novel components to achieve high accuracy with a lightweight design: (1) it strategically employs channel selection using Normal Mutual Information and Recursive Feature Elimination (NMI+RFE) to identify the most informative channels and (2) it uses ensemble stacking techniques integrate bagging K-Nearest Neighbour (KNN), bagging Random Forest (RF), and bagging Support Vector Machine (SVM) with an Artificial Neural Network (ANN) meta-classifier. The study conducts comprehensive experiments on the stress-based MAT dataset and further validates the framework on the stress-based SAM40 and anxiety-based DASPS datasets to demonstrate its effectiveness. KRAFS-ANet achieved the highest accuracies and F1-scores of 98.63% and 98.82% on the MAT dataset, 97.25% and 97.24% on the SAM40 dataset, and 94.92% and 95.15% on the DASPS dataset, respectively. This framework advances the practical application of EEG-based stress detection for portable devices such as wearables and smartphones. It enables real-time monitoring and interventions to enhance mental health in daily life, thus proving its efficacy in real-world scenarios.