Multi-Band EEG Spectrogram Decomposition with Residual Attention Network for Enhanced Stress Classification

Stress can affect both physical and mental health, and it is important to support faster intervention. EEG can record brain activity directly, but EEG signals are complex, noisy, and difficult to handle. This study explores how EEG spectrogram decomposition can improve stress classification accuracy using deep learning models. Decomposition was accomplished by splitting a single EEG spectrogram into five distinct segments based on frequency range. Deep neural networks resembling ResNet are well-suited for spectrogram data, as the iterative feature extraction across layers facilitates the identification of hidden patterns. Incorporating an attention module before the classification layer further strengthens the model's capabilities by highlighting the most pertinent features during the training process. The baseline architecture employed in this study was ResNet-152. By incorporating a Multi-Head Attention mechanism prior to the Fully Connected layer, the modified network is denoted as RAN-152. The combination of spectrogram decomposition, ResNet, and attention has been proven to improve classification accuracy in complex EEG data. Without these three together, the accuracy obtained was only 0.5479, while the combination of the three achieved the highest accuracy of 0.9026. Evaluation of other metrics such as precision, recall, and F1-score also confirms that the attention module is good enough to highlight important features while reducing noise, thereby making classification more balanced across classes. These findings show that the combination of EEG decomposition and attention can be a promising approach for stress detection.