Deep Learning-Based Classification of Cognitive and Resting States Using Electroencephalography Signals

This paper investigates the challenge of distinguishing between a resting brain state and various cognitive task states using electroencephalography (EEG). The authors aim to determine if a deep learning-based pipeline can reliably classify an eyes-open resting condition against three specific mental activities: continuous mathematical subtraction, memory recall, and mentally singing a song. To achieve this, the researchers propose a framework that integrates signal processing—specifically bandpass filtering, Independent Component Analysis (ICA), and Short-Time Fourier Transform (STFT)—with a hybrid CNN-GRU feature extraction module. This module is designed to capture both local spatial patterns and temporal dependencies within the EEG data. Finally, they introduce a custom 2D-Net architecture to perform the binary classification.

The proposed 2D-Net architecture, when combined with the CNN-GRU feature extraction pipeline, consistently outperformed standard machine learning and deep learning baselines across all three tasks. The framework achieved classification accuracies of 83.177% for resting-versus-mathematical tasks, 76.107% for resting-versus-memory tasks, and 83.432% for resting-versus-music tasks. The authors demonstrate that the CNN-GRU feature extraction stage significantly boosts the performance of various classifiers, including SVM, XGBoost, and Decision Trees, suggesting that the learned representations are highly effective for identifying task-related neural signatures.

Understanding how brain activity shifts between resting and active cognitive states is fundamental for developing robust brain-computer interfaces (BCIs) and neurofeedback systems. By demonstrating that a combined signal-processing and deep learning approach can extract informative features from EEG signals to differentiate these states, this research provides a scalable methodology for real-time cognitive state monitoring. The results highlight the potential of integrating time-frequency analysis with deep learning to improve the accuracy of mental state classification in non-invasive clinical and research settings.