Biomedical Signal Processing in EEG Analysis

Abstract

Electroencephalography (EEG) is the most widely used non-invasive neural recording modality, providing millisecond-resolution temporal dynamics of cortical activity for epilepsy diagnosis, brain-computer interfaces, sleep staging, and cognitive neuroscience. However, EEG signals are inherently noisy, non-stationary, and high-dimensional, requiring sophisticated signal processing pipelines for clinically reliable interpretation. This study developed and benchmarked a comprehensive EEG processing and classification framework comparing five approaches -- bandpass filtering with common spatial patterns (CSP), wavelet packet decomposition (WPD), Riemannian geometry on covariance matrices, EEGNet (compact CNN), and a proposed hybrid attention-augmented temporal convolutional network (ATCNet) -- across three clinical EEG tasks: motor imagery classification (BCI Competition IV-2a, n = 9), epileptic seizure detection (CHB-MIT, n = 23), and automated sleep staging (Sleep-EDF, n = 153). ATCNet achieved the highest accuracy for motor imagery (84.8 +- 6.2%; kappa 0.798) and seizure detection (sensitivity 96.4%, FPR 0.042/h), while Riemannian geometry provided the best cross-session transfer (accuracy drop < 3.2% between sessions). For sleep staging, ATCNet achieved 86.4% accuracy (Cohen's kappa 0.812) across 5 AASM classes, outperforming all baselines. Artifact rejection by independent component analysis (ICA) with ICLabel automated classification removed 94.2% of ocular and muscular artifacts while preserving 98.4% of neural signal power. These results establish ATCNet as the new state-of-the-art for multi-task EEG classification and provide a reproducible, open-source processing pipeline for clinical EEG applications.