Empowering Cross-Patient Epilepsy Diagnosis from Diverse-Sampling Low-Quality EEG Signals

Electroencephalograms (EEGs) are essential tools for detecting epilepsy using artificial intelligence, despite their inherent variability in format and fidelity. Most studies focus on patient-specific, high-quality EEGs with fixed sampling rates. However, the real-world challenge lies in diagnosing epilepsy across diverse patients with varying sampling rates, low-quality signals, and incomplete data. To tackle these challenges, we introduce the CPEDNet model, designed for Cross-Patient Epilepsy Diagnosis using EEGs that may have diverse sampling, missing data, high noise, insufficient labels, and volatility. Specifically, we employ a latent Neural Ordinary Differential Equation (NODE) method to enhance EEG signals, mitigating the challenges posed by irregular sampling, missing data, and outliers. Then, we represent the enhanced EEG signals as brain network flows, capturing neural representations across spatial and temporal dimensions. Furthermore, we integrate a score-based, two-stage self-supervised strategy to strengthen multi-channel EEG feature learning and ensure stability in the temporal dynamics of the brain network flow. Extensive testing on two real-world datasets shows that our CPEDNet surpasses current methods, effectively learning spatiotemporal patterns from diverse EEG formats. Abundant ablation studies validate the effectiveness of each module within our model.