摘要
Objective: Chronic migraine (CM), characterized by headaches on 15 or more days per month, imposes a substantial burden on daily functioning. Although electroencephalography (EEG) shows promise for detecting subtle migraine-related neural signals, data-driven approaches are often hampered by noisy labels arising from diagnostic bias, annotation errors, data collection challenges, and signal instability, ultimately diminishing data representativeness and limiting their generalizability. Methods: This study presents REAL, a robust EEG analysis and labeling framework integrating (1) global feature extraction via contrastive learning for capturing shared representations, (2) local feature extraction to consolidate within-subject segments, and (3) iterative noisy-label filtering to remove unreliable data. Results: Empirical evaluations indicate that REAL effectively identifies low-confidence samples, presumed to stem from noisy labels, and boosts classification performance, particularly in differentiating CM from healthy controls (CTL). The heterogeneity between low- and high-confidence data has been underscored by label-flipping experiments. Additionally, a marked difference between CTL and CM has been revealed by an entropy-based analysis of high-confidence samples, with higher entropy in CM potentially reflecting an excitatory-inhibitory imbalance and offering deeper insights into CM mechanisms. Conclusion: Comparisons with existing methods underscore REAL's superior ability to handle noisy labels and preserve discriminative features in EEG signals. Significance: These findings highlight REAL's potential to mitigate label noise, advancing EEG-based migraine research and aiding the development of reliable diagnostic tools for clinical practice. By improving data integrity and capturing subtle neural signatures, the effects of noisy labels are mitigated by REAL, thereby supporting accurate migraine detection.