REAL: Robust EEG Analysis and Labeling for Chronic Migraine Patients Amidst Label Noises

Area of Science:

• Neuroscience
• Biomedical Engineering
• Data Science

Background:

• Chronic migraine (CM) significantly impacts daily life, necessitating reliable diagnostic tools.
• Electroencephalography (EEG) shows potential for detecting migraine-related neural signals.
• Noisy labels in EEG data hinder the development of accurate, data-driven diagnostic approaches.

Purpose of the Study:

• To present REAL, a robust EEG analysis and labeling framework designed to overcome the challenges posed by noisy labels.
• To enhance the reliability and generalizability of EEG data for migraine research.
• To improve the accuracy of differentiating chronic migraine from healthy controls using EEG.

Main Methods:

• REAL integrates global and local feature extraction using contrastive learning.
• It employs iterative noisy-label filtering to remove unreliable data segments.
• The framework captures shared representations and consolidates within-subject segments for comprehensive analysis.

Main Results:

• REAL effectively identifies and filters low-confidence samples, presumed to originate from noisy labels.
• Classification performance, especially for differentiating CM from healthy controls (CTL), is significantly boosted.
• Entropy-based analysis of high-confidence samples reveals differences between CM and CTL, suggesting potential insights into CM pathophysiology.

Conclusions:

• REAL demonstrates superior performance in handling noisy EEG labels compared to existing methods.
• The framework preserves discriminative neural features crucial for accurate migraine detection.
• REAL holds significant potential for advancing EEG-based migraine research and developing reliable clinical diagnostic tools.