EEG-MACS: Manifold Attention and Confidence Stratification for EEG-based Cross-Center Brain Disease Diagnosis under Unreliable Annotations

A team of medical researchers and data scientists introduced a novel deep learning framework called EEG-MACS in a revised technical paper published on the arXiv preprint server on May 15, 2024, to improve the accuracy of cross-center neurodegenerative disease diagnosis. This technological advancement addresses the long-standing problem of 'noisy' data where brain signal recordings and medical labels vary significantly across different hospitals and clinics. By implementing Manifold Attention and Confidence Stratification, the researchers aim to filter out unreliable annotations and overcome the heterogeneity inherent in Electroencephalogram (EEG) data collected from diverse clinical settings. The core challenge in diagnosing neurodegenerative diseases via EEG lies in the subtle nature of abnormal neural dynamics. Unlike more obvious brain traumas, the signatures of early-stage cognitive decline are often faint and easily obscured by mechanical interference or differences in recording equipment used at different medical centers. Traditionally, AI models trained on data from one hospital struggle to maintain performance when applied to another. EEG-MACS solves this by focusing on the underlying data manifold, allowing the system to distinguish between genuine physiological signals and artifacts caused by variations in data collection environments. Beyond data heterogeneity, the framework specifically targets the issue of annotation unreliability. In complex neurological cases, human experts may disagree on a diagnosis, leading to training labels that are inconsistent or partially incorrect. The 'Confidence Stratification' component of the new framework effectively ranks the reliability of these expert annotations during the learning process. By prioritizing high-confidence data points while carefully managing uncertain ones, the model achieves a level of diagnostic robustness previously difficult to attain in small-group clinical settings. This breakthrough paves the way for more reliable, AI-assisted screening tools that can be deployed globally, regardless of local variations in medical hardware or diagnostic standards.