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

TL;DR

The paper tackles the challenge of cross-center EEG data heterogeneity and unreliable annotations in neurodegenerative disease diagnosis. It introduces MACS, a transferable framework that fuses Euclidean and Riemannian manifold representations through modules—Augmentor, Switcher, Encoder, Projector, and Stratifier—while enforcing confidence-guided learning. The authors develop a constrained, dual-head learning objective with multi-view contrastive and discriminative losses to handle unreliable labels, and demonstrate subject-independent transferability across PD, MCI, and AD datasets with cross-center validation. These results highlight MACS’s potential to improve cross-center EEG diagnostics under imperfect annotations and suggest broader applicability to heterogeneous multimodal data analysis.

Abstract

Cross-center data heterogeneity and annotation unreliability significantly challenge the intelligent diagnosis of diseases using brain signals. A notable example is the EEG-based diagnosis of neurodegenerative diseases, which features subtler abnormal neural dynamics typically observed in small-group settings. To advance this area, in this work, we introduce a transferable framework employing Manifold Attention and Confidence Stratification (MACS) to diagnose neurodegenerative disorders based on EEG signals sourced from four centers with unreliable annotations. The MACS framework's effectiveness stems from these features: 1) The Augmentor generates various EEG-represented brain variants to enrich the data space; 2) The Switcher enhances the feature space for trusted samples and reduces overfitting on incorrectly labeled samples; 3) The Encoder uses the Riemannian manifold and Euclidean metrics to capture spatiotemporal variations and dynamic synchronization in EEG; 4) The Projector, equipped with dual heads, monitors consistency across multiple brain variants and ensures diagnostic accuracy; 5) The Stratifier adaptively stratifies learned samples by confidence levels throughout the training process; 6) Forward and backpropagation in MACS are constrained by confidence stratification to stabilize the learning system amid unreliable annotations. Our subject-independent experiments, conducted on both neurocognitive and movement disorders using cross-center corpora, have demonstrated superior performance compared to existing related algorithms. This work not only improves EEG-based diagnostics for cross-center and small-setting brain diseases but also offers insights into extending MACS techniques to other data analyses, tackling data heterogeneity and annotation unreliability in multimedia and multimodal content understanding.

Figures (10)

• Figure 1: Overview of the MACS framework and subject-independent experiments across diseases and centers. The Projector has a dual-head structure for latent space representation and classification. The Stratifier categorizes samples by label confidence, constraining brain variants for contrastive learning, and discriminative loss for distrusted samples. Refer to Figures \ref{['fig: Figure3']} and \ref{['fig: Figure4']} for Switcher and Encoder, respectively.
• Figure 2: The Switcheremploys conditional interpolated blending for trusted samples and bypasses distrusted ones to mitigate overfitting on incorrectly labeled samples.
• Figure 3: The Encodercombines Riemannian and Euclidean metrics for feature extraction, leading to a manifold-based attention mechanism that effectively captures characteristics of spatiotemporal complexity and dynamic synchronization.
• Figure 4: Qualitative comparison of MACS with SOTA frameworks using t-SNE visualization based on latent distribution.
• Figure 5: Evaluation of MACS's Cross-center Transferability: (a) Direct testing results of MACS, trained on Center A's PD data applied to Center B's PD data; (b) Fine-tuning testing results of MACS, trained on Center C's MCI data applied to Center D's AD data, using only a limited percentage of labels.