Manifold representation learning for Alzheimer’s disease detection with EEG-based functional brain connectivity

26/08/2026

AI Summary

CAE-Transformer framework combines a convolutional autoencoder with Transformer-based manifold regularization to learn low-dimensional Grassmann manifold representations of EEG functional networks.
CAE-Transformer improved intra-group similarity and inter-group diversity, achieving 98.50% accuracy and 99.49% sensitivity in subject-wise five-fold cross-validation.
Low-dimensional latents preserved AD-related connectivity, with alpha band functional networks yielding the highest diagnostic performance.

TL;DR (AI Generated)

In this study, we proposed a deep manifold representation learning network to characterize the alteration of functional brain connectivity for AD detection.

The proposed framework used a convolutional autoencoder (CAE) to infer low-dimensional manifold representations of functional brain networks reconstructed by phase correlation analysis of electroencephalogram (EEG) signals.

Experimental results with EEG data showed that CAE-Transformer model could map high-dimensional functional brain networks into low-dimensional representations on Grassmann manifold and improved intra-group similarity and inter-group diversity.

Basic summary

Neuroscience. 2026 Aug 25:S0306-4522(26)00549-X. doi: 10.1016/j.neuroscience.2026.08.020. Online ahead of print.

ABSTRACT

Alzheimer’s disease (AD) is a common neurodegenerative disease characterized by severe cognitive dysfunctions and brain disorders. The emergence of deep learning methods provides a feasible way to develop effective representations for clinical diagnosis of brain diseases. In this study, we proposed a deep manifold representation learning network to characterize the alteration of functional brain connectivity for AD detection. The proposed framework used a convolutional autoencoder (CAE) to infer low-dimensional manifold representations of functional brain networks reconstructed by phase correlation analysis of electroencephalogram (EEG) signals. To enhance the global associations among latent features, a Transformer-based manifold regularization module was designed to optimize the structure of the learned low-dimensional manifold representations. Experimental results with EEG data showed that CAE-Transformer model could map high-dimensional functional brain networks into low-dimensional representations on Grassmann manifold and improved intra-group similarity and inter-group diversity. Compared with other advanced methods, CAE-Transformer demonstrated better performance in AD detection, achieving an average accuracy of 98.50% and a sensitivity of 99.49% under subject-wise five-fold cross-validation on a single-center EEG dataset containing 2800 samples derived from segmented recordings of 51 participants. The low-dimensional latent representations learned by CAE-Transformer preserved AD-related brain connectivity patterns. Moreover, functional brain networks in the alpha frequency band showed the highest diagnostic performance for AD detection. Our findings showed the variation of brain networks in AD and provided an efficient method for identifying AD patients.

PMID:42641910 | DOI:10.1016/j.neuroscience.2026.08.020

Document this CPD