DGNet: Self-Supervised Delta2Gamma Multi-Band EEG Representation Learning for Dementia Classification

Chanwoo Park, Chanwoo Kim

Keywords: Band-head, self-supervised learning, EEG, regularization, adaptive temperature, dementia, representation learning

TL;DR: We propose a novel multi-head contrastive learning model that uses EEG frequency bands to improve early detection and monitoring of dementia from EEG signals.

Abstract: As the global population ages and dementia cases rise, there is an urgent need for effective early diagnosis and monitoring of neurodegenerative diseases. Electroencephalogram (EEG)-based technologies are increasingly important due to their portability, affordability, and suitability for widespread screening compared to other neuroimaging methods. However, EEG signals present challenges such as low signal-to-noise ratio, high inter-subject variability, and limited labeled data, especially in elderly or dementia patients, which restricts the effectiveness of traditional supervised learning approaches. Leveraging the neurophysiological significance of the five EEG frequency bands (delta, theta, alpha, beta, gamma), this study introduces an innovative multi-head Simple Framework for Contrastive Learning of Visual Representations (SimCLR) architecture. The proposed Delta2Gamma (DGNet) model combines frequency-band specific representation learning, enabling more precise detection of subtle EEG changes linked to brain disorders like dementia. Our self-supervised learning (SSL) adaptive multi-band heads model achieved a 31.5\% relative performance improvement over training from scratch, and a 25.4\% improvement over the single-head approach. To the best of our knowledge, our proposed method achieved state-of-the-art performance in multi-head approaches. The source code is available at GitHub by https://anonymous.4open.science/r/iclr2026-7FE2.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 2552