개선된 대조 표현학습을 이용한 EEG 치매 진단 및 채널 기여도 분석 = EEG-based Dementia Diagnosis and Channel Contribution Analysis Using Enhanced Contrastive Representation Learning

다국어 초록 (Multilingual Abstract)

This study proposes Contrastive Multi-scale EEGNet (CM-EEGNet) for electroencephalography (EEG)-based classification of Alzheimer’s disease (AD) and cognitively normal (CN) subjects, as well as channel-level contribution analysis. EEG is a useful non-invasive signal for assisting dementia diagnosis because it reflects functional brain activity with high temporal resolution. However, EEG-based classification remains challenging due to signal nonstationarity, noise sensitivity, and inter-subject variability. To address these issues, the proposed CM-EEGNet extends the conventional EEGNet structure by applying multi-scale temporal convolutions to capture EEG patterns at different time scales and depthwise spatial convolutions to learn inter-channel spatial relationships.
The proposed model consists of an encoder, a projection head, and a classification head. The classification head predicts AD and CN classes, while the projection head generates embeddings for supervised contrastive learning. The model is trained using a joint loss that combines cross-entropy loss and supervised contrastive loss, thereby improving both classification performance and class separability in the embedding space. Experiments were conducted using the OpenNeuro ADFTD EEG dataset. The frontotemporal dementia group was excluded, and only AD and CN subjects were used for binary classification. A subject-independent 60:20:20 split was applied to prevent subject leakage, and 19-channel EEG signals were segmented into non-overlapping 8-second epochs.
In the subject-independent test set, CM-EEGNet achieved a segment-level accuracy of 97.30%, with F1-scores of 97.22% for AD and 97.37% for CN. The ablation study showed that the joint loss setting with =0.5 and =0.5 produced the best performance. In addition, t-SNE visualization confirmed that the learned encoder features and projection embeddings showed separable distributions between AD and CN. For interpretability, this study applied Contrastive Prototype Attribution (CPA), which estimates channel-wise contributions by measuring changes in contrastive evidence scores after channel masking. The CPA results showed that O2, F7, and P3 contributed mainly to AD classification, while T5, O2, and F4 contributed mainly to CN classification. These findings indicate that the proposed method can not only improve EEG-based AD-CN classification performance but also provide an interpretable channel-level basis for model decisions in the contrastive embedding space.

목차 (Table of Contents)

• 제1장 서론 1
• 제1절 연구 배경 및 중요성 1
• 제2절 연구 내용 및 구성 3
• 제2장 관련 연구 4
• 제1절 EEG 기반 치매 진단 연구 4
• 제2절 딥러닝을 이용한 EEG 기반 치매 진단 연구 6
• 제3절 대조 표현학습 기반 연구 9
• 제3장 대조 멀티스케일 EEGNet 11
• 제1절 EEGNet 11
• 제2절 제안된 CM-EEGNet 모델 13
• 제3절 Joint Loss 기반 학습 방법 17
• 제4절 대조 프로토타입 기여도 기반 채널 기여도 분석 19
• 제5절 제안 방법의 전체 절차 22
• 제4장 실험 및 결과 24
• 제1절 실험 데이터셋 및 실험 환경 24
• 제2절 분류 성능 및 결과 27
• 제3절 CPA 기반 채널 기여도 분석 결과 31
• 제5장 결론 34
• 참고문헌 36