Scholarlink

Transfer learning in dementia research uses advanced techniques to generalize knowledge from, for example, Alzheimer’s disease (AD) to frontotemporal dementia (FTD) using EEG data, potentially through transformer-based models to enhance this process. This study explores the feasibility and effectiveness of transferring insights gained from AD versus healthy EEG patterns to classify FTD versus healthy EEG patterns. The transformer model’s ability to capture complex temporal relationships in EEG signals is harnessed to adapt learned representations to a different type of dementia. Our experiments demonstrate promising results in cross-domain transfer, where the model trained on AD-related features achieves competitive performance in distinguishing FTD cases from healthy controls. Using only 10% of available FTD sample , transfer learning from AD data yields a 17.16% error rate (F1 score: 86.32) and increasing this further to only 30% significantly reduces the error rate to 12.29% (F1 score: 90.07). This approach may not only enhance diagnostic accuracy but also provide insights into shared and distinct neurophysiological markers across dementia subtypes. The findings underscore the potential of transformer-based transfer learning to facilitate early detection and differential diagnosis of related neurodegenerative diseases using EEG data.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From Alzheimer’s Disease to Frontotemporal Dementia: Transfer Learning in EEG-Based Diagnosis of Dementia

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Paulino Salmon,
  • Jaymar Soriano

摘要

Transfer learning in dementia research uses advanced techniques to generalize knowledge from, for example, Alzheimer’s disease (AD) to frontotemporal dementia (FTD) using EEG data, potentially through transformer-based models to enhance this process. This study explores the feasibility and effectiveness of transferring insights gained from AD versus healthy EEG patterns to classify FTD versus healthy EEG patterns. The transformer model’s ability to capture complex temporal relationships in EEG signals is harnessed to adapt learned representations to a different type of dementia. Our experiments demonstrate promising results in cross-domain transfer, where the model trained on AD-related features achieves competitive performance in distinguishing FTD cases from healthy controls. Using only 10% of available FTD sample , transfer learning from AD data yields a 17.16% error rate (F1 score: 86.32) and increasing this further to only 30% significantly reduces the error rate to 12.29% (F1 score: 90.07). This approach may not only enhance diagnostic accuracy but also provide insights into shared and distinct neurophysiological markers across dementia subtypes. The findings underscore the potential of transformer-based transfer learning to facilitate early detection and differential diagnosis of related neurodegenerative diseases using EEG data.