Scholarlink

<p>Epilepsy ranks as the second most prevalent neurological disorder globally. Current EEG-based seizure detection methods struggle with inter-patient variability. While unsupervised domain adaptation (UDA) techniques can address this, existing approaches require source patient data access, violating privacy requirements. We propose a privacy-preserving UDA framework for cross-patient seizure detection that: (1) quantifies cross-patient domain discrepancy via distribution uncertainty distance, (2) enables channel-wise knowledge transfer through EEG channel-wise transferability assessment, (3) employs confidence-weighted pseudo-labeling based on distance-based confidence metric. Our method achieves an accuracy of 94.81%, a false detection rate of 0.39, and an average detection delay of 3.29 seconds on the CHB-MIT dataset. On the Siena dataset, an accuracy of 94.1% and a sensitivity of 90.77% are obtained. The results validate the effectiveness of our approach in achieving robust cross-patient seizure detection under privacy-preserving constraints.</p>

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

Cross patient seizure detection via unsupervised domain adaptation based on uncertainty estimation

  • 期刊论文

  • 发表日期：

    2025年7月18日

查看全文
  • Shuai Wang,
  • Hongbin Lv,
  • Hailing Feng,
  • Hao Peng,
  • Wenqian Feng,
  • Chenxi Nie,
  • Yanna Zhao

摘要

Epilepsy ranks as the second most prevalent neurological disorder globally. Current EEG-based seizure detection methods struggle with inter-patient variability. While unsupervised domain adaptation (UDA) techniques can address this, existing approaches require source patient data access, violating privacy requirements. We propose a privacy-preserving UDA framework for cross-patient seizure detection that: (1) quantifies cross-patient domain discrepancy via distribution uncertainty distance, (2) enables channel-wise knowledge transfer through EEG channel-wise transferability assessment, (3) employs confidence-weighted pseudo-labeling based on distance-based confidence metric. Our method achieves an accuracy of 94.81%, a false detection rate of 0.39, and an average detection delay of 3.29 seconds on the CHB-MIT dataset. On the Siena dataset, an accuracy of 94.1% and a sensitivity of 90.77% are obtained. The results validate the effectiveness of our approach in achieving robust cross-patient seizure detection under privacy-preserving constraints.