Scholarlink

<p>Epilepsy is a common neurological disease, and in some patients, abnormal changes in brain activity typically begin before the onset of a seizure. Electroencephalography (EEG) is a practical method for recording electrical activity of brain and plays a crucial role in the diagnosis of epilepsy. Previous studies relied on multi-channel EEG signals and large deep neural networks, which require powerful hardware and reduce user convenience. In this study, teacher-student based knowledge distillation technique on deep neural networks is employed to find the best single EEG electrode to improve user convenience and decrease the network complexity. The teacher model employes Mel-spectrograms of all EEG electrodes as input of 3D convolutional layer of neural network. Subsequently, using response-based knowledge distillation techniques the information from 22 input electrodes of teacher model is transferred to the student model utilized one electrode. The proposed teacher and student models were evaluated on the CHB-MIT dataset, which contains scalp EEG signals recorded using 22 electrodes in a bipolar montage from 24 patients. The experimental results enabled us to identify two electrodes with superior performance (electrode no. 20 achieved accuracy of 84.49%, sensitivity of 86.85%, specificity of 82.58%, and an F1-score of 83.58%, while electrode no. 22 reached accuracy of 84.30%, sensitivity of 86.45%, specificity of 82.93%, and an F1-score of 83.35%). Both electrodes demonstrated the capability to predict seizure onset 30&#xa0;min prior to its occurrence and have superior performance in terms of accuracy, sensitivity, specificity, and F1-score.</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

Efficient epileptic seizure prediction using single channel EEG signal and knowledge distillation on deep neural networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月25日

查看全文
  • Hana Nyamoradi,
  • Abdolhossein Fathi

摘要

Epilepsy is a common neurological disease, and in some patients, abnormal changes in brain activity typically begin before the onset of a seizure. Electroencephalography (EEG) is a practical method for recording electrical activity of brain and plays a crucial role in the diagnosis of epilepsy. Previous studies relied on multi-channel EEG signals and large deep neural networks, which require powerful hardware and reduce user convenience. In this study, teacher-student based knowledge distillation technique on deep neural networks is employed to find the best single EEG electrode to improve user convenience and decrease the network complexity. The teacher model employes Mel-spectrograms of all EEG electrodes as input of 3D convolutional layer of neural network. Subsequently, using response-based knowledge distillation techniques the information from 22 input electrodes of teacher model is transferred to the student model utilized one electrode. The proposed teacher and student models were evaluated on the CHB-MIT dataset, which contains scalp EEG signals recorded using 22 electrodes in a bipolar montage from 24 patients. The experimental results enabled us to identify two electrodes with superior performance (electrode no. 20 achieved accuracy of 84.49%, sensitivity of 86.85%, specificity of 82.58%, and an F1-score of 83.58%, while electrode no. 22 reached accuracy of 84.30%, sensitivity of 86.45%, specificity of 82.93%, and an F1-score of 83.35%). Both electrodes demonstrated the capability to predict seizure onset 30 min prior to its occurrence and have superior performance in terms of accuracy, sensitivity, specificity, and F1-score.