Scholarlink

<p>Efficient epilepsy diagnosis and monitoring of patients require the accurate detection of epileptic seizures based on scalp electroencephalogram (EEG) signals. Nevertheless, EEG variational differences across individuals and the unbalanced nature of seizure and non-seizure events are crucial issues concerning automated detection systems. We are systematically comparing the five supervised machine learning models, Kernelized Support Vector Machine (KSVM), Multi-Layer Perceptron (MLP), K-Nearest Neighbor (KNN), Random Forest, and AdaBoost to their accuracy in patient-specific and non-patient-specific seizure detection on multichannel scalp EEG recordings in the CHB-MIT Scalp EEG database. A full spectrum of time-domain and frequency-domain features was obtained after undergoing a thorough preprocessing process, including filtering, segmentation into 10-second epochs, feature normalization, and feature dimensionality reduction. The true positive rates (TPR) varied among all models between 70% and 90%, whereas the false positive rates (FPR) were less than 0.2%. We used rigorous comparisons between five machine learning methods on patient- and non-patient-specific seizure classification. Kernel SVM performed best among them, recording &gt; 95% TPR using both scenarios of the classification, whereas it was better than a commercial non-patient seizure detector. In addition, we also discovered that the adjustment of class weights proved to be remarkably beneficial in detecting minority classes in highly imbalanced data sets. These results indicate that the use of scalp EEG-based seizure detection systems has potential as a clinical decision support system, especially in epilepsy monitoring units to provide continuous patient monitoring. The results indicate the possible use of advanced supervised learning algorithms, specifically KSVM, to supplement the accuracy of EEG-based seizure detection systems in clinical practice and upcoming wearable practitioner instruments.</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

Supervised Machine Learning Evaluation for Patient-Specific and Non-Patient-Specific Epileptic Seizure Detection with Multichannel Scalp EEG

  • 期刊论文

  • 发表日期：

    2025年7月24日

查看全文
  • Vijay C. P.,
  • Mahendra G.,
  • Bharath K. N.,
  • Suhaas K. P.,
  • Bharath Kumara,
  • Divyashree R. H.,
  • Sudheesh K. V.

摘要

Efficient epilepsy diagnosis and monitoring of patients require the accurate detection of epileptic seizures based on scalp electroencephalogram (EEG) signals. Nevertheless, EEG variational differences across individuals and the unbalanced nature of seizure and non-seizure events are crucial issues concerning automated detection systems. We are systematically comparing the five supervised machine learning models, Kernelized Support Vector Machine (KSVM), Multi-Layer Perceptron (MLP), K-Nearest Neighbor (KNN), Random Forest, and AdaBoost to their accuracy in patient-specific and non-patient-specific seizure detection on multichannel scalp EEG recordings in the CHB-MIT Scalp EEG database. A full spectrum of time-domain and frequency-domain features was obtained after undergoing a thorough preprocessing process, including filtering, segmentation into 10-second epochs, feature normalization, and feature dimensionality reduction. The true positive rates (TPR) varied among all models between 70% and 90%, whereas the false positive rates (FPR) were less than 0.2%. We used rigorous comparisons between five machine learning methods on patient- and non-patient-specific seizure classification. Kernel SVM performed best among them, recording > 95% TPR using both scenarios of the classification, whereas it was better than a commercial non-patient seizure detector. In addition, we also discovered that the adjustment of class weights proved to be remarkably beneficial in detecting minority classes in highly imbalanced data sets. These results indicate that the use of scalp EEG-based seizure detection systems has potential as a clinical decision support system, especially in epilepsy monitoring units to provide continuous patient monitoring. The results indicate the possible use of advanced supervised learning algorithms, specifically KSVM, to supplement the accuracy of EEG-based seizure detection systems in clinical practice and upcoming wearable practitioner instruments.