Scholarlink

<p>Electroencephalography (EEG) is pivotal in diagnosing epilepsy by capturing brain electrical activity. Neurologists visually analyse EEG recordings to detect seizure-related segments. However, manual examination of the extensive data from ambulatory EEGs used for continuous monitoring is time-consuming and can vary with expertise. To address these challenges, we introduce three hybrid architectures for binary seizure classification. The first two architectures combine convolutional neural networks (CNN) with bidirectional gated recurrent unit (Bi-GRU) and bidirectional long short-term memory (Bi-LSTM). These hybrid models were selected based on their ability to capture both spatial features and temporal dependencies, which are crucial for accurate seizure detection. The third model, the ‘convolutional self-attention model’, blends CNN with the Transformer architecture, pioneering transformers in seizure detection. Deviating from single architecture models, our hybrid approach leverages the strengths of two architectures, augmenting the performance of detection models. This methodology strategically employs CNNs to discern spatial patterns, RNNs to identify sequential patterns, and transformers to unveil global relationships within EEG data. Evaluation on the CHB-MIT dataset yielded promising results. CNN-Bi-GRU achieved 98.21% accuracy, 92.63% sensitivity, and 98.31% specificity; CNN-Bi-LSTM achieved 98.88% accuracy, 92.98% sensitivity, and 99% specificity; and the convolutional self-attention model exhibited superior performance with 99.04% accuracy, 96.06% sensitivity, and 99.11% specificity. Our models outperformed existing seizure detection methods on the CHB-MIT dataset, advancing automated seizure detection for better diagnostic accuracy and patient care. The novelty of the study lies in integrating diverse neural network architectures, highlighting the need for personalized models and future research directions.</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

An innovative approach harmonizing convolution and self-attention to enhance EEG-based seizure detection

  • 期刊论文

  • 发表日期：

    2025年6月3日

查看全文
  • Resmi Cherian,
  • E. Grace Mary Kanaga

摘要

Electroencephalography (EEG) is pivotal in diagnosing epilepsy by capturing brain electrical activity. Neurologists visually analyse EEG recordings to detect seizure-related segments. However, manual examination of the extensive data from ambulatory EEGs used for continuous monitoring is time-consuming and can vary with expertise. To address these challenges, we introduce three hybrid architectures for binary seizure classification. The first two architectures combine convolutional neural networks (CNN) with bidirectional gated recurrent unit (Bi-GRU) and bidirectional long short-term memory (Bi-LSTM). These hybrid models were selected based on their ability to capture both spatial features and temporal dependencies, which are crucial for accurate seizure detection. The third model, the ‘convolutional self-attention model’, blends CNN with the Transformer architecture, pioneering transformers in seizure detection. Deviating from single architecture models, our hybrid approach leverages the strengths of two architectures, augmenting the performance of detection models. This methodology strategically employs CNNs to discern spatial patterns, RNNs to identify sequential patterns, and transformers to unveil global relationships within EEG data. Evaluation on the CHB-MIT dataset yielded promising results. CNN-Bi-GRU achieved 98.21% accuracy, 92.63% sensitivity, and 98.31% specificity; CNN-Bi-LSTM achieved 98.88% accuracy, 92.98% sensitivity, and 99% specificity; and the convolutional self-attention model exhibited superior performance with 99.04% accuracy, 96.06% sensitivity, and 99.11% specificity. Our models outperformed existing seizure detection methods on the CHB-MIT dataset, advancing automated seizure detection for better diagnostic accuracy and patient care. The novelty of the study lies in integrating diverse neural network architectures, highlighting the need for personalized models and future research directions.