Scholarlink

<p>Deep learning (DL) has significantly advanced automated epilepsy seizure detection and prediction, enabling improved feature learning and temporal modeling from EEG signals. However, progress remains constrained by heterogeneous datasets, inconsistent evaluation protocols, and limited cross-study comparability, often leading to optimistic performance estimates that do not translate to real-world settings. This review provides a protocol-aware synthesis of 129 studies across major public and clinical EEG datasets, systematically examining how dataset characteristics, temporal definitions, and preprocessing strategies influence model performance and generalization. We introduce a unified evaluation framework that standardizes task definitions, validation protocols (Intra-P, Inter-P/LOPO, cross-dataset), and clinically relevant time-aware metrics, including prediction horizon, seizure occurrence period (SOP), false positives per hour (FP/h), and time in warning (TIW). A taxonomy of DL architectures spanning CNNs, RNNs, hybrid CNN RNN models, transfer learning approaches, and graph-based networks is analyzed through protocol-aware comparisons. Cross-study synthesis reveals that spectrogram-based CNNs are highly effective for seizure detection, while hybrid CNN RNN and LSTM/GRU models better capture long-range temporal dependencies required for prediction. Transfer learning improves performance in data-scarce settings, whereas connectivity-aware and graph-based models enhance interpretability by modeling inter-channel relationships. Despite these advances, consistent limitations emerge, including weak cross-dataset generalization, sensitivity to noise and annotation variability, imbalance in datasets, inconsistent temporal framing, and limited interpretability. Furthermore, ethical and deployment challenges such as privacy, bias, regulatory compliance, and computational constraints remain critical barriers to clinical adoption. This review highlights key research directions, including multimodal and context-aware modeling, personalized and continual learning, privacy-preserving frameworks, and lightweight architectures for real-time wearable systems. By aligning datasets, architectures, evaluation protocols, and clinical requirements, this work provides a structured roadmap toward robust, interpretable, and clinically deployable DL-based seizure monitoring systems.</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

Deep Learning for EEG-based epilepsy seizure detection and prediction: a comprehensive review of datasets, architectures, and clinical challenges

  • 期刊论文

  • 发表日期：

    2026年5月18日

查看全文
  • Hamed Alqahtani,
  • Paras Ahuja,
  • Wesam Shishah,
  • Gulshan Kumar

摘要

Deep learning (DL) has significantly advanced automated epilepsy seizure detection and prediction, enabling improved feature learning and temporal modeling from EEG signals. However, progress remains constrained by heterogeneous datasets, inconsistent evaluation protocols, and limited cross-study comparability, often leading to optimistic performance estimates that do not translate to real-world settings. This review provides a protocol-aware synthesis of 129 studies across major public and clinical EEG datasets, systematically examining how dataset characteristics, temporal definitions, and preprocessing strategies influence model performance and generalization. We introduce a unified evaluation framework that standardizes task definitions, validation protocols (Intra-P, Inter-P/LOPO, cross-dataset), and clinically relevant time-aware metrics, including prediction horizon, seizure occurrence period (SOP), false positives per hour (FP/h), and time in warning (TIW). A taxonomy of DL architectures spanning CNNs, RNNs, hybrid CNN RNN models, transfer learning approaches, and graph-based networks is analyzed through protocol-aware comparisons. Cross-study synthesis reveals that spectrogram-based CNNs are highly effective for seizure detection, while hybrid CNN RNN and LSTM/GRU models better capture long-range temporal dependencies required for prediction. Transfer learning improves performance in data-scarce settings, whereas connectivity-aware and graph-based models enhance interpretability by modeling inter-channel relationships. Despite these advances, consistent limitations emerge, including weak cross-dataset generalization, sensitivity to noise and annotation variability, imbalance in datasets, inconsistent temporal framing, and limited interpretability. Furthermore, ethical and deployment challenges such as privacy, bias, regulatory compliance, and computational constraints remain critical barriers to clinical adoption. This review highlights key research directions, including multimodal and context-aware modeling, personalized and continual learning, privacy-preserving frameworks, and lightweight architectures for real-time wearable systems. By aligning datasets, architectures, evaluation protocols, and clinical requirements, this work provides a structured roadmap toward robust, interpretable, and clinically deployable DL-based seizure monitoring systems.