Scholarlink

Abstract <p>Ripples-on-spikes (RonS) have been recognized as a more promising biomarker than spikes. They possess a high degree of specificity and sensitivity in diagnosing Benign Epilepsy with Centrotemporal Spikes (BECTS), a common childhood epilepsy syndrome. The existing automatic detection methods for RonS still lack sufficient robustness when dealing with noise and outliers, and the selection of the threshold is affected by individual differences. RonS events account for only a small portion of the EEG recordings, resulting in an imbalance in the samples. Aiming at the problems of the existing CNN-based RonS detection methods, such as poor detection performance of ripples in the high-frequency band and the inability to accurately identify temporal features to meet the detection requirements, this paper proposes a RonS detection algorithm based on a dual-path multi-kernel causal convolutional neural network (MK-CNN) with dynamic PSD channel weighting fusion, which is called <b>MK-PNet</b>. The framework fuses multi-band features through a multi-head attention network and encoding layers for RonS segment classification. The loss is calculated by combining downsampling with a loss function based on the Area Under the Curve (AUC). The proposed <b>MK-PNet</b> can achieve the precision of 94.84%, the accuracy of 97.67% and the F1-score of 94.33% on scalp EEG data collected from 31 subjects at Children’s Hospital of Zhejiang University (CHZU).</p> Graphical abstract <p></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

Multi-kernel convolutional neural network with attention mechanism for RonS detection

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Tianle Zhu,
  • Dinghan Hu,
  • Tiejia Jiang,
  • Shuangpeng Zhu,
  • Yunyun Zhao,
  • Jiuwen Cao

摘要

Abstract

Ripples-on-spikes (RonS) have been recognized as a more promising biomarker than spikes. They possess a high degree of specificity and sensitivity in diagnosing Benign Epilepsy with Centrotemporal Spikes (BECTS), a common childhood epilepsy syndrome. The existing automatic detection methods for RonS still lack sufficient robustness when dealing with noise and outliers, and the selection of the threshold is affected by individual differences. RonS events account for only a small portion of the EEG recordings, resulting in an imbalance in the samples. Aiming at the problems of the existing CNN-based RonS detection methods, such as poor detection performance of ripples in the high-frequency band and the inability to accurately identify temporal features to meet the detection requirements, this paper proposes a RonS detection algorithm based on a dual-path multi-kernel causal convolutional neural network (MK-CNN) with dynamic PSD channel weighting fusion, which is called MK-PNet. The framework fuses multi-band features through a multi-head attention network and encoding layers for RonS segment classification. The loss is calculated by combining downsampling with a loss function based on the Area Under the Curve (AUC). The proposed MK-PNet can achieve the precision of 94.84%, the accuracy of 97.67% and the F1-score of 94.33% on scalp EEG data collected from 31 subjects at Children’s Hospital of Zhejiang University (CHZU).

Graphical abstract