Scholarlink

Preventing serious cardiovascular events requires early diagnosis of cardiac arrhythmias utilizing electrocardiogram (ECG) signals. Despite their excellent classification accuracy, deep learning models are sometimes too computationally complex for wearable technology and embedded systems with limited resources. In order to categorize ECG beats from the MIT-BIH Arrhythmia Database, this study suggests a lightweight hybrid model that combines 1D Convolutional Neural Network (CNN), Gated Recurrent Units (GRU), and an attention mechanism. A carefully designed and effective preprocessing pipeline ensured high-quality inputs, enhancing the model’s ability to generalize to unseen patients. The proposed 276 KB model achieves a macro-F1 score of 0.9363 and precision above 92% across all major classes, enabling real-time arrhythmia detection on microcontrollers, demonstrating that accurate ECG beat classification can be achieved within the strict memory and processing limits of wearable and portable healthcare devices.

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

Low-Memory CNN–GRU–Attention Network for Accurate ECG Arrhythmia Detection on Edge Devices

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Saad Eddine Albourine,
  • Mohamed Ben Salah,
  • Azzedine Dliou,
  • Mohamed Zarboubi

摘要

Preventing serious cardiovascular events requires early diagnosis of cardiac arrhythmias utilizing electrocardiogram (ECG) signals. Despite their excellent classification accuracy, deep learning models are sometimes too computationally complex for wearable technology and embedded systems with limited resources. In order to categorize ECG beats from the MIT-BIH Arrhythmia Database, this study suggests a lightweight hybrid model that combines 1D Convolutional Neural Network (CNN), Gated Recurrent Units (GRU), and an attention mechanism. A carefully designed and effective preprocessing pipeline ensured high-quality inputs, enhancing the model’s ability to generalize to unseen patients. The proposed 276 KB model achieves a macro-F1 score of 0.9363 and precision above 92% across all major classes, enabling real-time arrhythmia detection on microcontrollers, demonstrating that accurate ECG beat classification can be achieved within the strict memory and processing limits of wearable and portable healthcare devices.