Scholarlink

<p>Real-time electrocardiogram (ECG) anomaly detection is essential for early diagnosis and continuous heart monitoring. However, deploying deep learning models in wearable or mobile environments is still limited by computational and memory constraints. In this paper, we introduce a lightweight QRS-attention network (QRSAN), optimized by quantization for efficient and low-latency arrhythmia detection. The proposed QRS-attention mechanism uses temporal convolution and causal attention to selectively highlight discriminative features within QRS complexes, enabling robust identification of abnormal heartbeats in streaming ECG signals. Experiments on the MIT-BIH Arrhythmia Database demonstrate that QRSAN achieves superior performance across multiple evaluation metrics, including accuracy, precision, recall, specificity, F1-score and latency. Compared with state-of-the-art lightweight baselines, QRSAN improves F1-score by up to twofold over LSTM and by approximately 2% over CNN–LSTM and EfficientNet1D, with comparable latency. These results indicate that QRSAN provides an effective and computationally efficient framework for continuous ECG anomaly monitoring in wearable and edge-computing 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

Lightweight QRS-attention network for efficient real-time detection of ECG anomalies

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月11日

查看全文
  • Namkyung Yoon,
  • Sanghong Kim,
  • Seunghyun Yoo,
  • Hwangnam Kim

摘要

Real-time electrocardiogram (ECG) anomaly detection is essential for early diagnosis and continuous heart monitoring. However, deploying deep learning models in wearable or mobile environments is still limited by computational and memory constraints. In this paper, we introduce a lightweight QRS-attention network (QRSAN), optimized by quantization for efficient and low-latency arrhythmia detection. The proposed QRS-attention mechanism uses temporal convolution and causal attention to selectively highlight discriminative features within QRS complexes, enabling robust identification of abnormal heartbeats in streaming ECG signals. Experiments on the MIT-BIH Arrhythmia Database demonstrate that QRSAN achieves superior performance across multiple evaluation metrics, including accuracy, precision, recall, specificity, F1-score and latency. Compared with state-of-the-art lightweight baselines, QRSAN improves F1-score by up to twofold over LSTM and by approximately 2% over CNN–LSTM and EfficientNet1D, with comparable latency. These results indicate that QRSAN provides an effective and computationally efficient framework for continuous ECG anomaly monitoring in wearable and edge-computing systems.