Scholarlink

<p>Electrocardiograms (ECGs) exhibit variability due to disease causes, locations, individual characteristics, and noise, necessitating deep neural networks (DNNs) for precise analysis. However, DNN training is hindered by limited ECG data and output bias, leading to imbalance issues. This study addresses imbalance by replicating rare outputs, improving dataset balance and evaluates the impact on ECG wave detection performance. Six filter-type neural networks (f-NNs) were designed to identify P, QRS, and T wave boundaries, with 15 outputs indicating target locations and one indicating absence. Training datasets were constructed from ECGRDVQ (Physionet), and rare outputs were replicated up to 40 times. Performance evaluation using ECG data from Kangwon National University Hospital (KNUH) showed that F1 scores for f-NNs trained on the original dataset ranged from 0.50 to 0.60. After augmentation, F1 scores improved significantly, ranging from 0.87 to 1.00, demonstrating enhanced neural network performance.</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

Improvement of filter-type neural network performance for ECG analysis by adjusting the output ratio of training data

  • 期刊论文

  • 发表日期：

    2025年7月1日

查看全文
  • Ui-in Lee,
  • Seong-wook Choi

摘要

Electrocardiograms (ECGs) exhibit variability due to disease causes, locations, individual characteristics, and noise, necessitating deep neural networks (DNNs) for precise analysis. However, DNN training is hindered by limited ECG data and output bias, leading to imbalance issues. This study addresses imbalance by replicating rare outputs, improving dataset balance and evaluates the impact on ECG wave detection performance. Six filter-type neural networks (f-NNs) were designed to identify P, QRS, and T wave boundaries, with 15 outputs indicating target locations and one indicating absence. Training datasets were constructed from ECGRDVQ (Physionet), and rare outputs were replicated up to 40 times. Performance evaluation using ECG data from Kangwon National University Hospital (KNUH) showed that F1 scores for f-NNs trained on the original dataset ranged from 0.50 to 0.60. After augmentation, F1 scores improved significantly, ranging from 0.87 to 1.00, demonstrating enhanced neural network performance.