Scholarlink

<p>A convolutional neural network (CNN)–enhanced electrocardiogram (ECG) has been reported for detecting mitral regurgitation (MR). This tool may be particularly useful for identifying candidates for echocardiography in patients with chronic atrial fibrillation (AF) to detect atrial functional MR early. The data from a single-center, prospective cohort study (Shinken Database 2010–2017, <i>n</i> = 19,170) were combined with an ECG database. Initially, a CNN model was developed to detect MR (Grade ≥ 3) across the entire cohort using fivefold cross-validation. The model was refined using sublabels, including primary MR, MR with chronic AF and left atrial dilatation, and MR with left ventricular remodeling, to create an integrated neural network (INN) model. We then analyzed the relationship between MR diagnosed by the INN and the MR prevalence in chronic AF patients. In the CNN model, the AUCs of the ROC curve and PR curve in 0.836 (SD: 0.022) and 0.196 (SD: 0.036), which numerically increased to 0.848 (SD: 0.014) and 0.198 (SD: 0.031) in the INN model. The Grad-CAM analysis revealed that the CNN algorithm appears to highlight nonspecific ECG features, such as P-waves in the leads V1 to V2 (or f-wave in the lead V1) and R-wave amplitude or ST-T changes in precordial leads, which may explain the high false-positive rate in the model. When applying the model to CAF patients, although the sensitivity was around 0.9 at the threshold determined by the ROC curve, PPR and F1 score was relatively low. These metrics slightly improved when adjusting the threshold to that corresponding to a sensitivity of 0.8 and further improved by restricting the target population to those with BNP ≥ 100&#xa0;pg/mL. The INN model improved MR detection performance compared to the initial CNN model, but the overall PPR remained suboptimal. High false-positive rates remained an issue, even in high-prevalence populations such as CAF patients or those with elevated BNP values.</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

Utility of convolutional neural network-enhanced electrocardiogram to diagnose and predict mitral regurgitation in patients with chronic atrial fibrillation

  • 期刊论文

  • 发表日期：

    2025年5月15日

查看全文
  • Mayu Sakuma,
  • Shinya Suzuki,
  • Naomi Hirota,
  • Jun Motogi,
  • Takuya Umemoto,
  • Hiroshi Nakai,
  • Wataru Matsuzawa,
  • Tsuneo Takayanagi,
  • Akira Hyodo,
  • Keiichi Satoh,
  • Takuto Arita,
  • Naoharu Yagi,
  • Mikio Kishi,
  • Hiroaki Semba,
  • Hiroto Kano,
  • Shunsuke Matsuno,
  • Yuko Kato,
  • Takayuki Otsuka,
  • Junji Yajima,
  • Yasuchika Takeishi,
  • Tokuhisa Uejima,
  • Yuji Oikawa,
  • Takeshi Yamashita

摘要

A convolutional neural network (CNN)–enhanced electrocardiogram (ECG) has been reported for detecting mitral regurgitation (MR). This tool may be particularly useful for identifying candidates for echocardiography in patients with chronic atrial fibrillation (AF) to detect atrial functional MR early. The data from a single-center, prospective cohort study (Shinken Database 2010–2017, n = 19,170) were combined with an ECG database. Initially, a CNN model was developed to detect MR (Grade ≥ 3) across the entire cohort using fivefold cross-validation. The model was refined using sublabels, including primary MR, MR with chronic AF and left atrial dilatation, and MR with left ventricular remodeling, to create an integrated neural network (INN) model. We then analyzed the relationship between MR diagnosed by the INN and the MR prevalence in chronic AF patients. In the CNN model, the AUCs of the ROC curve and PR curve in 0.836 (SD: 0.022) and 0.196 (SD: 0.036), which numerically increased to 0.848 (SD: 0.014) and 0.198 (SD: 0.031) in the INN model. The Grad-CAM analysis revealed that the CNN algorithm appears to highlight nonspecific ECG features, such as P-waves in the leads V1 to V2 (or f-wave in the lead V1) and R-wave amplitude or ST-T changes in precordial leads, which may explain the high false-positive rate in the model. When applying the model to CAF patients, although the sensitivity was around 0.9 at the threshold determined by the ROC curve, PPR and F1 score was relatively low. These metrics slightly improved when adjusting the threshold to that corresponding to a sensitivity of 0.8 and further improved by restricting the target population to those with BNP ≥ 100 pg/mL. The INN model improved MR detection performance compared to the initial CNN model, but the overall PPR remained suboptimal. High false-positive rates remained an issue, even in high-prevalence populations such as CAF patients or those with elevated BNP values.