Scholarlink

This study presents a novel approach to heart sound classification utilizing time domain feature extraction and the CatBoost algorithm. This study analyzed 3153 phonocardiogram recordings from the PhysioNet/Computing in Cardiology Challenge 2016 dataset. Method used first extracts statistical, entropy-based, and temporal features without relying on heart sound segmentation. The CatBoost classifier was employed for binary classification of normal and abnormal heart sounds. The approach achieved 91.2% accuracy, 88.7% sensitivity, and 92.4% specificity. The area under the ROC curve was 0.937 (95% CI: [0.921, 0.953]). Feature importance analysis revealed Shannon Energy and Sample Entropy as the most influential factors. These results demonstrate that time domain features, combined with advanced ensemble learning, can match or exceed the performance of more complex models. The interpretability of our method, coupled with its competitive performance, suggests potential for clinical applications in various settings. This research contributes to automated cardiac auscultation by demonstrating the efficacy of a segmentation-free, time domain analysis approach. Future work should investigate the model's generalizability across diverse patient populations and recording conditions.

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

Temporal Feature Extraction and Gradient Boosting for Binary Cardiac Acoustic Classification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ahmed Ali Farhan Ogaili,
  • Noor T. Al-Sharify,
  • Zainab T. Al-Sharify,
  • Mushtaq Talib Al-Sharify,
  • Alaa Abdulhady Jaber,
  • Emad Zuhair Gheni

摘要

This study presents a novel approach to heart sound classification utilizing time domain feature extraction and the CatBoost algorithm. This study analyzed 3153 phonocardiogram recordings from the PhysioNet/Computing in Cardiology Challenge 2016 dataset. Method used first extracts statistical, entropy-based, and temporal features without relying on heart sound segmentation. The CatBoost classifier was employed for binary classification of normal and abnormal heart sounds. The approach achieved 91.2% accuracy, 88.7% sensitivity, and 92.4% specificity. The area under the ROC curve was 0.937 (95% CI: [0.921, 0.953]). Feature importance analysis revealed Shannon Energy and Sample Entropy as the most influential factors. These results demonstrate that time domain features, combined with advanced ensemble learning, can match or exceed the performance of more complex models. The interpretability of our method, coupled with its competitive performance, suggests potential for clinical applications in various settings. This research contributes to automated cardiac auscultation by demonstrating the efficacy of a segmentation-free, time domain analysis approach. Future work should investigate the model's generalizability across diverse patient populations and recording conditions.