Scholarlink

In recent years, AI, specifically deep learning, has advanced automated analysis of the ECG signal. This study compares raw signal processing against pre-extracted features for multi-label diagnostic superclass classification using the public PTB-XL dataset and its PTB-XL+ extended dataset with feature sets (Uni-G, 12SL). Three deep learning architectures, namely, CNN-based signal-only, dense network features-only, and combined multi-modal models, have been evaluated using standard PTB-XL splits. Performance results on the test set show that the combined Uni-G/12SL feature-only models performed well (Macro F1 = 0.73, AUC = 0.92). However, it slightly underperformed a signal-only CNN (Macro F1 = 0.75, AUC = 0.93). The combined model performed best, but with only marginal gains over the signal-only model. These results suggest pre-extracted clinical features capture much of the predictive information, and integrating raw signals requires more sophisticated architectures to yield significant benefits.

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

Evaluation of Raw Signal and Feature-Based Deep Learning Models for Multi-Label ECG Diagnosis on PTB-XL

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Calvin Holloway,
  • Lakshmi Babu Saheer,
  • Mahdi Maktab Dar

摘要

In recent years, AI, specifically deep learning, has advanced automated analysis of the ECG signal. This study compares raw signal processing against pre-extracted features for multi-label diagnostic superclass classification using the public PTB-XL dataset and its PTB-XL+ extended dataset with feature sets (Uni-G, 12SL). Three deep learning architectures, namely, CNN-based signal-only, dense network features-only, and combined multi-modal models, have been evaluated using standard PTB-XL splits. Performance results on the test set show that the combined Uni-G/12SL feature-only models performed well (Macro F1 = 0.73, AUC = 0.92). However, it slightly underperformed a signal-only CNN (Macro F1 = 0.75, AUC = 0.93). The combined model performed best, but with only marginal gains over the signal-only model. These results suggest pre-extracted clinical features capture much of the predictive information, and integrating raw signals requires more sophisticated architectures to yield significant benefits.