东华大学学报（英文版）2026，Vol.43Issue(3)：113-129,17.DOI:10.19884/j.1672-5220.202602015
基于Transformer深度学习模型的心电信号早期诊断先天性心脏病
Early Diagnosis of Congenital Heart Disease Using Transformer-Based Deep Learning on Electrocardiogram Signals
RAHMAN Md Saifur 1ZAMAN Junaid 2KHAIRUL Md Sadi Iftia 3ISLAM Md Rakibul 4张义红5
作者信息
- 1. 东华大学 信息与智能科学学院,上海 201620
- 2. 淮阴工学院,江苏 淮安 223003
- 3. 南京邮电大学,江苏 南京 210023
- 4. 中国矿业大学 机械工程学院,江苏 徐州 221116
- 5. 东华大学 信息与智能科学学院,上海 201620||东华大学 数字化纺织服装技术教育部工程研究中心,上海 201620
- 折叠
摘要
Abstract
Congenital heart disease(CHD)is one of the most common birth defects worldwide and a major cause of pediatric morbidity and mortality.Early detection is essential for improving survival and long-term outcomes,yet timely diagnosis remains challenging,especially in primary-care and resource-constrained settings.Although electrocardiography(ECG)is inexpensive and widely available,subtle pediatric CHD abnormalities are difficult to detect through manual interpretation,which carries subjectivity and risk of misdiagnosis.Moreover,many existing deep-learning models rely on single-domain convolutional neural network(CNN)or recurrent neural network(RNN)architectures that insufficiently capture long-range temporal dependencies and frequency-domain features.To address these limitations,we propose pediatric AI for cardiac ECG recognition(PACER),a hybrid CNN-Transformer-discrete wavelet transform(DWT)-TabNet framework for automated CHD detection from standard 12-lead pediatric ECG signals.PACER integrates convolutional layers for local morphological extraction,Transformer-based self-attention for long-range temporal modeling,DWT for frequency representation,and TabNet for interpretable multimodal feature fusion.A tailored preprocessing and augmentation pipeline,including SMOTE and Gaussian noise enhancement,improves robustness to class imbalance.Evaluated on 10 344 pediatric ECG recordings using stratified cross-validation,PACER achieved an accuracy of 90.93%,an F1 score of 0.91,and an area under the receiver operating characteristic curve(ROC-AUC)of 0.95,outperforming CNN,RNN,hybrid,Transformer,and CHDdECG baselines.Ablation experiments and model interpretability analysis validate the effectiveness of each module and indicate the potential clinical utility of the proposed method.关键词
先天性心脏病/深度学习/Transformer网络/小儿心电图学/自动诊断Key words
congenital heart disease/deep learning/Transformer networks/pediatric electrocardiography/automated diagnosis分类
医药卫生引用本文复制引用
RAHMAN Md Saifur,ZAMAN Junaid,KHAIRUL Md Sadi Iftia,ISLAM Md Rakibul,张义红..基于Transformer深度学习模型的心电信号早期诊断先天性心脏病[J].东华大学学报（英文版）,2026,43(3):113-129,17.
