基于心音信号的常见先天性心脏病智能诊断算法研究

Deep learning based intelligent auscultation of heart sounds in neonates with congenital heart disease

目的对室间隔缺损、房间隔缺损、动脉导管未闭和卵圆孔未闭合并肺动脉高压4种常见先天性心脏病(简称先心病)心音信号进行分析,提出一种基于深度学习的智能听诊算法,实现心音信号的自动分类。方法基于数字信号处理技术,将一维时序信号分类问题转换为二维图像分类问题,利用深度神经网络实现心音的自动分类。采用该算法对浙江大学医学院附属儿童医院采集的941例心音数据进行训练、验证和测试,按照8∶1∶1的比例分为训练集、验证集和测试集。此外,本研究还收集了107例基于临床筛查环境的心音数据,用于验证智能听诊算法在实际临床应用中的效果。结果本文采用离散小波变换法对心音信号进行降噪处理,观察到降噪处理对模型性能的显著改善。与未经降噪处理的模型相比,经过降噪处理的模型在测试集上的准确率、灵敏度、特异度和F1分数分别提高了15.8%、32.6%、11.1%和27.3%。比较5种通用分类神经网络模型(Swin_transform、Vit、Mobilenet、Resenet和Vgg)的性能,F1分数分别为0.905、0.842、0.687、0.814和0.864。使用Swin_transform模型对107例外部数据集进行测试,得到0.833的准确率、0.872的灵敏度和0.801的特异度。结论先心病心音信号的自动分类模型性能受噪声与神经网络结构的影响较大。通过应用离散小波变换法对心音信号进行降噪处理,模型性能显著改善。比较多种通用分类神经网络模型发现Swin_transform模型展现出了最佳的分类性能。智能听诊算法在实际临床应用中有良好的有效性、准确率、灵敏度和特异度。基于深度学习的智能听诊算法在先心病心音信号自动分类方面具有潜在应用价值。

Objective To examine the heart sound signals of four common congenital heart diseases of ventricular septal defect,atrial septal defect,patent ductus arteriosus and patent foramen with associated pulmonary hypertension and propose a deep learning-based intelligent auscultation algorithm for automatic classification of heart sound signals.Methods The algorithm in this study was based upon digital signal processing technology of converting one-dimensional temporal signal classification into a two-dimensional image classification and further applying deep neural networks for automatic classification of heart sound signals.A total of 941 heart sound data samples collected from Children’s Hospital of Zhejiang University School of Medicine were employed for training,validation and testing with a ratio of 8:1:1 for training,validation and testing sets respectively.Additionally,107 heart sound data samples gathered from a clinical screening environment were collected to validate the effectiveness of intelligent auscultation algorithm in real-world clinical applications.Results In this study,discrete wavelet transformation was utilized to denoise the heart sound signals and there was a significant improvement in model performance.Compared to model without denoising,the denoised model achieved notable enhancements in accuracy,sensitivity,specificity and F1 score on testing set with improvements of 15.8%,32.6%,11.1%and 27.3%respectively.Furthermore,the authors compared the performance of several common classification neural network models,including Swin_transform,Vit,Mobilenet,Resenet and Vgg,with their respective F1 scores of 0.905,0.842,0.687,0.814 and 0.864.Finally,using Swin_transform model,tests on the external dataset of 107 cases yielded an accuracy of 0.833,a sensitivity of 0.872 and a specificity of 0.801.Conclusion This study highlights the significant impact of noise and neural network structure on the performance of automatic classification models for CHD heart sound signals.Through the application of discrete wavelet transform for denoising heart sound signals,a substantial improvement in model performance is observed.Among various common classification neural network models,Swin_transform model exhibits the best classification performance.Additionally,the validation of intelligent auscultation algorithm on an external dataset of 107 cases demonstrates its effectiveness in real-world clinical applications,yielding favorable accuracy,sensitivity and specificity results.In summary,this study demonstrates the promising potential of deep learning-based intelligent auscultation algorithms for automatic classification of CHD heart sound signals.