摘要
心律失常是心血管疾病中常见的临床表现形式,实现心律失常的自动分类在医学领域具有重要意义。在实际临床中,医生除了提供诊断结果,还要有详细的解释来支持自己的诊断,但是现有的大多数机器学习模型都忽略了结果的可解释性。同时,之前大部分研究致力于宏观分类,实际临床意义不大。为了解决这些问题,提出了一种可解释的基于注意力的混合深度学习模型(IAHM)。IAHM通过分别提取心拍级别和心律级别的注意力特征,将医学知识和心电数据相结合,使学习的模型具有高度的可解释性。实验在公开数据库MIT-BIH上进行,对五种心律失常分类以弥补宏观分类的短板。IAHM在准确率、特异性、敏感性和阳性预测值分别达到94.65%、98.69%、92.69%和92.60%,有助于临床医生对心律失常进行准确分类。
Automatic arrhythmia classification is of great significance in medical area with arrhythmias being common clinical manifestations of cardiovascular diseases.In the actual clinical environment,in addition to providing diagnostic results,cardiologists also need detailed explanations to support their diagnosis.However,most of the existing models ignore the interpretability.Meanwhile,much effort is devoted to the classification of macro-classes,which lacks of practical clinical significance.To address such issues,a novel interpretable attention-based hybrid deep learning model(IAHM)is proposed.By extracting the attention features of beat-level and rhythm-level respectively,IAHM combines the medical knowledge and ECG data,making the learned model highly interpretable.The experiment is conducted on the public database MIT-BIH,and 5 categories of arrhythmia are classified to make up for the shortcomings of the macro classification.The accuracy,specificity,sensitivity and positive predictive value of IAHM have reached 94.65%,98.69%,92.69%and 92.60%,respectively,which can thus help clinicians to classify arrhythmias accurately.
作者
罗望成
杨湘
陈艳红
LUO Wang-cheng;YANG Xiang;CHEN Yan-hong(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Wuhan Asia Heart Hospital,Wuhan 430022,China)
出处
《计算机技术与发展》
2022年第9期114-120,共7页
Computer Technology and Development
基金
国家自然科学基金(U1836118)。
关键词
心律失常分类
注意力机制
卷积神经网络
可解释性
长短期记忆网络
arrhythmia classification
attention mechanism
convolutional neural network
interpretability
long short term memory
