计算机技术与发展2025，Vol.35Issue(9)：182-191,10.DOI:10.20165/j.cnki.ISSN1673-629X.2025.0094
基于对比学习和掩码重构的无监督心梗检测算法
Contrastive Learning and Masked Reconstruction-based Unsupervised Algorithm for Myocardial Infarction Anomaly Detection
摘要
Abstract
Myocardial infarction(MI)is one of the leading global health burdens,and deep learning models for MI detection based on e-lectrocardiograms(ECG)play a crucial role in the early diagnosis of the disease.Traditional MI detection methods rely on large-scale labeled ECG data,which are scarce in real-world settings.Masked reconstruction techniques enable model pre-training by reconstructing masked regions from observed data.However,random masking of time points disrupts temporal information,limiting the effectiveness of representation learning.To address this,we propose an unsupervised MI detection algorithm(CLMR)based on contrastive learning and masked reconstruction,framing MI detection as an anomaly detection task trained exclusively on normal ECG data.Through lead segmentation and masking strategies,the model captures MI-related anomalies in the Q-wave,ST-segment,and T-wave.Furthermore,cross-lead contrastive learning between original and reconstructed limb and precordial leads enhances feature extraction within a shared latent space,while complementary information across multiple masked sequences improves reconstruction quality.Experimental results show that the CLMR method outperforms mainstream algorithms by 8.72 percentage points and 6.05 percentage points in AUC,and 10.69 percentage points and 3.55 percentage points in AP on the public PTBXL and real EMI12 datasets,respectively.These findings validate the model's effectiveness in MI anomaly detection tasks.关键词
心电图/心肌梗死/异常检测/对比学习/无监督学习/掩码重构Key words
electrocardiogram/myocardial infarction/anomaly detection/contrastive learning/unsupervised learning/mask reconstruction分类
信息技术与安全科学引用本文复制引用
陆王威,樊好义,陈姞伽,常慧慧..基于对比学习和掩码重构的无监督心梗检测算法[J].计算机技术与发展,2025,35(9):182-191,10.基金项目
河南省重大科技专项项目(241100310200) （241100310200）
科技创新2030——"新一代人工智能"重大项目(2021ZD0111000) （2021ZD0111000）
