迁移学习方法在精准识别室性异位QRS波群中的应用  

Application of transfer learning method in accurately identifying ventricular ectopic QRS complexes

作　　者：傅国华 储慧民[1] 孙汉泽 王彬浩 卓伟东 丰明俊[1] 高昉 戚莹波 FU Guohua;CHU Huimin;SUN Hanze;WANG Binhao;ZHUO Weidong;FENG Mingjun;GAO Fang;QI Yingbo(Arrhythmia Center,the First Affiliated Hospital of Ningbo University,Ningbo 315000,China;不详)

机构地区：[1]宁波大学附属第一医院心律失常诊疗中心,315000 [2]慈溪市人民医院医疗健康集团心内科

出　　处：《心电与循环》2024年第5期469-474,共6页Journal of Electrocardiology and Circulation

基　　金：浙江省基础公益研究计划项目(LGJ20H020001)。

摘　　要：目的探索迁移学习方法在精准识别室性异位QRS波群中的应用。方法回顾性提取2007年3月至2019年9月在宁波大学附属第一医院心律失常诊疗中心行射频导管消融术的545例特发性室性心律失常(VAs)患者,构成研究数据库(Ningbo DB)。采用小波散射网络特征提取算法从Ningbo DB中获取模型输入特征,使用美国麻省理工学院心律失常研究数据库(MIT-BIH DB)对深度学习模型进行训练,利用Ningbo DB验证该深度学习模型作为迁移学习方法。采用ROC曲线相关指标AUC、灵敏度、特异度、F1评分和准确度评估深度学习模型预测的效能。结果共提取心电图片段19695段,其中室性心律下QRS波群片段11988段(60.87%)、室性心律片段7471段(37.93%)、室上性期前收缩伴差异性传导片段236段(1.20%)。利用MIT-BIH DB建立深度学习模型,识别室性心律QRS波群的AUC、灵敏度、特异度、F1评分和准确度分别为0.997、0.996、1.000、0.997、0.996。采用Ningbo DB验证该模型识别室性心律QRS波群的AUC、灵敏度、特异度、F1评分和准确度分别为0.996、0.993、0.998、0.995、0.997。结论小波散射网络特征提取算法与残差网络深度学习模型结合的迁移学习方法能识别室性异位QRS波群。Objective To explore the application of transfer learning methods in accurately identifying ventricular ectopic QRS complexes.Methods Electrocardiographic data from 545 patients who underwent radiofrequency catheter ablation for idiopathic ventricular arrhythmias(VAs)between March 2007 and September 2019 at the Arrhythmia Center of the First Affiliated Hospital of Ningbo University were retrospectively extracted to construct a research database(Ningbo DB).The wavelet scattering network feature extraction algorithm was applied to obtain model input features from the Ningbo DB,and the deep learning model was trained using the Massachusetts Institute of Technology Beth Israel Hospital Arrhythmia Database(MIT-BIH DB).The Ningbo DB was then used to validate the deep learning model as a transfer learning method.ROC-AUC,sensitivity,specificity,F1 score,and accuracy were used to evaluate the diagnostic efficacy of the deep learning model.Results A total of 19695 electrocardiogram segments were extracted.Among them,there were 11988 segments(60.87%)of QRS complexes in sinus rhythm,7471 segments(37.93%)of ventricular ectopic complexes,and 236 segments(1.20%)of supraventricular premature contractions with aberrant conduction.The AUC,sensitivity,specificity,F1 score,and accuracy of the deep learning model established using MIT-BIH DB for identifying ventricular ectopic QRS complexes were 0.997,0.996,1.000,0.997,and 0.996,respectively,which were validated using Ningbo DB with values of 0.996,0.993,0.998,0.995,and 0.997,respectively.Conclusion The transfer learning method combining the wavelet scattering network feature extraction algorithm and the residual network deep learning model can accurately identify ventricular ectopic QRS complexes.

关 键 词：人工智能 室性异位QRS波群 迁移学习 深度学习模型 

分 类 号：R541.7[医药卫生—心血管疾病]