Analysis of the clinical value of deep learning algorithms in electrocardiogram diagnosis
CAO Chun-ge
SHEN Rui-qi
LIU Shu-jing
WANG Ya-nan
LI Jun-wei
YANG Cui-wei
WANG Hong-yu
Abstract：Objective To compare the diagnostic performance of the four deep learning models(ResNet18,GRU,1D-CNN and Transformer)in the dichotomous(normal/abnormal)and quadruple(normal,conduction abnormalities,ST-T alterations and myocardial infarction)tasks of 12-lead ECG and to evaluate their clinical application value.Methods The model was trained based on the PTB-XL dataset and validated across the database by the SXMU-2k dataset of the Second Hospital of Shanxi Medical University during the period from April 13,2018 to October 15,2023.After optimizing the hyperparameters,the model performance was evaluated in the PTB-XL 10th fold and SXMU-2k test sets,respectively.Results Among the PTB-XL quadruple classification tasks,GRU had the highest F1 score in the diagnosis of NORM,STTC and MI(0.88,0.66,0.73),and ResNet18 had the best performance in the diagnosis of CD(F1=0.75).GRU had the best overall performance,with quadruple and binary accuracy rates of 81％and 85％,respectively.In the cross-database verification,GRU has 84％and 97％accuracy of quadruple and binary classification in SXMU-2k.Conclusion The GRU model has the smallest accuracy deviation and the strongest generalization ability in the multi-dimensional analysis of 12-lead ECG,which has significant clinical practical value.The results of clinical validation across databases confirmed the generalization ability of the model.
Keywords：Deep learningElectrocardiogram diagnosisDiagnostic performanceClinical application
Publication Date：2025-06-20
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：8( 502-509 )