王晓娟 1余睿 1周著黄 1高小峰 2朱红玲 3宾光宇1
作者信息
- 1. 北京工业大学化学与生命科学学院,北京 100124
- 2. 北京麦迪克斯科技有限公司,北京 100095
- 3. 华中科技大学同济医学院附属同济医院心血管内科,武汉 430030
- 折叠
摘要
Abstract
Objective To propose a bias-correction method for cardiac age prediction based on residual convolutional networks and verify its effectiveness,and to investigate the application value of the corrected age gap(AG)for disease classification and mortality risk assessment.Methods First,from four public databases of the CODE-15%ECG dataset,the MIMIC-Ⅳ database,the PTB_XL ECG dataset and the UK Biobank,the 12-lead ECG data of 185 005 healthy samples were selected and divided into a healthy training set(n=148 022),a healthy validation set(n=18 479)and a healthy test set(n=18 504).Additionally,23 917 disease samples were enrolled into a disease test set.Second,a cardiac age prediction model was constructed using a residual convolutional network and then trained with the healthy samples;the mean prediction residuals across all age groups calculated with the validation set was used as the age bias coefficient,and the prediction results for the test set were corrected by subtracting the corresponding bias coefficient from the predicted cardiac age of each sample in the test set.Finally,changes in the AG distribution were evaluated on the healthy test set before and after correction,and the method proposed was compared with the linear correction method commonly used in brain age studies using the mean absolute error(MAE)and standard deviation(SD)metrics.An independent sample t-test was used to compare AG differences between the healthy and diseased test sets,and survival analysis and Cox regression were employed to explore the association between adjusted AG and the risk of all-cause mortality.Results The results of the bias correction indicated that,prior to correction,the model exhibited significant age bias in the healthy test set(among the participants with a physiological age of approximately 20 years,the mean AG was close to 10 years;among elderly participants aged approximately 80 years,the mean AG was approximately-10 years);after correction,the AG distribution for all age groups approached 0.A comparative analysis of methods showed that,compared with the linear correction method with the results in the older age range(71-85 years)(MAE=5.12,SD=4.44),the residual correction method proposed yielded a lower MAE(5.00)and smaller SD(4.30)in this range,indicating more thorough bias correction and a more concentrated AG distribution.Disease discrimination analysis showed that the corrected AG in the disease test set was significantly higher than that in the healthy test set;notably,the corrected AG shifted from negative to positive in elderly patients,which was consistent with clinical and pathological characteristics.Survival analysis results proved that the survival rate of the high-risk population,defined based on the corrected AG,was significantly lower than that of the low-risk population;after grouping AG into quintiles and incorporating it as an ordinal variable into the Cox regression model,AG was found to be significantly positively correlated with all-cause mortality risk,with the hazard ratio being 1.40 and the risk of death increasing by approximately 40%on average with each higher quintile.Conclusion The proposed residual correction method significantly reduces systematic age bias in cardiac age prediction and improves predictive accuracy.The corrected AG serves as a reliable biomarker for reflecting pathological cardiac aging,distinguishing between healthy and diseased states and predicting mortality risk,thereby offering a new approach for early screening and risk stratification of cardiovascular diseases.[Chinese Medical Equipment Journal,2026,47(4):1-12]关键词
残差卷积网络/心脏年龄/偏差校正/深度学习/心血管疾病/生物标志物Key words
residual convolutional network/cardiac age/bias correction/deep learning/cardiovascular disease/biomarker分类
医药卫生
