心脏年龄的深度学习预测模型及其临床应用初探
Deep learning-based"cardiac age"prediction model and its implications for clinical practice
钟高艳 1刘森 1邓新桃 2李春普 2霍洪业 2杨翠微3
作者信息
- 1. 复旦大学信息科学与工程学院生物医学工程系,上海 200433
- 2. 兴化市人民医院心内科,兴化 225700
- 3. 复旦大学信息科学与工程学院生物医学工程系,上海 200433;上海市医学图像处理与计算机辅助手术重点实验室,上海 200093
- 折叠
摘要
目的 开发一个基于深度学习的心脏年龄预测模型,并探讨其在各类心电图数据集上的应用效果,特别是在评估心房颤动(房颤)患者病情严重程度方面的潜力.方法 首先利用PTB-XL公开数据库的12导联心电图和相应年龄数据,设计了一种结合多尺度Transformer网络和Inception架构的双流网络模型,提升心脏年龄预测的性能.接着,使用SPH数据集评估该模型的稳定性,根据心律失常发生率将该数据集分为全健康组、半健康组和异常组,并对模型对于不同组别的心脏年龄预测性能进行了比较分析.最后,探索了该模型在真实世界采集的不同轻、重症房颤患者数据集上的应用.模型性能通过(MAE)、均方根误差(RMSE)、决定系数(R2)、解释方差分数(EV)和相关系数(COR)来评估模型预测与实际标签的平均偏差及相关性.结果 在PTB-XL数据集中,模型的MAE和RMSE分别为7.48和9.43,R2为0.626,COR为0.748,EV为0.633.在SPH数据集中,健康组的MAE和RMSE显著低于异常组,且异常组的R2和EV显著低于健康组,验证了模型在健康样本上的高准确性,同时表明受试者心律失常的发生率越高,预测结果与实际年龄的差异越大.对于不同严重程度的房颤患者,轻症组和重症组的MAE分别为11.61和12.55,显示了心脏年龄预测准确性与患者病情严重程度有显著相关性(P＜0.001),表明预测的心脏年龄与真实年龄差值可作为衡量房颤严重程度的可靠指标.结论 基于心电图和神经网络预测的心脏年龄可以作为评估不同心律失常心脏健康程度的指标.本心脏年龄预测模型仅需使用患者的心电图数据,可为临床提供一种更为便捷的房颤病情严重程度判别工具.
Abstract
Objective To develop a deep learning-based model for predicting"cardiac age",and to explore its effectiveness across various electrocardiogram(ECG)datasets,particularly in assessing the severity of conditions in patients with atrial fibrillation(AF).Methods Utilizing the PTB-XL public database of 12-lead ECG and corresponding age data,the study designed an innovative dual-stream network structure that combines multi-scale Transformer and Inception blocks,aimed at enhancing the performance of cardiac age prediction.Additionally,the model's stability was evaluated using the SPH dataset,which was divided into all-healthy,semi-healthy and abnormal groups based on the incidence rate of arrhythmias,and the model's performance in predicting cardiac age across these groups was compared.Lastly,the application of this model was explored in datasets of AF patients with varying severity levels collected from the real world.Model performance was determined by mean absolute error(MAE),root mean squared error(RMSE),coefficient of determination(R2),explained variance(EV)and correlation coefficient(COR)to assess the average deviation and correlation between model predictions and actual labels.Results In the PTB-XL dataset,the model achieved MAE and RMSE of 7.48 and 9.43,with R2 of 0.626,COR of 0.748,and EV of 0.633,respectively.In the SPH dataset,the MAE and RMSE were significantly lower in the healthy group compared to the abnormal group,and both R2 and EV were significantly higher in the healthy group,validating the model's high accuracy in healthy samples and indicating that a higher incidence rate of arrhythmias results in greater discrepancies between predicted and actual age.For AF patients of different severities,the MAE for mild and severe groups was 11.61 and 12.55,respectively,demonstrating a significant correlation between the accuracy of cardiac age prediction and the severity of the patient's condition(P＜0.001),indicating that the difference between the predicted cardiac age and the actual age can serve as a reliable indicator for measuring the severity of AF.Conclusion The cardiac age predicted based on ECG and neural network can be used as an indicator to assess the degree of heart health of different arrhythmias.Compared to traditional evaluation methods,the cardiac age prediction model proposed in this research only requires the patient's ECG data,which can provide a more accessible and efficient tool for the clinical evaluation of AF severity.
关键词
心房颤动/心脏年龄/深度学习/心电图/病情评估Key words
Atrial fibrillation/Cardiac age/Deep learning/Electrocardiogram/Disease progression assessment引用本文复制引用
基金项目
国家自然科学基金(62371138)
江苏省卫生健康委医学科研项目(ZDB2020025)
江苏省卫生健康委医学科研项目(Z2020075)
出版年
2024
NETL
NSTL
万方数据