摘要目的:应用人工智能方法建立心脏年龄预测模型，评估其在真实世界心电数据上的心脏年龄预测效果。方法:使用多个数据集对心脏年龄预测模型进行训练和验证。首先，获取20 s到数分钟心电图正常的12导联匿名心电图记录。该数据用于对模型的训练、验证和测试（测试集）。其次，在一个大型的开源心电数据集（PTB-XL数据集）上对训练完成的模型进行了外部验证。该数据集包含来源于18 869例患者的21 799条10 s的心电图记录。其中，测试集和PTB-XL数据集为12导联数据集，提取测试集和PTB-XL数据集中I导联构建单导联测试集和单导联PTB-XL数据集。最后，在一个包含184 749例真实世界采集到的单导联心电数据集（Heartvoice数据集）上对心脏年龄预测模型进行了真实世界验证。Heartvoice数据集来源于2020年6月至2022年12月手持式单导联心电记录仪采集到的184 749例受试者的心电记录。以受试者的生理年龄为金标准对心脏年龄预测结果和心龄变异性（HAV）进行验证，利用平均绝对误差（MAE）衡量心脏年龄预测模型的准确性。结果:测试集的MAE分别为4.79岁（12导联）和6.45岁（单导联），PTB-XL数据集的MAE分别为8.06岁（12导联）和10.08岁（单导联），Heartvoice数据集的MAE为7.28岁。相关性分析结果显示：测试集的相关系数分别为0.887（ P<0.000 1，12导联）和0.87（ P<0.000 1，单导联），PTB-XL数据集的相关系数分别为0.790（ P<0.000 1，12导联）和0.711（ P<0.000 1，单导联），Heartvoice数据集的相关系数为0.849（ P<0.000 1）。一致性分析结果显示：PTB-XL数据集的均值误差分别为（-6.00±21.46）岁（12导联）和（-4.88±17.46）岁（单导联），Heartivoice数据集的均值误差为（-5.99±14.98）岁。 结论:基于人工智能方法的心脏年龄预测模型在心脏年龄预测方面具有较高性能，心脏年龄预测结果的差异可对用户起到警示作用，有望实现心脏疾病早期预警。

abstractsObjective:To evaluate the performance of heart age prediction using an artificial intelligence proach on real-world electrocardiogram (ECG) data.Methods:The heart age prediction model was trained and validated using multiple datasets. First, 12-lead anonymized ECG recordings of 20 s to several minutes of normal ECG were obtained. This data was used for training, validation, and testing of the model (Test set) . Second, external validation of the trained model was performed on a large open-source ECG dataset (PTB-XL dataset) . This dataset contains 21 799 10 s ECG recordings originating from 18 869 patients. Among them, the Test set and PTB-XL dataset were 12-lead datasets, and the lead I of the test set and PTB-XL dataset was extracted to construct the single-lead test set and single-lead PTB-XL dataset. Finally, real-world validation of the heart age prediction model was performed on a real-world acquired single-lead ECG dataset containing 184 749 cases (Heartvoice dataset) . The Heartvoice dataset was derived from ECG recordings of 184 749 subjects acquired from June 2020 to December 2022 by handheld single-lead ECG recorders. Validation of heart age prediction results and heart age variability (HAV) were performed using subjects' physiologic age as the gold standard, and the accuracy of the heart age prediction model was measured using the mean absolute error (MAE) .Results:The MAEs were 4.79 (12-lead) and 6.45 (single-lead) years for the Test set, 8.06 (12-lead) and 10.08 (single-lead) years for the PTB-XL dataset, and 7.28 years for the Heartvoice dataset. The correlation coefficient were 0.877 ( P<0.000 1, 12-lead) and 0.783 ( P<0.000 1, single-lead) for the test set, 0.790 ( P<0.000 1, 12-lead) and 0.711 ( P<0.000 1, single-lead) for the PTB-XL dataset, and 0.849 ( P<0.000 1) for the Heartvoice dataset. The bland-altman mean error were (-6.00±21.46) (12-lead) and (-4.88±17.46) (single-lead) years for the PTB-XL dataset and (-5.99±14.98) years for the Heartvoice dataset. Conclusion:The heart age prediction model based on artificial intelligence methods has high performance in heart age prediction, and the difference in heart age prediction results can serve as a warning to users, which is expected to realize early warning of heart disease.