Many studies that rely on manual ECG interpretation as a reference use multiple ECG expert interpreters and a method to resolve differences between interpreters, reflecting the fact that experts sometimes use different criteria. The aim of this study was to show the effect of manual ECG interpretation style on training automated ECG interpretation.

Methods: The effect of ECG interpretation style or differing ECG criteria on algorithm training was shown in this study by careful analysis of the changes in algorithm performance when the algorithm was trained on one database and tested on a different database. Morphology related ECG interpretation was summarized in eleven abnormalities such as left bundle branch block (LBBB) and old anterior myocardial infarction (MI). Each of the two databases used in the study had a reference interpretation mapped to those eleven abnormalities. F1 algorithm performance scores across abnormalities were compared for four cases. First, the algorithm was trained and tested on randomly split database A and then trained on the training set of database A and tested on randomly chosen test set of database B. The previous two test cases were repeated for opposite databases, train and test on database B and then train on database B and test on the test set of database A.

Results: F1 scores across abnormalities were generally higher when training and testing on the same database. F1 scores were high for bundle branch blocks (BBB) no matter the training and testing database combination. Old anterior MI F1 score dropped for one cross-database comparison and not the other suggesting a difference in manual interpretation.

Conclusion: For some abnormalities, human experts appear to have used different criteria for ECG interpretation, as evident by the difference between cross-database and within-database performance. Bundle branch blocks appear to be interpreted in a consistent manner.

许多依赖人工心电图解释作为参考的研究使用多个心电图专家解释器和解决解释器之间差异的方法，这反映了专家有时使用不同标准的事实。本研究的目的是展示手动心电图解释方式对训练自动心电图解释的影响。

方法：本研究通过仔细分析算法在一个数据库上训练并在不同数据库上测试时算法性能的变化，显示了心电图解释风格或不同心电图标准对算法训练的影响。形态学相关的心电图解释总结为 11 种异常，例如左束支传导阻滞 (LBBB) 和陈旧性前壁心肌梗塞 (MI)。研究中使用的两个数据库中的每一个都有一个参考解释映射到这 11 个异常。比较了四个案例的 F1 算法在异常情况下的性能得分。首先，算法在随机拆分的数据库 A 上进行训练和测试，然后在数据库 A 的训练集上进行训练，并在数据库 B 的随机选择的测试集上进行测试。

结果：在同一数据库上进行训练和测试时，异常的 F1 分数通常较高。无论训练和测试数据库组合如何，束支块 (BBB) 的 F1 分数都很高。一个跨数据库比较的旧前 MI F1 评分下降，而另一个则表明人工解释存在差异。

结论：对于某些异常，人类专家似乎使用了不同的心电图解释标准，这从跨数据库和数据库内性能之间的差异就可以看出。束分支块似乎以一致的方式解释。