Abstract
BACKGROUND: The electrocardiogram (ECG) is an almost universally accessible diagnostic tool for heart disease. An ECG is measured by using an electrocardiograph, and today's electrocardiographs use built-in software to interpret the ECGs automatically after they are recorded. However, these algorithms exhibit limited performance, and therefore, clinicians usually have to manually interpret the ECG, regardless of whether an algorithm has interpreted it or not. Manual interpretation of the ECG can be time-consuming and requires specific skills. Therefore, better algorithms are clearly needed to make correct ECG interpretations more accessible and time-efficient. Algorithms based on artificial intelligence (AI) have demonstrated promising performance in various fields, including ECG interpretation, over the past few years and may represent an alternative to manual ECG interpretation by doctors. RESULTS: We trained and validated a convolutional neural network with an Inception architecture on a dataset with 88253 12-lead ECGs, and classified 30 of the most frequent annotated cardiac conditions in the dataset. We assessed two different loss functions and different ECG sampling rates and the best-performing model used double soft F1-loss and ECGs downsampled to 75Hz. This model achieved an F1-score of 0.420 ± 0.017 , accuracy = 0.954 ± 0.002 , and an AUROC score of 0.832 ± 0.019 . An aggregated saliency map, showing the global importance of all 12 ECG leads for the 30 cardiac conditions, was generated using Local Interpretable Model-Agnostic Explanations (LIME). The global saliency map showed that the Inception model paid the most attention to the limb leads and the augmented leads and less importance to the precordial leads. CONCLUSIONS: One of the more significant contributions that emerge from this study is the use of aggregated saliency maps to obtain global ECG lead importance for different cardiac conditions. In addition, we emphasized the relevance of evaluating different loss functions, and in this specific case, we found double soft F1-loss to be slightly better than binary cross-entropy (BCE). Finally, we found it somewhat surprising that drastic downsampling of the ECG led to higher performance than higher sampling frequencies, such as 500Hz. These findings contribute in several ways to our understanding of the artificial intelligence-based interpretation of ECGs, but further studies should be carried out to validate these findings in other datasets from other patient cohorts.