Deep Learning for 12-Lead ECG Classification: Uncovering Learned Features Aligned with Diagnostic Criteria

This study analyzes a deep learning model for 12-lead ECG classification, revealing learned features that closely resemble established diagnostic criteria. The model, trained on a large dataset of ECG recordings, demonstrated high accuracy in classifying different cardiac conditions. To understand the model's decision-making process, the research team investigated the learned features extracted by the deep neural network. The analysis showed that these features, represented as activation patterns in specific layers of the network, closely aligned with diagnostic criteria used by cardiologists for interpreting ECG signals. For example, the model learned to identify QRS complex abnormalities associated with myocardial infarction, as well as ST segment changes indicative of ischemia. This alignment between learned features and clinical diagnostic criteria enhances the interpretability of the deep learning model, suggesting that it may be able to provide valuable insights into the underlying pathophysiology of cardiac disease. The study further highlights the potential of deep learning for supporting clinical decision-making in cardiology.