Background: Diagnostic 12-lead ECG analysis is one of the most important clinical cardiology tests. However, one of the challenges with a 12-lead ECG is attaching multiple electrodes to obtain the signals. This study aims to use a deep learning model (DNN)-based approach to directly generate interpretation results from the input waveform for a reduced lead set, without relying on the criteria built for a full 12-lead, and to compare the results with those from a DNN-based full 12-lead ECG.
Methods: A DNN classification model was developed to make determinations for selected morphology classes using 10-second ECG recordings with input lead sets that included the full limb leads and two precordial leads, V2 and V4. A second DNN model was built for the standard 12-lead set [I, II, v1-v6], with signal preprocessing that formed the median beats for all leads. The training dataset consisted of two million 12-lead ECGs from the Mayo Clinic, and the test dataset included 500,000 ECGs, also from the Mayo Clinic. Classification performance was assessed using the Area-Under-the-Curve (AUC) and Precision-Recall (AUCPR) for detecting clinical abnormalities.
Results: The results, as shown in the table, indicated that all reduced lead sets exhibited the same excellent performance on major morphology-related abnormalities, such as bundle branch block (AUC > 0.98), ischemia (AUC > 0.97), and myocardial infarction (old and acute, AUC > 0.98), as the full 12-leads.
Conclusion: Based on the performances of both the reduced lead and full 12-lead classification models, the answer to the question 'Does a Reduced ECG Lead Set Contain Full 12-lead ECG Information for Interpretation?' is 'Yes.'