Abstract 15042: Heart Rhythm Classification From an Optimal Lead Subset of the 12-lead Electrocardiogram by Deep Learning
Introduction: Deep learning (DL) has achieved promising performance on common heart rhythms classification using 12-lead electrocardiogram (ECG). However, two major concerns hinder the DL’s application - lack of interpretability and overfitting caused by using the full 12-lead ECG as input. Objective: We proposed a hybrid DL model with enhanced interpretability to detect 9 common types of heart rhythms from an optimal ECG lead subset, and to quantitively analyze the overfitting. Methods: We used a multicenter dataset of 6,877 annotated 12-lead ECG recordings. The proposed model (Fig. 1A) consists of a feature extraction and a decision-making. The feature extraction used 12 separate neural networks to extract features from each lead. The features were then fed into a random-forest classifier in the decision-making step to classify heart-rhythm types. The classifier was used to interpret the correlations between the heart rhythms and the ECG leads, to find an optimal subset of ECG leads, and to analyze whether using 12-lead ECG added unnecessary complexity to the model and undermined its generalizability. Results: The proposed model detected the correlations between the heart-rhythm types and the ECG leads (Fig. 1B), and identified an optimal ECG lead subset (leads II, aVR, V1, V4). The optimal subset was, in comparison with using 12-lead ECG, significantly better (F1 =0.776 vs. F1 = 0.767, P=0.02) on the validation set for classifying the 9 common heart rhythms. There was no statistical difference on the test set. No overfitting caused by 12-lead ECG was detected in this study. Conclusion: The hybrid DL model based on an optimal 4-lead ECG can interpret rhythm types without significant loss of accuracy in comparison with the 12-lead ECG.