Curriculum Learning using Real and Simulated Data in Deep Learning Models for Electrocardiography Classification

Obtaining well-balanced datasets with quality-assured and clinically-validated labels remains a challenge in healthcare. Many datasets suffer from a class imbalance with a majority of healthy subjects. Synthetic data, which is artificially generated, might be a suitable solution to counter class imbalance; however, the degree to which synthetic data is "accurate enough" to actually improve model training is unclear in many areas. In this work, we evaluate in how far synthetic electrocardiography signals as a supplement to real-world data can improve performance of a deep neural network for electrocardiography classification. Real data stems from the public PTB-XL dataset and synthetic data is based on electrophysiological simulations from the recently-published MedalCare-XL dataset. We measure the classification performance of four different classes including three cardiac conduction abnormalities and evaluate different oversampling and training data shuffling strategies, borrowing from the concept of Curriculum Learning. Results indicate that adding synthetic data yields a modest yet consistent accuracy boost up to 0.7%. Even though synthetic signals are not as complex as real-world measurements, they might be promising to balance class distributions in the context of rare disease.