Performance improvement of deep learning based multi-class ECG classification model using limited medical dataset

Medical data often exhibit class imbalance, which poses a challenge in classification tasks. To solve this problem, data augmentation techniques are used to balance the data. However, data augmentation methods are not always reliable when applied to bio-signals. Also, bio-signal such as ECG has a limitation of standardized or normalized methods. The present study endeavors to tackle the difficulties associated with imbalanced and limited medical datasets. Our study is to compare different approaches for addressing class imbalance in medical datasets, and evaluate the efficacy of various techniques and models in overcoming these challenges. To this end, three experiments with different configurations were considered, that is, a change in the loss function (Experiment A), the amount of data in each class (Experiment B), and the applied grouping methods (Experiment C). Inception-V3 was used as our main model, and three dataset groups were utilized: an imbalanced dataset with a large amount of data, a balanced dataset with limited data, and a dataset with a subclass bundled with a small amount of data. We propose an improved method using focal loss for an imbalanced classification. The F1 score was 0.96 for Inception net with focal loss and 0.86 in a limited data environment with the same ratio.