The electrocardiogram (ECG) is considered the gold standard for cardiac arrhythmia diagnosis; however, its manual interpretation is slow, subjective, and prone to human error. Deep learning (DL) techniques, especially convolutional neural networks (CNNs), have demonstrated significant performance in automatic classification and detection of biomedical signals. Coleeg is an open-source framework that was originally designed for EEG signal classification; however, its modular design makes it easy to extend for other types of signals. This paper utilizes the Coleeg framework to investigate cross-subject classification using a one-dimensional CNN (1D-CNN) for three arrhythmia databases, which are MIT-BIH, INCART, and PTB-XL, and it presents the features introduced in Coleeg Versions 6.0 and 7.0, which are related to this research work. The main problem encountered in the classification is the extreme imbalance in the databases, where normal beats dominate database classes, which reduces the classification performance for other beat types. To mitigate the data imbalance, the normal beat class count was reduced, and class weights were implemented during training. Data augmentation using the Synthetic Minority Over-sampling Technique (SMOTE) was evaluated, but it did not yield a significant or consistent improvement in the evaluation metrics. Because of the fluctuation in validation loss, early stopping was used to obtain the best classification results. The model evaluation was performed in a systematic way for the three databases, where they were resampled to 250 Hz, and the subjects were divided into 5 groups for a fivefold evaluation. One group is used for validation, one for testing, and the remaining three are for training. Each database evaluation was repeated 5 times for 1-, 2-, and 3-s time windows. The details of the numerical values for the minimum, mean, maximum, and sample standard deviation of recall, precision, F1-score, and accuracy are presented in this paper.
更多
原文链接
分享
引用
