Classification of 27 heart abnormalities using 12-lead ECG signals with combined deep learning techniques

Abstract

An electrocardiogram (ECG) machine with a standard 12-lead configuration is the primary clinical technique for diagnosing abnormalities in heart function. Automated 12-lead ECG machines have the capacity to screen the general population and provide second opinions for physicians. However, expertise and time are required for manual ECG interpretation. Therefore, computer-aided diagnoses are of interest to the medical community. Hence, this study aims to build a deep learning (DL) model with an end-to-end structure that can categorize 12-lead ECG results into 27 different disorders. We use multivariate time-series data to construct a novel end-to-end DL model (based on combined convolutional neural networks (CNNs), long short-term memory, gated recurrent units, and a deep residual network structure) for feature representations and determining spatial relations among deep features. In addition, a dataset of 43,101 classified standard ECG recordings was collected from six different sources to guarantee the model’s ability to generalize and alleviate data divergence. As a result, the residual network-based model obtained promising outcomes and an accuracy of 0.97. According to the experimental data, it outperforms other methods.