On the Application of Convolutional Neural Networks for 12-lead ECG Multi-label Classification Using Datasets From Multiple Centers

Abstract

Cardiac arrhythmia is a group of conditions in which falls changes in the heartbeat. Electrocardiography (ECG) is the most common tool used to identify a pathology in the cardiac electrical conduction system. ECG analysis is usually manually performed by an expert physician. However, manual interpretation is time-consuming and challenging even for cardiologists. Many automatic algorithms relying on handcrafted features and traditional machine learning classifiers were developed to recognize cardiac diseases. However, a large a priori knowledge about ECG signals is exploited. To overcome this main limitation and provide higher performance, recently, deep neural networks were designed and applied for 12-lead ECG classification. In this study, we designed decoding workflows based on three state-of-the-art architectures for time series classification. These were InceptionTime, ResNet and XResNet. Experiments were conducted using the training datasets provided during the PhysioNet/Computing in Cardiology Challenge 2020. The best-performing algorithm was based on InceptionTime, scoring a training 5-fold cross-validation challenge metric of 0.5183±0.0016, while using a low number of parameters (510491 in total). Thus, this algorithm provided the best compromise between performance and complexity.