Comparison of Pre-Trained Deep Learning Algorithms for Quality Assessment of Electrocardiographic Recordings

Abstract

[EN] Convolutional neural networks (CNNs) have recently become popular for electrocardiogram (ECG) analysis. However, their ability for ECG quality assessment has still not been thoroughly assessed. This challenging topic, aimed at identifying poor-quality ECG intervals, could facilitate more accurate decisions on cardiac disorders, especially from wearable long-term ECG monitoring systems. Hence, this work introduces a comparative study about how several CNN algorithms discern between high and low-quality ECGs. Because no many datases with large amount of annotated ECG signals are freely available, three pretrained CNNs, such as AlexNet, VGG16 and GoogLeNet, have been compared. For that purpose, 2,000 5 second-length ECG excerpts were extracted from a public database and classified into two groups. Thus, 1,000 ECG intervals were annotated as noisy and other 1,000 segments presented sufficient quality for further analysis. For robust validation of the algorithms, five learningtesting cycles with random 80/20 splits of the dataset were conducted. Whereas all CNN models achieved mean accuracies about 90%, remarkable differences were seen in terms of computational time and memory usage. Thus, AlexNet was the fastest method classifying ECG excerpts, but GoogLeNet required the lowest memory content for its performance. Consequently, both algorithms reported an interesting trade-off between poor-quality ECG identification accuracy and computational load.