Comparative Study of Convolutional Neural Networks for ECG Quality Assessment

In the last years, convolutional neural networks (CNNs) have become popular in ECG analysis, since they do not require pre-processing stages, nor specific pre-training. However, their ability for ECG quality assessment has still not been thoroughly assessed. Hence, this work introduces a comparison about the ability of several CNN algorithms to classify between high and low-quality ECGs. Taking advantage of the concept of transfer learning, five common pre-trained CNNs were analyzed, such as AlexNet, GoogLeNet, VGG16, ResNet18 and InceptionV3. They were fed with 2-D images obtained by turning 5 second-length ECG segments into scalograms through a continuous Wavelet transform. To train and validate the algorithms, 1,168 noisy ECG intervals, along with other 1,200 ECG excerpts with sufficient quality for their further interpretation, were extracted from a public database. The obtained results showed that all CNNs provided mean values of accuracy between 89 and 91%, but notable difference in terms of computational load were noticed. Thus, AlexNet was the fastest algorithm, requiring notably less CPU usage and memory than the remaining methods. Consequently, this CNN exhibited the best trade-off between high-quality ECG identification accuracy and computational load, and it could be considered as the most convenient algorithm for ECG quality assessment.