Background and Aim. Convolutional neural networks (CNNs) have recently become popular for 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. This challenging topic, aimed at identifying poor-quality ECG intervals, could facilitate more accurate decisions on cardiac disorders. This work introduces a comparative study about how several CNN algorithms discern between high and low-quality ECGs.
Methods. To take advantage of the concept of transfer learning, five common pre-trained CNNs were analyzed, such as VGG16, ResNet18, AlexNet, GoogleNet, and InceptionV3. They were fed with 2D images obtained by turning 5 s-length ECG segments into scalograms through a continuous Wavelet transform. The ECG intervals were selected from the training set proposed for the PhysioNet/CINC Challenge 2017. Although this database is composed of 8,528 recordings, only a minority of them were labeled as noisy signals. Thus, whereas all available noisy 5 s-length ECG intervals (1,168) formed the poor-quality group, 1,200 high-quality segments were selected from the remaining recordings. A total of 2,368 ECG intervals were then studied.
Results. Five learning-testing cycles were conducted with random 80/20 splits of the dataset. Whereas all algorithms provided mean values of accuracy between 89 and 91%, notable differences were seen in terms of computation time and CPU usage. Thus, AlexNet was the fastest algorithm, moreover requiring 10% less of CPU usage. It should also be noted that the obtained values of accuracy were comparable or slightly better than those reported by other works introducing non-CNN-based techniques.
Conclusions. AlexNet has reported the best trade-off between poor quality ECG identification accuracy and computational load, and therefore it is the most convenient CNN-based approach for ECG quality assessment.