Assessing ECG Signal Quality Using a Pre-trained Audio Network

Novel ECG wearable systems offer the possibility of continuous cardiac monitoring by acquiring long periods of ECG signal uninterruptedly.However, ECG signals obtained from these wearable devices often suffer from severe noise, requiring automated quality assessment.Recent advancements have demonstrated the effectiveness of convolutional neural networks (CNNs) for that purpose.The transfer learning approach has been widely used in vision and natural language domains, however its application to an ECG signal interpreted as audio data remains relatively unexplored.Therefore, this study aims to analyze the applicability of this technique in the assessment of ECG signal quality.The ECG recordings were conditioned to be inputted to the pre-trained audio network YAMNet, which was fine-tuned to discern between low-and high-quality ECG excerpts.Briefly, after the training process with a proprietary balanced dataset, the model was validated and externally tested using a publicly available database.Results showed a slight unbalance between sensitivity and specificity values, obtaining rates of about 82% and 74% respectively.This performance is similar to that obtained by previous algorithms based on wider and deeper CNN architectures.Hence, audio-based transfer learning seems to work successfully in ECG quality assessment and its use in other healthcare applications could also be explored.