Comparative Assessment of Machine Learning Strategies for Electrocardiogram Denoising

An electrocardiogram (ECG) is an important non-invasive predictor of cardiovascular disease (CVD) used to support early diagnosis and detection of various heart problems. Monitoring ECG continuously is expected to lower mortality from CVD, but achieving this aspiration is constrained by the high cost of medical-grade ECG. Although advancements in wearable devices have made ECG monitoring in everyday environments possible, the resulting recordings are affected by severe noise corruption, for which traditional signal processing techniques fall short. Therefore, in recent years, the focus has been on machine learning (ML) techniques for ECG denoising. Despite recent advances, many unanswered questions and unsolved challenges exist, and a comparative study is missing. To address this gap, we comparatively assessed state-of-the-art ML models, namely Denoising Autoencoder (DAE), Convolutional Neural Network (CNN), Long Short Term Memory (LSTM), and Generative Adversarial Network (GAN), for ECG denoising using the MIT-BIH Arrhythmia and ECG-ID datasets. Three noise types were considered, including baseline wander, electron motion, and motion artifacts. Performance was assessed explicitly by comparing denoised and clean signals, and implicitly through the balanced accuracy of downstream tasks (beat classification, person identification). Furthermore, we investigated the models’ generalisation capabilities to unseen data (data transferability) and unseen noise types (noise transferability). Our findings suggested that in certain cases, explicit evaluation may be insufficient and implicit metrics need to be considered. Transfer learning improved data transferability, while all models could generalise to unseen noise types, albeit in different levels. Overall, CNN and GAN models achieved the best performance. Our results encourage the development of denoising and processing pipelines for healthcare applications based on wearable ECG.