A hybrid machine learning framework for single-lead ECG signal quality assessment

Wearable devices are increasingly being used to monitor electrocardiogram (ECG) signals in real time, enabling earlier diagnosis and more effective monitoring of heart health. However, single-lead ECG signals captured by these devices are often contaminated with noise and artefacts, which can degrade signal quality and compromise the accuracy of subsequent diagnostic steps. To prevent this, it’s essential to assess the quality of ECG signals before further processing. Manual evaluation of signal quality can be laborious and prone to human error, especially in continuous monitoring contexts with large volumes of data. As a result, developing a streamlined approach for classifying ECG signal quality is critical to improving clinical workflows, reducing human error, and ensuring that diagnostic algorithms receive high-quality input for accurate and timely health assessments. This thesis focuses on developing a robust and efficient system for the automatic classification of single-lead ECG signals based on their quality. The datasets used in this work included publicly available single-lead ECG signals, each pre-processed to ensure consistency across various data acquisition environments. Pre-processing steps were applied separately to each dataset to ensure that data integrity was preserved. Additionally, each trace was handled carefully to ensure that no data was shared between the training, validation, and test sets, thereby maintaining the reliability of the performance evaluation. The primary goal of this study is to develop a system capable of accurately distinguishing between high-quality, borderline, and unacceptable ECG signals in single-lead settings. To this end, the study employs a convolutional neural network (CNN) model for initial quality classification, combined with a random forest (RaF) algorithm to refine the classification of borderline signals. The CNN model was enhanced using Generative Adversarial Network (GAN)-based data augmentation to balance the dataset and improve generalization, with traditional data augmentation also evaluated. The CNN model, when trained using GAN-based data augmentation, achieved an accuracy of 90%, an Area under the curve (AUC) score of 96%, and a Recall of 97% on the test set. The RaF classifier, used to further enhance the classification of borderline signals, demonstrated a validation accuracy of 86% and an AUC score of 94%. The ensemble of CNN and RaF models performance indicated robust results, particularly for high-quality signals, where the model achieved a precision of 91%, recall of 86%, and F1-score of 88%, resulting in an overall accuracy of 89% with near-perfect recall (99.9%) for unacceptable signals. In addition to classification performance, the computational efficiency of the proposed method was evaluated. The total processing time for a single 5-second ECG signal was approximately 0.2458 seconds, requiring only 1.42 MMACs and 2.89 MFLOPs. This thesis makes a significant contribution by providing an effective three-class classification algorithm for ECG signal quality assessment in resource-constrained settings, with potential applications in real-time health monitoring. Future work will focus on further optimizing the model for deployment on wearable devices, considering factors such as memory usage and energy consumption on a stand-alone wearable device.