Abstract
Artificial intelligence (AI) algorithms have demonstrated remarkable efficiency in analyzing 12-lead clinical electrocardiogram (ECG) signals. This has sparked interest in leveraging cost-effective and user-friendly smart devices based on single-lead ECG (SL-ECG) for diagnosing heart dysfunction. However, the development of reliable AI model is influenced by the limited availability of publicly accessible SL-ECG datasets. To address this challenge, presented study introduces a novel approach that utilizes 12-lead clinical ECG datasets to bridge this gap. We propose a hierarchical model architecture designed to translate SL-ECG data while maintaining compatibility with 12-lead signals, ensuring a more reliable framework for AI-driven diagnostics. The proposed sequential model utilizes a convolutional neural network enhanced with three integrated translational layers, trained on individual 12-lead clinical ECG, to significantly improve classification performance on SL-ECG. The experimental analysis is conducted using three benchmark datasets: Physikalisch-Technische Bundesanstalt (PTB-XL), Computing in Cardiology Challenge 2017, and China Physiological Signal Challenge 2018. This study also evaluates the effects of denoising techniques and lead polarity variations, including biphasic and negative deflections. Results show that the model achieved over 82% test accuracy on unseen SL-ECG signals, with an area under the receiver operating characteristic of 0.81, sensitivity of 76.60%, and specificity of 83.44% when trained on clinical lead I. Additionally, leads II, V4, and V5 demonstrated potential for effective AI model training. The work supports advancement of smart devices by enhancing SL-ECG classification and assists clinicians in assessing heart abnormalities more effectively.