Signal Reconstructions for ECG-Transmembrane Voltage Potentials Using Transformers

Transmembrane voltage potential (TMP) data play a vital role in detecting heart conditions such as coronary heart diseases, arrythmias, and heart failures. However, defects and limitations in electrocardiogram devices can adversely affect data collection, often resulting in data that are noisy and time-step limited. Such corrupted data can impair a clinician's ability to correctly diagnose and further understand a patient's heart condition. In this paper, we propose two machine learning architectures for recovering the true TMP signal from corrupted ECG data. In particular, we consider (1) a long short - term memory model, which takes in the entire ECG measurement data and produces the TMP signal in a 1-to-1 fashion; and (2) a transformer-based model, which takes in the ECG signal and produces the TMP signal in a generative fashion. Numerical experiments demonstrate that these approaches out-perform current state-of-the-art 1D convolutional neural network.