A Novel Single Lead to 12-Lead ECG Reconstruction Methodology Using Convolutional Neural Networks and LSTM

Abstract

The Electrocardiogram (ECG) is a useful diagnostic tool to diagnose cardiovascular diseases (CVD). Standard 12-Lead ECG setup is most commonly used by doctors for the diagnosis. But the promising type of wearable ECG device uses minimal wire setup on the body to increase patients' comfort resulting in fewer recorded leads, mainly single lead. There is a need to reconstruct the remaining leads from these less recorded leads. Accounting for this, we are proposing a novel Single Lead to 12-Lead ECG reconstruction methodology using convolution neural networks (CNN) and long short term memory (LSTM). In the proposed methodology, lead-II is taken as the basis lead to reconstruct the remaining independent leads (I, V1, V2, V3, V4, V5, and V6). Seven individual models corresponding to the above mentioned seven independent leads have been trained, where each model takes lead-II as input and gives I/V1/V2/V3/V4/V5/V6 as output. Leads III, aVR, aVL, and aVF are reconstructed using a standard approach using original lead II and reconstructed lead I signals, without the need for deep learning models. The proposed methodology was evaluated on myocardial infarction data from PTBDB using R2 statistics, correlation coefficient, and regression coefficient. The mean values averaged across all the 11 leads of the stated performance metrics obtained were 93.62%, 0.973, and 0.959, respectively.