Reconstructing 12-lead ECG from reduced lead sets using an encoder-decoder convolutional neural network

The standard 12-lead electrocardiogram (ECG) is the gold-standard clinical tool for assessing the heart's electrical activity. The primary goal of this study is to reduce the number of recording sites needed to capture the same amount of information as a 12-lead ECG. This approach will simplify monitoring procedures, improve patient comfort, and broaden ECG accessibility across various healthcare settings. In this work, a patient-specific framework is investigated to map a multivariate input to a multivariate output, namely reconstructing the standard 12-lead ECG from any subset of three independent standard ECG leads (i.e. $I,II,V1:V6$ ).The algorithm is evaluated on two datasets: non-public data from 15 patients and the PTB Diagnostic ECG Database, with models trained and tested individually for each patient. The proposed method involves a series of preprocessing steps aimed at eliminating noise and segmenting the data into heartbeats, followed by Short-Time Fourier Transform (STFT) and an encoder-decoder convolutional neural network (CNN) model. The calculated correlation coefficient (CC) and root mean square error (RMSE) confirm the model's accuracy in reconstructing the 12-lead ECG. The proposed model achieves average correlations of 97.6% for the first dataset and 98.9% for the PTB Database using three leads. With a single lead as input, the average correlations reach 97.3% and 98.4% for the respective datasets. These results demonstrate the effectiveness of the proposed methodology in accurately reconstructing 12-lead ECGs using either any three independent leads or any single lead as input.