Abstract
Electrocardiogram (ECG) reconstruction involves synthesizing leads from a reduced or alternative lead set. While ECG leads are generally considered linearly related, recording distortions and individual differences make perfect replication difficult, leading researchers to explore deep learning (DL) methods. This paper challenges DL methods by introducing wave masking, a novel preprocessing technique adapted from image recognition, where sections of the input are masked to highlight segments most relevant to improving reconstruction. Applied to ECG, it emphasizes key parts of the time-series signal. The study compares the performance of wave masking combined with linear regression against traditional preprocessing for both linear and DL models, using 10,000 normal ECG records from the CODE-15% database (trimmed to 10 s, resampled to 500 Hz, and denoised). Results show mean correlation values of 0.869 ± 0.201 for the linear pipeline, 0.880 ± 0.190 for the wave masking pipeline, and 0.894 ± 0.168 for the DL pipeline. Wave masking significantly improves linear regression performance by over 0.01 and produces results comparable to DL models, though not superior. These findings highlight wave masking as a promising, low computation preprocessing step for ECG reconstruction. Further research is needed to explore its potential benefits when integrated with deep learning models and diverse demographic records.