On the redundancy of denoising in electrocardiogram reconstruction

Electrocardiogram (ECG) reconstruction involves synthesizing specific leads from other recorded leads, addressing cases where standard ECGs have missing, noisy, or unrecorded leads. Traditionally, reconstruction pipelines include a denoising step, which could be time-consuming and may distort the original signal morphology. Recent studies have shown that reconstruction can be successfully performed without denoising by using multiple models tailored to different input noise levels. This paper builds on that approach by evaluating whether a single, unified model can achieve similar performance.

The study compares the multiple-model pipeline (MMP) and the single-model pipeline (SMP) using 10,000 normal ECG records from the CODE-15% database (trimmed to 10 s, resampled to 500 Hz, and denoised to act as a reference point for when artificial noise is added prior to reconstruction).

Results show that the SMP performs comparably to the MMP, achieving average correlations above 0.85, consistent with previous research, and maintaining stable performance across noise levels (∼0.04 variation). Although the model performs poorly when tested on abnormal ECG, fine-tuning the model through transfer learning with as little as ten seconds data returns its performance to the expected level by making the generic model patient specific.

These findings have significant implications for wearable medical devices, where minimizing storage and computational demands is critical. While the MMP requires over five times the memory due to multiple models, the SMP (and transfer learning) offers a more streamlined, efficient, and cost-effective solution by eliminating both the need for multiple models and separate denoising steps.