Reconstructing 12-Lead ECGs from Reduced Leads

Abstract

The standard 12-lead electrocardiogram (ECG) is the clinical gold standard for cardiac assessment because it captures the heart's electrical activity from multiple spatial viewpoints. However, obtaining a full 12-lead ECG requires ten electrodes and careful placement, which can limit convenience in ambulatory, bedside, and wearable monitoring settings. If the missing leads can be reconstructed accurately from fewer measured electrodes, reduced-burden ECG systems may retain much of the information in a conventional 12-lead recording while simplifying acquisition. Although prior studies have reported strong reconstruction accuracy, many evaluate a limited combination of input leads or evaluate reconstruction only over isolated beats or very short waveform segments.

This thesis investigates whether a full 12-lead ECG can be reconstructed accurately from only four measured electrodes and which additional chest lead contributes the most information when the limb electrodes are already available. Using 176 ECGs from the PTB-XL dataset, patient-specific multilayer perceptrons were trained on the first 4 seconds of each recording and evaluated on the remaining 6 seconds. Three architectures were compared: a baseline ReLU MLP, a ReLU MLP with NeRF-style positional encoding, and a SIREN network with sinusoidal activations. Three input regimes were evaluated: all limb leads only, a single precordial lead only, and all limb leads plus one precordial lead. Reconstruction quality was assessed using correlation, normalized mean squared error, peak signal-to-noise ratio, and timing shifts.

The results show that, although the 12-lead ECG contains substantial redundancy, reconstruction performance depends more strongly on lead selection than on model architecture. Limb leads alone reconstruct the lateral chest leads well but perform worst on V2 and V3, while a single chest lead alone is less reliable overall. Adding one precordial lead to the limb leads, especially V2 or V3, produces the strongest four-electrode reconstruction, with average correlations near 0.98 across reconstructed leads. Among single-lead input models, V4 performs best. Across architectures, SIREN and positional encoding provide only modest improvements over a plain ReLU baseline. Together, these findings extend previous reduced-lead ECG studies by evaluating multi-second reconstructions in electrode-constrained settings and by identifying which added chest measurements are most valuable for practical reduced-electrode 12-lead ECG reconstruction.