The 12-lead electrocardiogram (ECG) is a fundamental tool in clinical cardiology, providing multi-lead perspectives of the cardiac electrical activity from different anatomical landmarks. We propose a neural network-based model that approximates each lead from the others by leveraging intrinsic redundancies among the leads. The goal is to decompose each lead into its redundant components (which can be recovered from the other leads) and lead-specific components (which are unique to each lead). Hypothetically, both components hold clinical value and cardiac anomalies impact what can or cannot be inferred from other leads.
The model architecture comprises three stages. First, a linear regression, which generalizes the well-known Dower's transform. Second, a multilayer perceptron (MLP) layer, which learns the residual nonlinear inter-lead dependencies. Third, a temporal recurrent neural network (RNN) layer, which captures temporal correlations of the lead residues. The outputs of all stages are aggregated to reconstruct the target lead.
The model is evaluated on the PTB-XL dataset (21,837 12-lead ECGs). The average lead reconstruction signal-to-noise ratio (SNR) improves across the linear, MLP and RNN layers and varies across the leads. The best performance is observed on lead V5 (average SNR = 16.16 dB), followed by V4, V6, V3, V2, V1, I, with II being the most difficult to reconstruct (average SNR = 7.33 dB).
The proposed model can serve as a building block for more sophisticated ECG diagnostic networks. The architecture can be used as pretrained layers for transfer learning or foundation models, allowing the layers to be frozen, and later fine-tuned for specific applications. By disentangling shared from lead-specific information, it enhances lead-specific reasoning and helps identify which leads contribute unique diagnostic value. The model also facilitates ECG compression by reducing inter-lead redundancy, making it well-suited for both diagnostic pipelines and efficient transmission and storage in resource-constrained applications.