LeadWeaver: Topology-Aware Masked-Lead Modeling for Robust 12-Lead ECG Diagnosis

Abstract

Clinical electrocardiography is organized as a structured set of correlated leads, yet most neural ECG classifiers assume that all twelve leads are present, correctly ordered, and acquired under the same device conditions at training and deployment. We propose LeadWeaver, a topology-aware ECG representation model designed for diagnosis under missing, corrupted, or device-shifted leads. LeadWeaver represents each lead as a temporal token stream, couples leads through a fixed-plus-learned physiological lead graph, and performs masked-lead latent completion rather than waveform synthesis. A reliability router estimates which observed leads should dominate each diagnostic decision, while a consistency objective aligns complete-lead and incomplete-lead predictions. We evaluate the model on PTB-XL and CPSC-style settings using complete-input diagnosis, arbitrary missing-lead subsets, lead corruption, cross-dataset transfer, calibration, and selective prediction. The central hypothesis is that robust ECG diagnosis requires explicitly modeling both inter-lead topology and uncertainty about unavailable evidence, rather than treating missing channels as zero-valued inputs.