I-GraphECG: Observability-Based Lead Selection and Interpretable Disease Prediction from the 12-Lead ECG Using a...

Wearable ECG increasingly records fewer than the standard 12 leads, so a model must be accurate, interpretable, and explicit about the information lost under lead reduction. I-GraphECG is a gray-box graph electrophysiology surrogate: an encoder maps a 12-lead median beat to a bounded vector of 47 equivalent electrophysiological descriptors, a constrained eight-node conduction-graph decoder reconstructs the signal, and disease prediction uses the descriptors rather than the waveform. On a clean PTB-XL four-class subset, median beats are reconstructed at median correlation 0.91, and macro-AUROC reaches ≈0.90, near the black-box references (0.920-0.924). Myocardial infarction is under-detected (recall ≈ 0.49), consistent with the ST-source's low observability, so the model must not be used as a stand-alone rule-out for infarction. Observability-based sensor selection makes explicit what each lead set can resolve: lead removal provably cannot lower any Cramér-Rao bound; under the data-derived noise models, every optimal three-lead set retains one of the precordial leads V2-V5, and the conventional reduced set is never observability-optimal across seven noise models, though the exact montage is noise-model-dependent. Reconstruction and the observability analysis of the fixed model transfer zero-shot to US and Chinese cohorts; parameter-only disease prediction degrades externally, and infarction, untestable in those cohorts, remains internally demonstrated.