ECG Recoverability

Feature-level recoverability and uncertainty calibration for reduced-lead ECG reconstruction.

Research question

Reduced-lead ECG reconstruction is an ill-posed inverse problem. A low average waveform error does not by itself identify which P, QRS, ST, or T features remain trustworthy on a particular lead. This project studies feature-level recoverability rather than treating reconstruction quality as a single scalar.

Approach

The research code combines an ECG dipole model, a conditioning measure for lead configurations, and distribution-free calibration. It separates model-supported components from statistically estimated and unrecoverable components, then evaluates those distinctions using public PTB-XL data, synthetic checks, and a conditional-diffusion illustration.

Repository-reported experiment comparing perceptual reconstruction behavior with a held-out recoverability analysis.

Evidence and boundaries

The repository includes theorem-linked tests and experiments on a public research dataset. It is a research artifact and preprint in preparation, not a clinically validated device, diagnostic system, treatment recommendation, or published medical result. Dipolarity varies by feature and pathology; calibration is tied to the stated data and grouping assumptions. Nothing on this page is medical advice.

Artifacts and provenance