Artificial Intelligence Electrocardiogram Triage for Rheumatic Heart Disease Under Portable Device Lead-Configuration Differences: A Transfer-Learning Evaluation of Diagnostic Calibration Across Echocardiography Referral Pathways

Portable single-lead and reduced-lead electrocardiogram devices are increasingly proposed as triage instruments for rheumatic heart disease in settings where echocardiography is scarce, yet the artificial intelligence models that interpret them are almost always trained on twelve-lead recordings acquired in controlled conditions. the present analysis treats the resulting mismatch as a calibration problem rather than a simple accuracy problem, and asks what happens to the operating characteristics of a transferred triage model when the lead configuration at deployment differs from the lead configuration at training. A formal framework is developed in which lead-configuration shift is represented as a structured perturbation of the input distribution, and in which the clinically relevant quantity is the post-transfer calibration curve that determines how many true cases are referred for echocardiography at any fixed referral budget. The framework yields three propositions: that discrimination and calibration degrade at different rates under lead shift; that the degradation is asymmetric across disease prevalence bands; and that recalibration on a small target-configuration sample recovers more of the referral pathway's operating value than retraining on the same sample. A design-based evaluation is specified to test these propositions, with the before-after-control-impact estimator characterised under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64) across 2000 replications per design cell. The contribution is a calibration-centred account of transfer failure that reframes portable electrocardiogram triage as a problem of referral-pathway operating characteristics rather than of raw classification accuracy.