Transport What Shares the Conduction Pathway: Normal-Anchored SPD Geometric Transport for Source-Free Patient Adaptation of Holter and Patch ECG

Abstract

Continuous electrocardiogram (ECG) streams are dominated by normal beats (87.9-98.6% of the label-free adaptation prefix in our four databases), and patient-specific waveform distortion is not shared by every beat class; under this combination, source-free and test-time adaptation methods that assume class-balanced target batches collapse. We propose Normal-Anchored Geometric Transport (NAGT, GA_NS), which estimates a closed-form symmetric positive-definite congruence map from label-free normal anchors in the first 5 min of each recording and applies it only to the normal and supraventricular prototypes, keeping the ventricular and fusion prototypes fixed. Decision rules were set on MIT-BIH and the rate-invariant RR head was adopted after the development databases (INCART, SVDB) had been opened; the frozen model was then evaluated once under a pre-specified locked evaluation protocol on the held-out Icentia11k chest-patch cohort. In 493 patients (973 recordings, 3,615,388 evaluated beats), NAGT raised the arrhythmia macro-F1 J from 34.5% to 48.7% ($\Delta J$ = +14.1 [+8.6, +19.3] pp; non-AF 43.9% to 59.8%) and outperformed all five Holm-adjusted comparators, including AdaBN, Tent and T3A. In post hoc analyses of the frozen outputs, ventricular false alarms fell by 68% (41.1 to 13.2 per hour) and 98.3% of normal beats stayed outside the alarm classes. Adaptation takes 5.8 ms per record without back-propagation.