Low-rank matrix features and quality-controlled RR-interval models in short multilead ECGs: Matched construct validation and external rhythm comparisons

Low-rank matrix features summarize multilead ECG structure, but their incremental value beyond RR-interval features and performance across competing rhythms remain uncertain.

We analyzed a matched PTB-XL construct-validation cohort (250 development and 100 held-out pairs). Separate RR-only and RR-plus-static models were developed using 951 age- and sex-matched pairs from PTB-XL folds 1-8 and tested without refitting in the independent, multicenter PhysioNet ECG Arrhythmia release against seven rhythm groups.

In the PTB-XL construct-validation analysis, held-out AUROC was 0.692 for static, 0.604 for dynamic, and 0.958 for quality-controlled RR features. Development cross-validated AUROC was 0.989 for both RR-only and RR-plus-static models. In external matched comparisons, RR-only AUROC ranged from 0.527 for AF versus atrial flutter to 0.985 versus sinus bradycardia. Adding static features changed AUROC by -0.004 (95% CI, -0.006 to -0.002) for pooled competing rhythms and +0.089 (95% CI, +0.078 to +0.100) for supraventricular tachycardia.

Static features showed moderate construct validity but did not add discrimination consistently beyond RR features. External performance varied by competing rhythm, with poor separation from atrial flutter. The supraventricular-tachycardia increment is exploratory and requires confirmation.