Post-pandemic generalization of a deep-learning ECG biomarker of sudden cardiac death: a pre-registered replication and extension testing convergent electrical remodelling and COVID-19 substrate shift
Amr Kamel Khalil Ahmed, Khulood Almutairi
Abstract
Structured abstract
Background
A deep-learning model trained on pre-pandemic Swedish electrocardiograms (2010 to 2016) identifies a biomarker of sudden cardiac death, a slurred terminal R-wave downstroke in lead . 1aVL with reduced first and second QRS differences, attributed to diffuse myocardial fibrosis, shifts the cardiovascular 2,319 produces diffuse myocardial inflammation and fibrosis-COVID, and leaves measurable, if modest, electrocardiographic changes after 2,4survivorsrisk profile of . Whether the specific aVL biomarker behaves as a convergent, 5,6severe infection-to-moderatemulti-substrate signature, and whether it shifts in the post-COVID myocardium, is untested.Hypotheses
H1 (open data, primary mechanistic test): the reduced-derivative aVL signature is present across multiple non-COVID disease substrates rather than a single fibrosis-specific group. H2 (clinical, primary): among patients after moderate-to-severe or hospitalised COVID-19 with cardiac magnetic resonance evidence of myocarditis or fibrosis, the aVL biomarker and standard repolarisation measures differ from matched non-COVID controls. The direction is hypothesised but the magnitude is treated as unknown and may be null.Methods
11XL (21,799 ECGs)-data analysis on PTB-specified components. Study A: an open-Two precomputing the aVL first and second differences across diagnostic superclasses with a slur-versus-notch falsification. Study B: a retrospective, multi-centre, matched cohort of digital 12-lead ECGs from adults after moderate-to-severe or hospitalised COVID-19 with paired cardiac magnetic resonance, compared with comorbidity-matched non-COVID controls, with the same biomarker plus repolarisation endpoints.
Outcome-neutral statement
The study is powered to detect a moderate effect and is reported in full whatever the result. A null finding constrains the deployment risk of the original model and is itself informative.