A self-supervised transformer model trained on over 7.2 million ECGs quantifies how difficult it is to reconstruct a partially masked cardiac signal — and that reconstruction error independently predicts all-cause mortality. Across five cohorts (CODE-15%, MIMIC-IV-ECG, HEEDB, Innsbruck, CHRIS), each one-standard-deviation increase in reconstruction error corresponded to hazard ratios of 1.23–1.41 for death after adjustment for age and sex. A binary high-risk threshold replicated in a sixth cohort, the UK Biobank (HR 1.27, 95% CI 1.08–1.50), with follow-up windows spanning 1.4 to 11 years.
What makes this architecturally significant is the self-supervised framing: most AI-ECG risk tools train against external labels like age or biological sex, effectively learning a proxy rather than intrinsic cardiac signal complexity. Reconstruction error instead captures how far any given ECG deviates from learnable normal electrical patterns — a direct measure of cardiac irregularity agnostic to the labeling choices of the developer. This sidesteps a known generalizability problem in supervised ECG AI, where models can quietly optimize for demographic surrogates.
The multi-cohort external validation — spanning critical care, hospital, cardiology clinic, and population settings across three countries — is unusually rigorous for this class of model. Limitations include retrospective design, reliance on single-timepoint ECGs, and unknown causal mechanisms linking reconstruction difficulty to mortality pathways. As a preprint not yet peer-reviewed, effect sizes and methodological choices warrant independent scrutiny before clinical adoption. Still, the cross-setting consistency suggests this approach is more than incremental.