Sudden cardiac death claims hundreds of thousands of lives each year, often striking without warning in people who felt reasonably healthy. The standard screening tool, left ventricular ejection fraction (LVEF), identifies only a small fraction of those at risk. Most people who die suddenly have normal LVEF. A study published June 24 in Nature by Ziad Obermeyer (UC Berkeley), Markus Lingman (Region Halland, Sweden), Sendhil Mullainathan, and colleagues changes that picture fundamentally.
The team trained a deep learning model on 382,095 electrocardiograms from 111,042 patients in Sweden, then validated it on a separate 251,858 ECGs from Sharp HealthCare in San Diego and a Taiwanese registry of 4,268 tracings from the National Taiwan University Hospital, roughly 634,000 ECGs in total across approximately 250,000 patients. What the model found was not a black-box prediction but a recognizable, previously undescribed ECG morphology that the researchers could extract and interpret.
What the model sees
The newly identified signature combines two features visible on a standard 12-lead ECG: left axis deviation (specifically left anterior fascicular block, LAFB) and posterior rotation (poor R-wave progression). But the most distinctive feature is a slurred terminal aspect of the R wave in lead aVL, replacing the normally sharp negative S wave with a gradual, blunted descent. The researchers quantified this by measuring the summed first and second differences of QRS amplitude from the R peak to the QRS end.
When applied to the Swedish lockbox dataset, the model achieved an AUC of 0.872 (95% CI 0.843-0.899) for predicting sudden cardiac death from death certificates, far outperforming the AHA/ACC 10-year risk score (AUC 0.697) and a standard deep-learning model (AUC 0.655). The highest-risk 2.2% of patients had a 7.0% annual sudden cardiac death rate (95% CI 4.9-9.5%). At the extreme 0.2% of risk, the annual rate reached 11.0%.
The real-world performance held up across continents. At Sharp HealthCare in San Diego, a zero-shot external validation with no retraining, the model achieved an AUC of 0.822 (95% CI 0.812-0.831) for predicting ventricular fibrillation or ventricular tachycardia. The top 2.2% of patients had a 29.1% annual incidence of these arrhythmias. In Taiwan, the AUC for arrhythmic cardiac arrest was 0.767 (95% CI 0.706-0.823), while a placebo test using non-arrhythmic arrests yielded only 0.582 (p<0.001 for the difference), confirming the model's specificity for arrhythmic events.
Catching what LVEF misses
Perhaps the most clinically important finding is what LVEF, the current standard, misses. Only 13.9% of model-identified high-risk patients had reduced LVEF (35% or below). That means 86.1% of the patients the model flagged as high risk would never have been flagged by current screening. And the model’s high-risk group had a higher sudden death rate (7.0%) than patients with reduced LVEF (4.6%, p=0.02). Patients positive on both markers fared worst: a 10.7% annual sudden cardiac death rate.
In an observational analysis from the Swedish cohort, high-risk ECG patients who received defibrillators showed a 54.4% reduction in mortality (6.65% predicted to 3.03% actual, p<0.001). For patients with reduced LVEF who received defibrillators, the reduction was 67.5% (4.43% to 1.44%).
What the biomarker likely reflects
Cardiac MRI review of a subset of high-risk patients found a higher prevalence of subtle, diffuse late gadolinium enhancement, a marker of myocardial fibrosis. The authors propose that the novel ECG signature picks up diffuse fibrotic changes in the heart muscle that current imaging catches only when they become severe enough to reduce LVEF. A single-lead version of the model achieved nearly the same AUC as the full 12-lead version, suggesting the signal reflects a diffuse myocardial process rather than a focal defect.
Caveats
The defibrillator benefit analysis is observational, not randomized, so confounding by indication is possible, patients who got defibrillators may have differed in other ways. As the authors note, “A randomized trial in high-risk patients will be crucial: many promising predictors in the past have failed to identify patients who benefit from defibrillators.” The Taiwan validation set is relatively small and the AUC lower there, raising questions about ethnic generalizability. And the association with myocardial fibrosis is preliminary, based on retrospective MRI review.
Changxin Lai (Johns Hopkins University), writing in an accompanying Nature News and Views article, calls the study a demonstration of “the power of artificial intelligence in uncovering subtle patterns in physiological data that human experts miss” and says it “opens the door to precision prophylaxis for sudden cardiac death.”
Source: Obermeyer, Z., Schubert, A., Ross, J. et al. An ECG biomarker for sudden cardiac death discovered with deep learning. Nature (2026). DOI: 10.1038/s41586-026-10674-6