Deep Learning Identifies Novel ECG Biomarker for Sudden Cardiac Death, Published in Nature
核心洞察
A deep learning model discovered a previously unrecognized ECG biomarker that identifies patients at high risk of sudden cardiac death (搜索), with a 7.0% annual event rate versus 4.6% for reduced LVEF.
The model flagged 86.1% of high-risk patients who were missed by the current LVEF-based standard, and high-risk patients receiving defibrillators showed a 54.4% reduction in mortality.
Validated across three continents (Sweden, USA, Taiwan), the model achieved an AUC of 0.872 in the Swedish lockbox and zero-shot AUCs of 0.822 (USA) and 0.767 (Taiwan).
A deep learning model has uncovered a novel electrocardiogram biomarker that markedly improves prediction of sudden cardiac death (搜索), according to research led by Ziad Obermeyer at the University of California, Berkeley and published in Nature in July 2026. The model identifies a high-risk population that the current standard of care—left ventricular ejection fraction (搜索) (LVEF)—systematically overlooks, potentially opening the door to more targeted, life-saving defibrillator (搜索) therapy.
The study linked comprehensive ECG records from Region Halland, Sweden, with corresponding death certificates, enabling the development of a predictive model trained on 262,554 ECGs from 75,157 patients. Results were generated in a 40% data lockbox comprising 119,541 ECGs from 35,885 patients under 80 years old, which remained untouched until provisional acceptance of the manuscript.
A High-Risk Group Hidden in Plain Sight
The model's preferred high-risk group comprised 2.2% of the analyzed sample and experienced an annual sudden cardiac death (搜索) rate of 7.0% (95% CI: 4.9–9.5%). This contrasts sharply with the 4.6% annual rate observed in patients with reduced LVEF, the current cornerstone for defibrillator (搜索) eligibility. Critically, 86.1% of the model's high-risk patients were not flagged by LVEF, underscoring a substantial gap in contemporary risk stratification.
The model achieved an area under the receiver-operating-characteristic curve (AUC) of 0.872 (patient-level bootstrapped 95% CI: 0.843–0.899). By comparison, the American Heart Association–American College of Cardiology 10-year risk score yielded an AUC of 0.697, and a validated ECG-based deep-learning model for heart disease (SEER) achieved 0.655.
Cross-Continental Validation
The model's performance was validated in two external datasets without any fine-tuning. In a US cohort from Sharp HealthCare (San Diego, CA) comprising 251,858 ECGs from 139,613 patients, the zero-shot AUC for ventricular fibrillation or ventricular tachycardia (VF/VT) was 0.822 (95% CI: 0.812–0.831). Using the same 2.2% high-risk threshold, the annual VF/VT incidence reached 29.1% (95% CI: 26.5–31.9%), compared with a base rate of 3.8%.
In a Taiwanese hospital-based registry at National Taiwan University Hospital, the model distinguished future arrhythmic arrests from controls with a zero-shot AUC of 0.767 (95% CI: 0.706–0.823). A placebo test using non-arrhythmic arrests yielded an AUC of 0.582 (95% CI: 0.529–0.636), significantly worse (P < 0.001), confirming the model's specificity for arrhythmic deaths.
Mortality Reduction with Defibrillators
Although the study lacked randomization, an observational analysis provided suggestive evidence of clinical benefit. High-risk patients who received defibrillators experienced a 3.62 percentage point reduction in sudden cardiac death (搜索) relative to their predicted rate of 6.65 percentage points—a 54.4% reduction (P < 0.001). For all-cause mortality, the reduction was 12.6 percentage points (39.0%) relative to a predicted rate of 32.4 percentage points.
"High-risk patients seem to represent a new, previously unsuspected population with frequent, preventable death," the authors note, while emphasizing that a randomized trial will be crucial to confirm these findings.
Visualizing the Discovery
To translate the model's correlation into a clinically interpretable finding, the researchers paired the predictive model with a generative variational auto-encoder. This "morphing" procedure produced counterfactual ECG waveforms that progressively transitioned from low-risk to high-risk beats, isolating the risk signal.
The high-risk morph exhibited left axis deviation consistent with left anterior–superior fascicle block and posterior rotation. Most notably, lead aVL revealed a slurred terminal aspect of the R wave—a morphology that, unlike axis deviation, has not been previously described in the literature. Quantifying this feature using first and second differences of QRS amplitude confirmed it as a robust, independent predictor of sudden cardiac death (搜索) across both Swedish and US cohorts.
A Mechanistic Hypothesis: Diffuse Fibrosis
Blinded review of cardiac MRIs in a subset of high-risk patients revealed a significantly higher prevalence of subtle, diffuse late gadolinium enhancement throughout the left ventricle, a pattern most often associated with myocardial fibrosis. The researchers hypothesize that diffusely distributed fibrotic obstacles to conduction could explain both the ECG biomarker and the elevated arrhythmic risk.
"Fibrosis—a diffusely distributed phenomenon that alters conduction—is one mechanism that ties together the ECG biomarker and sudden cardiac death (搜索)," the authors write, while acknowledging the speculative nature of this hypothesis and calling for future work correlating ECGs with endomyocardial biopsy.
The study's findings suggest that deep learning can surface clinically actionable signals embedded in routine medical tests, identifying patients at lethal risk who are invisible to current guidelines.