Abstract Sudden cardiac death is, in theory, preventable with defibrillators. But every year, many patients die without defibrillators because doctors fail to predict their risk 1 . The only predictive biomarker in wide use, cardiac left ventricular ejection fraction (LVEF), misses most sudden cardiac deaths 2 , and flags many low-risk patients for futile defibrillators that never fire 3,4 . Here we apply deep learning to a dataset linking all electrocardiograms (ECGs) in a Swedish region to death certificates. The resulting model isolates a high-risk group (2.2% of the sample) with a 7.0% annual rate of sudden cardiac death, higher than those with reduced LVEF (1.9% of the sample; 4.6% annual rate). Notably, 86.1% of the model’s high-risk patients were not flagged by LVEF. High-risk ECG patients with defibrillators implanted were 54.4% less likely to die than expected, suggesting a mortality benefit. We externally validate the model in a US health system, in which it predicts ventricular arrhythmias that cause sudden death; and a Taiwanese hospital registry, in which it specifically predicts future arrhythmic cardiac arrests. To visualize the waveform morphology ‘discovered’ by the predictive model, we pair it with a generative model of the ECG waveform. Together, they reveal a biomarker that is easily visible and robustly predicts sudden cardiac death, but has not to our knowledge been previously described. Tying the biomarker’s shape to electrophysiological first principles, we form and preliminarily test a new hypothesis on the mechanism of sudden cardiac death.
Strengths
The Swedish lockbox was created before analysis, held untouched through peer review, and opened only after provisional acceptance. This materially strengthens the credibility of the headline predictive-performance claim.
↳ Methods, Study cohort and outcomes: Sweden; Predictive performance in the Sweden data lockbox
The model identifies a 2.2% high-risk group with a 7.0% annual sudden cardiac death rate, while only 13.9% of that group had known reduced LVEF. This supports the claim that the ECG model finds a population not captured by the dominant clinical biomarker.
↳ Abstract; Comparison of ECG model and LVEF; Figure 3
The model is tested without fine-tuning in US health-system data and a Taiwanese cardiac-arrest registry, using related but different arrhythmic outcomes. The Taiwan placebo AUC analysis against non-arrhythmic arrests adds a targeted specificity check.
↳ Model generalization to US and Taiwan data; Figure 2
The generative morphing analysis highlights a slurred R-wave downstroke in lead aVL and links it to testable hypotheses about conduction obstacles and fibrosis. This creates a scientific path beyond black-box prediction.
↳ An ECG biomarker discovered by the model; Clinical and scientific implications; Figures 4–5
Limitations
The 54.4% mortality reduction for high-risk ECG patients with defibrillators comes from observational OLS regression, not randomization. The paper acknowledges confounding, so this result should guide future trials rather than clinical adoption.
↳ Defibrillators and mortality in high-risk patients; Table 2
The generative model and morphing procedure are explicitly run on the initial training/validation split rather than the Swedish lockbox. This makes the biomarker-visualization component more exploratory than the headline predictive validation.
↳ Methods, Generative model and morphing procedure; An ECG biomarker discovered by the model
The Swedish data, US data, and Swedish-derived deep-learning model access are restricted through collaboration or partnership pathways. These restrictions are understandable for clinical data but limit independent replication of the full pipeline.
↳ Data availability; Code availability
The paper proposes monitoring, defibrillator evaluation, and randomized trials for high-risk patients, but it has not tested the model in a prospective clinical workflow. Translation therefore depends on whether prospective action changes outcomes safely.
↳ Clinical and scientific implications
The adjudicated scores are driven by a strong predictive core: a pre-specified Swedish lockbox, large regional ECG data, and external zero-shot validation in US and Taiwanese cohorts. The contribution score reflects both the non-overlap with LVEF in Figure 3 and the visible aVL biomarker described in Figures 4–5. Methodological Rigour is lower than the most positive reading because the defibrillator analysis is observational, the generative biomarker work is not repeated in the lockbox, and core data/model access is restricted. The impact score is high because the paper targets a clear clinical decision pathway, but translation remains short of deployment because randomized and prospective evidence is still needed.
Nabu’s assessment, alongside the field’s view.
Are you an author of this paper?
Strong4.2
Confidence highThe paper advances sudden cardiac death prediction by combining population-scale ECG learning, comparison with LVEF, and discovery of a visible aVL waveform feature. Its contribution is substantial, though the defibrillator-benefit and mechanistic claims remain provisional rather than practice-changing on their own.
86.1% of the model’s high-risk patients were not flagged by LVEF
The design uses a pre-specified Swedish lockbox, patient-level splitting, external zero-shot validation, multiple outcome definitions, and specificity checks. The score is limited by incomplete main-text detail on model selection, non-lockbox execution of the generative biomarker work, observational OLS inference for defibrillator benefit, and restricted reproducibility of core data/model components.
The lockbox remained untouched from model development through peer review
The manuscript presents a complex chain from clinical problem to prediction, external validation, LVEF comparison, biomarker visualization, and mechanistic hypothesis in a mostly legible order. Reporting is reduced by dense reliance on supplementary methods and by practice-facing phrasing that can compress observational and exploratory evidence into stronger implications.
Without randomization, we cannot know whether defibrillators would lower mortality
The paper situates its work against LVEF limitations, prior ECG and AI approaches, endpoint uncertainty in sudden cardiac death, and the historical need for randomized confirmation. Positioning is somewhat less complete on implementation issues such as deployment, alerting workflows, and equity across untested populations.
many promising predictors in the past have failed
Lower confidence on Methodological Rigour — domain match limited.
Caveats4 of 4 checks
The main predictive results are internally coherent and transparently tied to the Swedish lockbox and external validation cohorts. Reliability watch-outs concern stronger practice-facing language around observational defibrillator benefit and the integration of exploratory, non-lockbox biomarker analyses with locked-box predictive claims.
Ethics oversight, data provenance, and data/code access routes are stated, including specific Lund approval identifiers and named data-access pathways. Independent reproduction is constrained because the central Swedish and US data and parts of the Swedish-derived deep-learning code/model are not publicly accessible.
Flags: 4 declared / 5 total
All 58 cited references resolved to real publications via PubMed, Crossref, or OpenAlex.
References not verified automatically are resolved manually within 24 hours.
No retraction or expression of concern found in Crossref (includes Retraction Watch).
Sources: Crossref (incl. Retraction Watch) ✓
Where this paper’s evidence sits on the path from initial observation to real-world use.
The findings are beyond proof-of-concept because the model is tested in a Swedish lockbox and two external cohorts without fine-tuning. Readiness remains incomplete because no prospective clinical deployment or randomized defibrillator-selection trial has tested whether acting on the model improves outcomes.
With no fine-tuning or other modification
AI-generated, human-governed. Something look off? Contact us to request a review.
Keep reading
Other papers Nabu has evaluated blind, starting with the same field.
Maria Mironova, Bachir A. Taouli, Richard L. Ehman, et al. · 2026 · Clinical medicine
MRI proton density fat fraction (MRI-PDFF) is the most accurate noninvasive measure of steatosis (thresholds of 6.4%, 17.4% and 22.1% for grades ≥1, ≥2 and 3), and a ≥30% relative reduction is associated with histologic response.
SupportedRiedl, Farkas, Aygören-Pürsün, et al. · 2024 · Clinical medicine
Oral sebetralstat provided faster times to the beginning of symptom relief, reduction in attack severity, and complete attack resolution than placebo.
Well-supportedRiedl, Tachdjian, Lumry, et al. · 2024 · Clinical medicine
Donidalorsen at a dose of 80 mg administered subcutaneously every 4 or 8 weeks significantly reduced the hereditary angioedema attack rate
Supported