🔘 About this critique
A risk score can rank electrocardiograms long before anyone can say what the model sees in a heartbeat. In this study the authors follow the score back to a particular shape in lead aVL. They also ask a harder clinical question. Among people whose ECG raises concern, who might live longer if offered a defibrillator?
This assessment examines the version of record and its supplementary analyses. We checked the reported denominators, operating characteristics and mortality comparison against the published tables. The patient records and trained models were not independently analysed.
Critique
| Examined version | Peer-reviewed article, Nature 655, 210–218. Published online on 24 June 2026. |
| Materials examined | Version of record and Supplementary Information, particularly Tables I.A.1, III.B.1 and V.G.1. |
| Recommendation | Read Figure 1 alongside the sensitivity in Supplementary Table I.A.1, then follow the proposed waveform marker through Figures 4 and 5. |
Unit 1 — What does a high-risk result predict?
Article data. The Swedish model was trained to predict an out-of-hospital cardiac death recorded on a death certificate within one year after an ECG. Patients were allocated to separate development and test sets. The principal analysis comprised 113,072 ECGs from 35,417 patients under 80 without a defibrillator. Some patients contributed several recordings. The area under the receiver operating characteristic curve was 0.872 (95% confidence interval 0.843–0.899).
Reading. At the authors’ preferred threshold, 2.2% of recordings were flagged. The event rate was 7.0% in that group, against 0.58% across the test sample. Yet sensitivity was 26.5%: nearly three quarters of the ECGs followed by a qualifying death fell below the threshold. A death certificate also leaves uncertainty about the rhythm at the time of death.
Critical conclusion. The score identifies a small group at substantially elevated risk. A low score at this threshold offers limited grounds for ruling out a future event, and multiple ECGs from one patient must not be counted as separate lives.
| Claim type | Risk stratification and ascertainment of the outcome |
| Exposure | High. Confusing the event rate among flagged recordings with the fraction of events detected would overstate the reach of screening. |
Unit 2 — What travelled across the three health systems?
Article data. The US validation included 251,858 ECGs from 139,613 patients. The outcome was a recorded episode of ventricular fibrillation or ventricular tachycardia. The unchanged model achieved an area under the curve of 0.822 for that outcome, compared with 0.717 for the corresponding outcome in Sweden. In Taiwan the model distinguished adjudicated arrhythmic arrests from emergency department controls, with an area under the curve of 0.767.
Reading. The US records lacked linked causes of death. In Taiwan the investigators selected cases and controls, so the proportion of cases in the study could not establish an annual risk in the population. Even the definition of a positive screen matters. Applying the Swedish score cut-off in the US flagged 3.4% of recordings. Selecting the highest 2.2% locally produced a different group.
Critical conclusion. Evidence from three settings supports an association between the score and serious cardiac events. The calibration of an annual sudden-death probability in a new clinical population remains a separate question.
| Claim type | External validation across outcomes and sampling designs |
| Exposure | High. An estimate obtained from one outcome or sampling design could be used to counsel patients about another. |
Unit 3 — Which part of the heartbeat carries the signal?
Article data. A generative model altered the shape of individual beats in the direction that increased a predictive model’s score. In lead aVL, the downward slope at the end of the R wave became less abrupt. The authors then quantified the shape in observed ECGs using the mean absolute first and second differences in voltage from the R-wave peak to the end of the QRS complex.
Reading. Figures 4 and 5 show how a model-derived contrast led to a measurement that another investigator could examine. In Taiwan, both waveform measures were associated with arrhythmic arrest in unadjusted analyses. After adjustment for conventional ECG features, both estimates remained negative, but neither was statistically significant. For sudden cardiac death in Sweden, only the first-difference measure retained a significant association after adjustment. Both measures retained significant associations with ventricular fibrillation or tachycardia in Sweden and the US. The paired traces in Figure 5 illustrate a generated change to a beat. They do not document a patient’s heart progressing from one state to the other.
Critical conclusion. The work offers a testable waveform marker. Its relation to conduction through diseased myocardium merits direct physiological study, and the performance of the full model cannot be attributed to this one feature alone.
| Claim type | Interpretation of the waveform and its possible mechanism |
| Exposure | Moderate. A visible and measurable feature can acquire a more definite physiological interpretation than the observations support. |
Unit 4 — Would acting on the score prevent deaths?
Article data. Only 13.9% of the high-risk ECG group were known to have a left ventricular ejection fraction of 35% or less. Ejection fraction had been measured in 22.5% of the analysis sample. In a separate observational comparison, the estimated interaction between a high-risk ECG and an existing defibrillator was −3.62 percentage points for sudden cardiac death. Relative to the authors’ predicted rate of 6.65%, this gives the reported 54.4%.
Reading. An unmeasured ejection fraction cannot be classified as normal. Patients who already had a defibrillator had been selected for implantation in clinical practice. Their underlying disease, care and eligibility could differ from those of newly identified patients. The regression provides a numerical comparison, but treatment allocation leaves the counterfactual outcome uncertain.
Critical conclusion. The study makes a case for a prospective trial of ECG-guided selection. Demonstrating clinical benefit would require comparing outcomes under the proposed strategy and usual care, including all-cause mortality and device complications.
| Claim type | Added clinical information and benefit from treatment |
| Exposure | High. Treating the reported 54.4% association as the benefit of a new implantation rule could affect an invasive decision. |
Overall assessment
The paper’s most distinctive contribution is the passage from a prediction to a waveform measurement. A reader can follow the transformation in Figure 4, identify the interval being measured in Figure 5 and consult the regressions on observed ECGs. That sequence makes a specific question available for independent testing: does this part of the QRS complex retain its predictive value across instruments, populations and definitions of the outcome?
The choice of a clinical threshold needs its own evidence. The Swedish results quantify both how much risk was concentrated and how much lay outside the selected group. The external cohorts broaden the evidence for discrimination. A trial would answer the question that remains for a patient offered an implant on the strength of a routine ECG: how does that decision change the chance of surviving, and at what cost in complications?
🔘 Paper page: https://www.nature.com/articles/s41586-026-10674-6
Abstract
Sudden cardiac death is, in theory, preventable with defibrillators. But every year, many patients die without defibrillators because doctors fail to predict their risk1. The only predictive biomarker in wide use, cardiac left ventricular ejection fraction (LVEF), misses most sudden cardiac deaths2, and flags many low-risk patients for futile defibrillators that never fire3,4. 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.
Authors’ abstract reproduced from Nature under the Creative Commons Attribution 4.0 International licence. Reference numbers follow the original article.
Authors
Ziad Obermeyer, Alexander Schubert, James Ross, Sendhil Mullainathan, Markus Lingman
Liked this post? Follow this blog to get more.