AI and sudden cardiac death: what should clinicians do with an ECG risk prediction?

Christopher Ford

October 25, 2026

A short ECG can contain information about a dangerous event that has not yet happened. In research published in Nature on June 24, 2026, a deep-learning model identified a small group with a high subsequent rate of sudden cardiac death, including many people who were not already flagged by reduced cardiac pumping function. [1]

The investigators went further than predicting risk. By pairing predictive and generative models, they made part of the signal recognizable as a waveform feature that could be examined and investigated biologically. That combination makes the study important both as a possible route to earlier risk recognition and as a way to ask new questions about the heart. [1,2]

What the investigators discovered

Ziad Obermeyer and colleagues trained their model using clinical ECGs from Region Halland in Sweden, linked to health records and death certificates. They reserved patients and all their ECGs in an untouched evaluation dataset. The principal analysis examined 113,072 ECGs from 35,417 people younger than 80 who did not have defibrillators. [1]

At the investigators’ selected threshold, 2.2% of ECG observations fell into the high-risk group. That group had an observed one-year sudden cardiac death rate of 7.0%, with a 95% confidence interval of 4.9–9.5%. Only 13.9% of the high-risk group were known to have reduced left ventricular ejection fraction, the measure of pumping function used in the study’s comparison. The remainder included people with normal measurements and people whose ejection fraction had not been measured. [1]

A prediction alone did not explain what the model was seeing. The researchers therefore paired a model that assessed risk with a generative model that could reconstruct and alter individual ECG beats. Starting with a real, lower-risk beat, they repeatedly adjusted its generated representation in the direction that the predictive model associated with higher risk. The resulting sequence showed how a waveform could change as predicted risk rose, rather than simply comparing recordings from different people. These were synthetic variations, not recordings of a patient’s heart deteriorating over time. [1,2, sections V.A–B]

Alongside familiar changes in the direction of the heart’s electrical activity, the process revealed a distinctive slurring near the end of the R wave in lead aVL—one view of the heart’s electrical activity. The researchers quantified the shape by measuring how voltage changed from the R-wave peak to the end of the QRS complex. Those measures added predictive information in the Swedish and US data after accounting for standard ECG measurements. The visible feature represented part of the model’s signal, rather than a substitute for the full model. [1,2, sections V.G–H]

That recognizable feature gave the investigators a biological question to pursue. Could small, widely distributed obstacles to electrical conduction cause the advancing electrical wave to scatter and change direction, producing the observed shape? Myocardial fibrosis—the deposition of collagen between heart muscle cells—was one possible explanation because it can interfere with conduction. This was the authors’ hypothesis, developed from the waveform and a simplified conduction thought experiment. [1,2, section VI.A]

They then examined cardiac MRI scans already available in the health records. A cardiologist, blinded to model predictions and patient outcomes, reviewed scans from nine patients in the highest-risk tenth and fifteen lower-risk patients. The higher-risk group more often showed subtle, diffuse late gadolinium enhancement in the left ventricle, a pattern of contrast uptake consistent with fibrosis. This small, selected comparison offered a clue about the biomarker’s possible biological significance; it did not establish that fibrosis caused the waveform or the subsequent deaths. The exploratory waveform and MRI work used the original training and validation datasets, rather than the untouched dataset used for the final predictive evaluation. [1, Methods, “Generative model and morphing procedure”; 2, section VI.B]

The advance is therefore both predictive and investigative: a model identified otherwise unrecognized risk, and its signal led to a visible feature and a testable biological hypothesis. For cardiovascular software developers, the next task is to decide which clinical use of that information to pursue.

Define the clinical action before the product claim

A company developing such software could envisage several roles: identifying patients for further assessment, prioritizing specialist review, or supporting a decision about a preventive intervention.

They require different evidence because the result has a different influence in each pathway. A flag that prompts a fuller assessment leaves additional clinical work between the prediction and treatment. A result intended to support implantation of a defibrillator has a more direct relationship to an invasive intervention.

Zaliance’s development recommendation is to define that proposed action while planning the evidence program. “Predicts sudden cardiac death” describes an outcome of interest. It does not yet explain who will use the prediction, what they will do differently, or why that change should help the patient.

The defibrillator finding supports a trial, not a treatment rule

The paper includes an important observational analysis. Among patients classified as high risk, those who already had defibrillators experienced less sudden cardiac death than the regression model would otherwise have predicted. The authors report a difference of 3.62 percentage points against an expected rate of 6.65%, expressed as a 54.4% relative reduction. [1, Table 2]

These patients had been selected for defibrillators under existing clinical practice, not assigned to treatment on the basis of the ECG score. Differences in who received a device, and other unmeasured factors, could affect the comparison. The analysis therefore does not establish that treatment selected using the model would improve outcomes in the broader population it identifies. The authors call for a randomized trial in high-risk patients. [1,2]

For a developer, this separates two evidence questions: whether the software identifies meaningful risk, and whether acting on its result produces a beneficial change in care. The study provides a strong reason to investigate the second question, with the intended clinical pathway determining what needs to be evaluated.

External validation tested different parts of the promise

The international results strengthen the case that the signal is worth pursuing, but they are not three repetitions of the same clinical claim. The researchers applied the model without fine-tuning to US and Taiwanese data. Each dataset answered a different question. [1,2]

In Sweden, the main outcome was sudden cardiac death within a year, defined using death-certificate causes and the circumstances of death. Such records can include deaths that were not caused by a treatable arrhythmia. The investigators therefore examined additional evidence, including recorded ventricular fibrillation and ventricular tachycardia.

The US validation used ECGs from a San Diego health system and subsequent documented ventricular fibrillation or ventricular tachycardia. It did not validate one-year sudden cardiac death directly, because cause-of-death linkage was unavailable. The 29.1% rate reported in its highest-risk 2.2% group concerns those recorded arrhythmias; it is not a US replication of the Swedish 7.0% mortality rate. [1]

Taiwan supplied a hospital case–control comparison using earlier ECGs from people who later presented with cardiac arrest and controls from the same emergency department. Detailed investigation distinguished arrhythmic from non-arrhythmic arrests. That supports examination of the signal’s relationship to arrhythmic events, but the selected case–control sample cannot establish an absolute event probability for a general screening population. [1,2]

For product development, those differences define what remains to be demonstrated. A model may retain useful ability to rank risk in a new setting without preserving the same meaning of a numerical probability or treatment threshold. The proposed claim needs an outcome, time horizon, population and clinical setting that the evaluation actually addresses.

The recording conditions are part of the evidence

The main model used ten-second, twelve-lead clinical ECGs, represented at 500 Hz. The Swedish data came from a Philips system; the US and Taiwanese data came from GE MUSE systems. Some US recordings were acquired at 250 Hz and interpolated to 500 Hz. This provides useful evidence across those acquisition and storage contexts. [1,2]

It does not automatically support a watch, ambulatory patch or different lead placement. The supplement does explore models trained on individual leads from the existing ECG data, with lower positive predictive values than the twelve-lead model. That is a useful research direction, rather than validation of a different recording device in its intended environment. [2, section V.H]

The population boundary deserves the same attention. The main analyses focused on people under 80 who had clinical ECGs; the supplement also examined all ages. Neither description is interchangeable with an unselected population invited for screening. A company changing the acquisition configuration or population needs an evidence plan for that proposed use, rather than an assumption that the published performance travels with the algorithm.

Make uncertainty usable in care

The selected Swedish threshold illustrates the communication challenge. It concentrated risk into a small group, but its sensitivity was 26.5%: at that threshold, most subsequent sudden cardiac death outcomes in the evaluation were not flagged. Conversely, membership of the high-risk group did not mean that an event was certain. [2, section III.B]

A proposed report therefore needs to communicate more than a red or green category. The product team would need to establish whether clinicians understand the predicted outcome and time window, the applicable population, and the difference between a risk estimate and an existing diagnosis. A result below the selected threshold should not be presented as evidence that symptoms or other clinical findings can be disregarded.

The evaluation should also examine what happens after the report is read. Does it change referral, further assessment or intervention decisions in the way the manufacturer intends? What happens when the prediction and other findings disagree? Those questions connect the model’s performance to the work clinicians and patients would actually undertake.

This study provides a compelling reason to investigate earlier recognition of arrhythmic risk. The development task is to turn that opportunity into a defined clinical role, with evidence for the action the result is meant to support. A more informative ECG becomes a more useful product when the next decision is justified as carefully as the prediction.

References

[1] Obermeyer Z, Schubert A, Ross J, Mullainathan S, Lingman M. An ECG biomarker for sudden cardiac death discovered with deep learning. Nature 655, 210–218 (2026). Originally published June 24, 2026. See the predictive-performance, external-validation, defibrillator, biomarker and Methods sections. Full paper.

[2] Supplementary Information accompanying the paper. Sections III.A–C (performance and dataset differences), IV.A (defibrillator analysis), V (biomarker exploration and individual-lead models), VI (mechanistic hypothesis and cardiac MRI investigation), VII (dataset construction), VIII (all-age analysis) and IX (model input and training). Supplementary Information.