AI detects sudden cardiac death risk in ECGs that doctors cleared as normal

A 75-year-old woman in Sweden came to the emergency room with dizziness and was admitted. Her echocardiogram came back normal. Her blood vessels were clear. Doctors scheduled her for a defibrillator implant — but she died of sudden cardiac arrest before the procedure. She had undergone a routine electrocardiogram four months earlier at an outpatient visit. When researchers later ran that ECG through an AI model, it flagged her as high risk.

A study published in the 2026 issue of Nature by a research team led by Ziad Obermeyer, a professor at UC Berkeley, found that AI can identify patients at high risk of sudden cardiac death using ECG data alone.

The model outperformed echocardiography — currently the only accepted screening standard — and identified high-risk patients even among those whose echocardiograms had been read as normal.

Defibrillators can prevent cardiac death, but predicting who needs one remains elusive

Deaths from sudden cardiac arrest are theoretically preventable.

Defibrillators — devices that detect an irregular heartbeat and restore normal rhythm with an electric shock — have existed since the 1980s.

The problem is that doctors have no reliable way to know in advance who needs one implanted.

The only criterion used in clinical practice has been the ejection fraction, the share of blood the heart pumps out with each beat. Patients whose ejection fraction falls below a threshold are classified as high risk and considered for a defibrillator implant.

However, this criterion has significant gaps. Most people who die of sudden cardiac death had normal ejection fraction readings beforehand. Conversely, two out of three patients who received a defibrillator implant based on a low ejection fraction never needed the device to activate — meaning they underwent unnecessary surgery.

AI trained on 441,614 ECGs achieves greater accuracy

The research team conducted a comprehensive analysis of medical data from one region of Sweden, linking 441,614 ECGs performed there between 2010 and 2016 with death certificates and electronic medical records.

The AI was trained on this dataset, with the goal of predicting whether a patient would die of sudden cardiac death within one year based on ECG data alone.

A graph ranking patients by risk level based on a validation dataset of 113,072 ECGs. The horizontal axis shows the top percentage of patients by risk score; the vertical axis shows the actual annual rate of sudden cardiac death. The small inset box in the upper left enlarges the top 10 percent of the risk distribution.
The two horizontal dashed lines crossing the middle of the graph represent existing clinical thresholds for high-risk classification: the average risk level of the control group in defibrillator clinical trials (4.9 percent, pink dashed line) and the risk level of patients with reduced left ventricular ejection fraction (LVEF) (4.6 percent, blue dashed line).
The dark red dot indicated by the red arrow marks the optimal high-risk cutoff proposed by the research team. Applying this threshold, ECG analysis alone can accurately identify hidden high-risk patients whose sudden cardiac death risk equals or exceeds that of patients currently classified as high risk under existing criteria. [Nature, 2026] — A graph ranking patients by risk level based on a validation dataset of 113,072 ECGs. The horizontal axis shows the top percentage of patients by risk score; the vertical axis shows the actual annual rate of sudden cardiac death. The small inset box in the upper left enlarges the top 10 percent of the risk distribution. The two horizontal dashed lines crossing the middle of the graph represent existing clinical thresholds for high-risk classification: the average risk level of the control group in defibrillator clinical trials (4.9 percent, pink dashed line) and the risk level of patients with reduced left ventricular ejection fraction (LVEF) (4.6 percent, blue dashed line). The dark red dot indicated by the red arrow marks the optimal high-risk cutoff proposed by the research team. Applying this threshold, ECG analysis alone can accurately identify hidden high-risk patients whose sudden cardiac death risk equals or exceeds that of patients currently classified as high risk under existing criteria. [Nature, 2026]

The trained AI was then applied to a validation dataset of 113,072 ECGs.

The actual rate of sudden cardiac death among patients the AI flagged as high risk was about 1.5 times higher than among those classified as high risk under the existing ejection fraction standard. The AI identified high-risk individuals more accurately than echocardiography, the sole criterion used until now.

A more striking figure: 86.1 percent of the patients the AI identified as high risk had been assessed as normal under echocardiography-based criteria — patients who, under current standards, would have been considered to have no indication for a defibrillator implant.

The research team applied the same AI model, without adjustment, to 251,858 ECGs from a hospital in the United States and to data from National Taiwan University Hospital. In the United States, roughly three in 10 patients the AI classified as high risk developed arrhythmia within one year. In Taiwan, the model also proved effective at distinguishing between cardiac arrest caused by arrhythmia and cardiac arrest from other causes.

'If it had been caught earlier': a case the research team highlighted

The research team described a real patient case from Sweden in the paper.

The patient was a 75-year-old woman with a history of lymphoma. She visited the emergency room after recurring episodes of dizziness and was admitted after doctors confirmed her heart was beating irregularly. Her echocardiogram was normal. Vascular tests were clear. Further workup failed to identify a cause.

She was discharged with an outpatient appointment scheduled for a defibrillator implant, but died of sudden cardiac arrest before the procedure date arrived.

An ECG taken four months earlier at a routine outpatient visit — not because of any cardiac concern — was on file. When the research team ran that ECG through the AI, it classified her as high risk.

The research team said the case represented a typical pattern among the high-risk patients the AI identified.

However, the team said randomized clinical trials in high-risk patients identified by the model are still needed. In the past, markers that appeared highly accurate in prediction studies have failed to demonstrate a benefit from defibrillator implantation when put to the test in clinical trials.

Reference

DOI: 10.1038/s41586-026-10674-6

Obermeyer, Z., Schubert, A., Ross, J. et al. An ECG biomarker for sudden cardiac death discovered with deep learning. Nature (2026).

dbsdn1110@heraldcorp.com