Machine learning identifies new electrical signatures for sudden cardiac risk

Sudden cardiac death claims approximately 300,000 lives annually in the United States, often affecting individuals whose routine diagnostic tests show no signs of dysfunction. While implantable defibrillators can prevent lethal arrhythmias, current screening methods—primarily ultrasound measures of pumping capacity—frequently fail to identify candidates who need them most. A new study published in Nature suggests that a more precise marker has been hidden in plain sight within the standard electrocardiogram for decades.

A research team led by Ziad Obermeyer at the University of California, Berkeley, utilized 440,000 ECG recordings from Swedish patient records to train a 64-layer residual neural network. Their objective was to find patterns in the heart’s electrical activity that correlate with sudden death. The model successfully identified a high-risk group that traditional ultrasound tests missed. In this cohort, 86 percent of the individuals flagged by the algorithm were not considered at risk under current medical guidelines. The findings held constant across separate datasets from the U.S. and Taiwan, indicating the signal is a physiological reality rather than a demographic quirk.

To translate these findings into clinical practice, the team employed a second, generative AI model to visualize what the first network was seeing. The machine highlighted a subtle slurring in the aVL lead, a feature previously undescribed in a century of cardiology. This electrical fragmentation appears to correspond with diffuse fibrosis, or microscopic scarring in the heart muscle, which can be seen on specialized MRI scans but is too costly for routine population screening. By extracting this signature from a 10-second ECG, the technology offers a potential path toward utilizing mass-market sensors, such as those in smartphones or smartwatches, to triage at-risk patients for further clinical intervention.