A new artificial intelligence model trained on more than 440,000 electrocardiograms has identified patterns in routine heart tests that predict sudden cardiac death with striking accuracy—patterns that current clinical screening methods consistently miss.
Researchers at UC Berkeley, publishing their findings in Nature, developed an algorithm capable of flagging high-risk patients even when their hearts appear normal by conventional measures. The discovery points to a significant gap in how the medical field currently assesses cardiac risk.
Sudden cardiac death claims hundreds of thousands of American lives annually, often striking without warning in people with no known heart disease. Once cardiac arrest occurs outside a hospital setting, survival rates plummet. Early identification of at-risk patients could fundamentally change outcomes.
The traditional approach relies on a metric called left ventricular ejection fraction, or LVEF, which indicates how efficiently the heart pumps blood with each beat. When LVEF falls below a certain threshold, patients typically qualify for an implantable cardioverter-defibrillator—a device that can electrically correct dangerous heart rhythms. However, this method leaves substantial blind spots. Many people who experience sudden cardiac death never underwent the deeper cardiac evaluation that would have identified their reduced LVEF.
The UC Berkeley model revealed a striking disparity in risk stratification. Patients flagged as high-risk by the AI showed a 7 percent annual rate of sudden cardiac death, compared with 4.6 percent annually for the standard LVEF-based high-risk group. Moreover, the vast majority of patients identified by the algorithm would have been missed entirely by conventional screening.
The researchers worked to understand what the algorithm was actually detecting. Using a secondary AI system, the team compared ECG patterns between low-risk and high-risk populations. The analysis pointed to a specific feature within the aVL lead—one of the standard electrical views cardiologists examine—located within the QRS complex, the segment reflecting the heart's primary electrical impulse. This pattern emerged as a strong predictor of sudden cardiac death and had not been previously documented in medical literature.
Validation across multiple health systems strengthens the finding considerably. The team tested their model on patient data from the United States and Taiwan, separate from the Swedish dataset used for training. The algorithm's consistency across these distinct healthcare environments suggests genuine robustness.
Implantable defibrillators save lives, but they carry real costs and risks. The procedure is invasive and expensive. Many devices implanted under current clinical guidelines never fire—meaning patients undergo surgery for an intervention they never needed.
The AI tool could help physicians calibrate that balance. Rather than immediately moving high-risk patients toward implantation, physicians might intensify monitoring. Patients flagged by the algorithm could wear ambulatory heart-monitoring patches to detect dangerous rhythms before they become fatal.
The research team is advancing toward clinical implementation. Partnerships with health systems in Sweden, Taiwan, and the United States are underway to test the algorithm against hospital ECG databases in real-world conditions. The next phase will likely involve contacting patients identified as high-risk and offering them extended monitoring before considering more invasive treatment.
The broader adoption of medical AI raises governance questions. Developing this algorithm required approximately a decade of data collection and curation—a reminder of the substantial resources required for serious clinical AI development. That concentration of data creates power imbalances. Questions about data ownership, control and privacy become urgent as patient health records contribute to training algorithms.
For now, patients cannot independently assess their cardiac risk using this algorithm. The tool exists in research and early clinical testing phases, not in consumer applications.
A routine ECG—a test millions of Americans have already undergone—may contain prognostic information that artificial intelligence can extract where human pattern recognition cannot. Realizing that benefit depends not only on algorithmic accuracy but also on building public trust through transparent, ethical data practices.
