How Artificial Intelligence Discovered A New Way To Detect Patients At Risk Of Cardiac Death

I was excited to learn about a new use of artificial intelligence to make a potentially significant breakthrough in detecting patients at risk of sudden cardiac death.

Sudden cardiac death (SCD) is a serious problem, responsible for 10-15% of deaths globally and resulting in over 300,000 fatalities in the US every year. One current tool for helping prevent these deaths is detecting patients at high risk for SCD and surgically placing a device into their chest called an implantable cardioverter‐defibrillator (ICD). This can help “shock” the heart if it suddenly stops and keep them alive long enough to get to the emergency room.

However, it’s difficult to identify those who will benefit from the ICD devices. Like all surgeries, there is a risk of complications. Physicians would love to find a way to place ICDs into patients who truly need them, while avoiding the risks and expenses of placing them into patients who likely will never benefit from them.

At present time, the best screening test to look for appropriate candidates for ICD surgery has been the echocardiogram (an ultrasound of the beating heart). If the echocardiogram of the patient’s heart showed that a particular measurement known as the “left ventricle ejection fraction” was too low, then these patients had a roughly 4.6% chance of sudden cardiac death in the next year and might be good candidates for receiving an ICD. Even then, the echocardiogram is a very imperfect test. Not everyone who had the abnormal measurement experiences a cardiac event. Conversely, a significant number of patients who experience sudden cardiac death likely did not have that abnormal measurement.

However, recent research by Dr. Ziad Obermeyer (University of California, Berkeley School of Public Health) and colleagues offers hope for a significant improvement in detecting suitable candidates for the surgery. Dr. Obermeyer and his team used artificial intelligence to analyze 6 years of data from the Swedish health system, including 110,000 EKG readings on 35,000 patients. By correlating the EKG data with the known patient outcomes (i.e., which patients experienced sudden cardiac death and which did not), they were able to train the AI to flag patients at increased risk of a cardiac event.

Interestingly, the AI was able to identify a high-risk group of patients with a 7.0% annual risk of sudden cardiac death—significantly higher than the 4.6% risk using standard echocardiograms. Furthermore, 80% of the patients identified as high risk would not have been detected by the current echocardiograms. These patients would have been sent home without ever getting an ICD.

The researchers also confirmed that these results held up when applied to independent patient data sets from the US (San Diego) and from Taiwan.

One problem with medical AI diagnoses is that the systems often cannot explain their “reasoning.” Instead, the AI system operates much like a “black box” generating answers in a way that humans cannot replicate or understand. To address this, Obermeyer and colleagues did something clever. They used a second AI system to generate arbitrary EKGs to be analyzed by the first system, trying to pin down the feature(s) that yielded a “high risk” score. The second AI was systematically “rewarded” for modifying the synthetic EKGs to produce increasingly higher risk scores. After many iterations, the researchers determined that a key feature was a specific alteration in the shape of the electrical signal of one of the EKG leads called “aVL” (augmented vector left).

Although the exact reason this altered EKG pattern is correlated with cardiac death is still not fully understood, one plausible hypothesis is that scarring or tissue damage in a certain part of the heart will disrupt the flow of the internal electrical signals in a way that disrupts the normal heart rhythms—and also shows up in the EKG in this particular way.

This is the first time that an AI system has detected a new abnormality in EKGs that appears correlated with cardiac risk and is detectable visually by a human physician. Physicians and scientists will need to do more work to prove the precise causal linkage between the EKG finding and the risk of sudden cardiac death, but this is feasible with current tools such cardiac MRI and selective cardiac muscle biopsies.

Dr. Obermeyer’s research is not yet ready for routine clinical application. He and his colleagues are working to validate and refine their results in a fashion that ordinary practicing cardiologists and family physicians can use in day-to-day practice. But once the science is established, one advantage of screening patients with EKGs is that the test is relatively simple and inexpensive, as opposed to MRI scans or invasive coronary angiograms.

I am encouraged at this innovative use of artificial intelligence to advance both medical treatments as well as to improve our understanding of basic underlying mechanisms of heart disease. I’m looking forward to learning more from Dr. Obermeyer’s group, and for learning of similar uses of AI across all branches of medicine.

(If you are a patient interested in participating in Dr. Obermeyer’s research, you can learn more at his patient website.)