Bringing AI-Assisted Cardiology into the Future

Machine learning-based algorithms can now be used to help detect atrial fibrillation, asymptomatic left ventricular systolic dysfunction, and cardiac amyloidosis. Even more digital diagnostic tools are on the drawing board.

By John Halamka, M.D., President, Mayo Clinic Platform, and Paul Cerrato, senior research analyst and communications specialist, Mayo Clinic Platform.

Ever since William Harvey discovered the essential role of the heart in systemic circulation, clinicians and medical scientists have been seeking new ways to improve the detection and treatment of life-threatening disorders like myocardial infarction, heart failure, and related conditions. Like most diseases, these disorders do not appear out of nowhere. Typically, they are preceded by numerous warning signs and symptoms, many of which are too subtle for clinicians to detect without assistance. Over the years, stethoscopes, ECGs, and numerous other devices have improved early detection, but each has its limitations. Machine learning is taking early diagnosis to a new level of sophistication, ushering in what some thought leaders are calling a “golden age” in medicine.

In a video posted on Anumana’s website, Paul Friedman, M.D., Professor of Medicine and Chair of the  Department of Cardiovascular Medicine at Mayo Clinic, explains that the body is constantly providing physiological clues about cardiovascular status, including heart rate, heart rate variability, electrical signals, and changes in respiration, all of which can be monitored and integrated into an AI-driven algorithm to improve early detection. Mayo Clinic created Anumana to offer its cardiology algorithms to healthcare providers around the world. By tapping our 11-million patient data set, these algorithms have been able to detect subtle changes in cardiovascular status, which in turn can help detect atrial fibrillation, cardiac amyloidosis, and asymptomatic left ventricular systolic dysfunction (ALVSD). The researchers and physicians at Mayo Clinic and Anumana are also developing digital tools for the early detection of pulmonary hypertension, hypertrophic cardiomyopathy, myocarditis, aortic stenosis, and hyperkalemia.

In previous blogs, we discussed the EAGLE trial, a clinical trial that demonstrated the value of an AI/ECG tool to improve the detection of ALVSD. Asymptomatic ventricular dysfunction may not sound serious to non-physicians, but the disorder, which affects about 3% of the population, indicates the heart’s pumping ability has been compromised. It’s been linked to reduced quality of life and increased mortality.

In the EAGLE study, Zachi Attia and associates at Mayo Clinic used a 12-lead ECG in combination with a convolutional neural network to identify patients with the disorder, which they defined as an ejection fraction (EF) at or below 35%.  Attia et al found: “When tested on an independent set of 52,870 patients, the network model yielded values for the area under the curve, sensitivity, specificity, and accuracy of 0.93, 86.3%, 85.7%, and 85.7%, respectively. In patients without ventricular dysfunction, those with a positive AI screen were at 4 times the risk…of developing future ventricular dysfunction compared with those with a negative screen.”

To understand how these statistical findings can be used by clinicians at the bedside, it’s important to delve a little deeper into the meaning of terms like sensitivity, and compare it to another important metric: positive predictive value (PPV). A sensitivity rating of 86.3% indicates that about 8.6 out of 10 patients who are told they are at risk of ventricular dysfunction actually are at risk; in other words, they are true positives. While that finding can give providers confidence in the value of the AI/ECG tool, it doesn’t address the more pressing questions for clinicians working with individual patients.

To answer that question, PPV is more useful because it compares patients who are true positives against the total number of patients who test positive; in other words, it’s the probability that a patient with a positive test result actually has the condition of interest. Put another way, it’s the number of true positives, divided by true positives plus false positives. And that metric varies based on the incidence of the disorder and the cutoff point used for the test. 

As Attia et al explain: “The accuracy, sensitivity, and specificity of the networks were all excellent, but the test’s positive predictive value was only 33.8%. This is in part because we selected an EF cutoff of 35%. We selected this detection threshold owing to the well-established outcome and therapeutic implications of this value. However, identification of an EF of < 50% is still clinically significant, as this reflects an abnormal EF. When considering a higher cut-off for abnormal function of < 50%, the positive predictive value of the network was 63.4%.”

William Harvey could never have imagined what his discovery would eventually results in centuries later. And we in turn can only guess at the new diagnostic tools that will eventually become available a few short years down the road. It’s no exaggeration to say that we are in the middle of a digital health revolution that is profoundly impacting patients worldwide.