Feb. 01, 2019
Within Cardiovascular Medicine at Mayo Clinic, interdisciplinary teams apply artificial intelligence (AI) to some of the most challenging clinical problems. Examples are:
- Early risk prediction of conditions such as embolic stroke
- Heart monitoring and arrhythmia detection in smart clothing projects based on a textile computing platform
- Occult disease detection, such as identifying atrial fibrillation's earliest subclinical stages, through heart physiology signals transmitted by mobile electrocardiogram (ECG)
Explains Paul A. Friedman, M.D., chair of Cardiovascular Medicine at Mayo Clinic in Rochester, Minnesota: "With AI, we're developing new prediction and screening tools that improve clinical decision-making and identify diseases much earlier, when they are silent and without symptoms — and are most treatable. Our goal is to treat disease before it develops and prevent it from manifesting."
Asymptomatic left ventricle dysfunction
Asymptomatic left ventricle dysfunction (ALVD) is present in 3 to 6 percent of the general population, affecting more than 7 million Americans. Risk increases with age. ALVD is present in 9 percent of older adults. Progressive in nature, ALVD's weak pumping pathology and low ejection fraction (EF) are significantly associated with the risk of premature disability and death.
Effective therapies exist and, if administered early, can reduce hospitalizations and mortality significantly. Yet no ALVD screening tool exists that is inexpensive, widely available and noninvasive to facilitate broad and early intervention by general clinicians. Currently the gold standard screen for ALVD is echocardiogram (echo), an advanced, expensive technique not typically found in nonspecialty clinics or doctors' offices.
To improve care, the Mayo AI cardiovascular team envisioned screening for ALVD using the common and inexpensive ECG enriched by AI techniques. Team members hypothesized that the metabolic and structural derangements associated with the cardiomyopathic process would result in ECG changes they could reliably detect with a properly trained neural network.
Convolutional neural networks
A subspecialty of computer science, AI has grown tremendously in the last decade with growth in computing power needed to analyze enormous data sets. These advances speed the development of algorithms to classify data, identify patterns and make predictions. The Mayo Clinic AI cardiovascular team deployed an AI form of machine learning known as a convolutional neural network (CNN).
Itzhak Zachi Attia, M.S., team AI specialist, explains the CNN's importance this way: "Artificial intelligence has the power to look at things from a unique nonlinear perspective. It can make use of millions of data points and find patterns that are invisible to the human eye. These machines are free from most artificial boundaries humans have due to language and thinking in a linear matter. This, with the ability to learn in hours or days from more data than any human can learn from in a lifetime, makes it the ultimate tool for medicine. By harnessing these models to help patients, we can finally see the invisible and digitally reach out to patients that are beyond any physical reach."
Multilayered CNN merges information
Based on the brain's perceptual architecture, convolutional neural networks (CNNs) are used to analyze, recognize and classify visual imagery through algorithms trained on hundreds of thousands of images.
CNNs are a class of multilayered networks that merge (convolute) information. Based on the brain's perceptual architecture, CNNs are used to analyze, recognize and classify visual imagery through algorithms trained on hundreds of thousands of images. ECGs and echo images from Mayo Clinic's vast digital database were used in this ALVD project. From training inputs, CNNs generalize attributes to make predictions about images the network has never actually seen before.
Using Mayo Clinic stored digital data, researchers screened 625,326 paired ECGs and transthoracic echocardiograms to identify the population to be studied for analysis. They then trained and validated their CNN with patient data from 44,959 ECG-echocardiogram pairs. The network goal is to identify ejection fractions ≤ 35 percent to identify patients who have low EFs and refer them for more-detailed follow-up exams.
The Mayo Clinic AI team tested the network on 52,870 patients reserved for external validation. Of those, approximately 8 percent had an EF < 35 percent. The sensitivity, specificity and accuracy were 85 percent, 86 percent and 86 percent, respectively. "The network ranked 0.93 in its ability to flag weak heart pumps — out of a perfect score of 1.0. To put this in perspective, a mammogram is 0.85. These results show the network is very robust," Dr. Friedman says.
Additionally, the network detected an estimated fivefold increased risk of developing a future low EF. This suggests that the network may reveal silent subclinical, metabolic or structural abnormalities hidden in the ECG. A report on this project appears in the Jan. 7, 2019, issue of Nature Medicine.
To move their network-based screening tool into broad circulation where it could function as an effective low-cost ALVD alternative to echo exams, the Mayo Clinic AI team developed a smart stethoscope supported by a smartphone app.
The smart stethoscope was developed in partnership with Eko, a California device-maker that already has received Food and Drug Administration (FDA) clearance for a smart stethoscope designed for other purposes. Mayo's ALVD project qualifies as a new use for the stethoscope and will require its own FDA approval.
Says Dr. Friedman: "With our network on a stethoscope, we're basically putting an expert cardiologist in the pocket of health care providers around the world. This greatly expands the reach of improving heart health by offering an easy-to-use ALVD screen that identifies those who need follow-up."
For more information
Attia ZI, et al. Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nature Medicine. 2019;25:70.