Heart disease researchers say yes
Artificial intelligence, or AI, is all the rage. And it’s not just because of ChatGPT, self-driving cars, or even the smartphone apps that allow doctors to track a patient’s blood pressure. Researchers say AI, which uses computers to perform tasks that would normally require human intelligence, has the extraordinary potential to save lives – lots of them. And in the field of cardiovascular health, NHLBI-funded scientists are developing the tools to do just that. Their focus: using AI to help doctors more accurately predict risks for heart disease – the number one global killer – and diagnose serious cardiovascular problems.
“Recent advancements in AI have the potential to revolutionize how cardiovascular care is delivered,” said Girish Nadkarni, M.D., NHLBI-funded researcher and professor of medicine at the Icahn School of Medicine at Mount Sinai. “AI will detect disease earlier, improve access to care, and help develop personalized therapeutic plans that might prevent, control, or even cure cardiovascular disease for the millions of people who battle heart disease each year.”
It might seem other-worldly, scientists say, but it’s real.
So how did we get here – and seemingly so fast?
Patricia Bandettini, M.D., a senior research physician in NHLBI’s Division of Intramural Research and chief of the Cardiovascular Magnetic Resonance (CMR) Program, said it’s fair to start with two words: computational technology.
“Advances in computing power have made it possible to analyze huge amounts of data efficiently and reliably,” she explained. As a cardiologist, she’s trained to look at data from a patient’s electrocardiogram (ECG) – a readout of electrical output of the heart. “In a clearly abnormal ECG, I can reliably tell if a patient is having an acute heart problem,” she said. “But in another instance, from what might seem like a ‘normal’ ECG, I can’t always predict if that patient may have a risk for something more such as a heart attack.”
With AI, cardiologists will be able to use their years of training in combination with computational methods to discover nuances in the data – little signposts that might lead to more accurate diagnoses or predictions of future health outcomes. This data could include images such echocardiograms, an ultrasound that checks the structure and function of the heart; 3D heart x-rays known as CT scans; CMRs; ECGs; blood markers; patient symptoms; and even demographic information in a patient’s electronic medical record (EMR), such as age, weight, or ethnicity.
How it works
AI uses machine learning, which enables the computer to interpret and learn from the data. The key to machine learning is to find the pattern behind the data and build a model based on that information, which then allows the model to predict outcomes from future data. In practice, data scientists train the computer by analyzing extensive datasets, such as hundreds of thousands of echocardiograms or CT scans, to detect complex patterns that are not readily apparent to the human eye. Some patterns will mean the patient is on track to have good outcomes, while other patterns suggest a heart attack is looming. The result is that the computer can then read an image from any patient, compare it against known patterns, and detect a possible heart condition or predict future problems.
“There are thousands of patterns in heart data and multiple possible outcomes,” said Bandettini. “An AI algorithm can more quickly and reliably see what features make a patient higher risk for poor outcomes than I or even a team of well-trained clinicians ever could. That’s the power of AI.”
Testing AI’s potential
NHLBI-funded researchers have been developing AI tools for a slew of cardiovascular complications using a variety of data types. One recently published study, co-funded by the National Center for Advancing Translational Sciences, focused on identifying patients at high-risk for a poorly functioning right ventricle. This chamber of the heart serves an important role by sending blood to the lungs. The researchers trained the computer using ECG data, paired with CMR, to effectively predict right-side heart issues such inferior pumping or abnormal size.
“Our method might one day be employed in the clinic as a way to screen patients who are at risk of right heart failure – but without the need for expensive or complex testing,” said Son Duong, M.D., lead author of the study and assistant professor of pediatrics at Mount Sinai’s Icahn School of Medicine.
In another published study funded by NHLBI, researchers at Stanford University used body composition imaging biomarkers from CT scans of the abdomen region to train their AI model. Using their tool, along with features from the patients’ EMR, they were able to enhance the current risk assessment for ischemic heart disease to identify which patients are most at risk.
Finally, using data from echocardiograms, another team of NHLBI-funded researchers developed a model to interpret measurements from ultrasound images, such as size of the left atrium, thickness of the left ventricular wall, and how well the heart was pumping. The model was successfully able to predict heart failure, atrial fibrillation, heart attack, and even death.
“The computer models are pretty accurate, and the predicted heart measures by the model are associated with future heart disease,” said Jennifer Ho, M.D., senior author on the study and associate professor of medicine at Harvard Medical School.
Yet for all their promise in helping predict, diagnose, and treat disease, these kinds of early-stage discoveries still have to be put to the test using the highest scientific standards, said Ho. “We have to be cautious and assure robust validation and rigorous implementation,” she stressed.
When AI does become the norm, said Bandettini, cardiologists and other heart disease professionals should not fear for their jobs. “You will always need to have humans to do quality assurance and explore and manage the findings,” she said. “That part will never go away.”
Important considerations
While the use of AI is being developed at breakneck speed, health professionals and scientists have begun mulling over the many ethical and practical issues that need to be addressed before it becomes a reality in the clinic. For example, training the AI models requires very large, but representative and unbiased datasets.
“If the data the model was trained on outcomes from all older white men, then maybe it's not applicable to women or people of different races and ethnicities,” said Bandettini. “Or what if it’s only trained on patients with heart failure, but not congenital heart disease – will it still be relevant to those patients?”
Close communication between scientists who develop the models and analyze the data and health professionals who make sense of it will be essential, too.
“We’re on the precipice of something big. The potential is there,” said Bandettini. “Now we’re just waiting for the field to get to the comfort level we need to use it in everyday clinical practice.”
