When Apple first launched a feature on its smartwatch to identify irregular heart rhythms a couple of years ago, it was a game changer for wearable health technology.
The app, which provides information akin to a single lead electrocardiogram (ECG), may not have quite the same diagnostic value as the standard 12-lead tests used in clinics. Nonetheless, it has been deemed useful enough to be approved by health regulators in many countries.
Its emergence was a turning point for showing how artificial intelligence (AI) could be applied to the medical field of cardiology, using computers to recognize patterns in data and detect potential heart issues.
AI can pick up on subtle clues from a person’s physiological state such as their heart rate, the time differences between each heartbeat or the electrical signals their heart produces in order to identify irregularities that point to medical conditions.
“Being able to detect atrial fibrillation just by wearing a wristwatch all the time, that kind of relatively simple technology could actually have a massive impact,” explained consultant cardiologist Tim Fairbairn, cardiovascular imaging lead at Liverpool Heart and Chest Hospital in the UK.
AI is increasingly being explored by doctors to detect heart disease and, while it currently remains more of a research field than an everyday clinical application, Dr Fairbairn anticipates interest will grow as other Silicon Valley companies such as Google Health take an interest.
“I think as these big private healthcare companies become more and more interested and invest more that this journey that we are at the beginning of will start to accelerate quite quickly,” he told Inside Precision In Medicine.
His department currently sends data to HeartFlow, a company with an AI-powered algorithm system that can create a bespoke 3D model of a patient’s coronary arteries and show how blockages affect blood flow.
Dr. Fairbairn hopes that AI will eventually be able to power precision medicine in decision trees, with vast amounts of data individualized using biomarkers, imaging tests and even genetics to create an individual risk assessment for each patient.
“Some patients may respond better to medical therapy than other patients and being able to identify those individuals on a more personal basis I think is a great possibility of this kind of technology,” he told Inside Precision In Medicine.
Over in the US, the cardiovascular department at the Mayo Clinic in Rochester, Minnesota, has a built-in system that automatically creates an AI dashboard. When its chair of cardiovascular medicine Paul Friedman sees a patient, he can click a button and see every ECG they have ever had and the AI score for five or six conditions.
Historically, he said, when a computer was used to read an ECG it was very deterministic – the doctor would say, look for a T-wave and if it does this then it means this.
“Here, we turn it upside down and we simply train a computer through a repetitive process, through a convolutional neural network, where we say – starting off with common things – this is an ECG of someone having a heart attack, and then this is someone not having it, so you feed it a million times in [and] it learns the pattern directly. So, we don’t know what components of the ECG it’s looking at. It’s just adjusting a math model in an iterative process.”
The Mayo Clinic began asking questions that couldn’t normally be answered from an ECG but that must affect it because the heart muscles, and the electrical signals they generate, are affected.
They found that, after feeding it with ECG data and identifying different heart pumping abilities, the AI system was able to identify a weak heart pump – otherwise known as a low ejection fraction.
It was even able to identify whether the patient was a man or a woman, said Friedman.
“So, a computer [that] can read an ECG knows someone’s sex better than most of us can looking at someone walking down the street. I mean, it’s just a powerful tool.”
A key benefit is its ability to diagnose using the ECG, a standard part of most medical equipment that has been used for more than a century.
This makes it more widely accessible than other means of identifying a weak heart pump such as an echocardiogram or ultrasound and the Mayo Clinic is in talks for its use in India and Africa.
Intriguingly, the system can also predict conditions – with the Mayo Clinic finding cases where the ECG has identified amyloid heart disease a year before diagnosis was made.
It can also identify episodic atrial fibrillation, which can cause strokes, even if the ECG rhythm does not look unusual at the time.
“AI can tell, sure it looks normal now but an hour ago or a day ago it was abnormal, because there are subtle traces left behind,” explained Friedman.
“It’s like looking at a beach on a quiet day and saying was there a storm here yesterday. Now, how would you tell? They look for kelp and that kind of stuff.”
While the Mayo Clinic’s system is not yet available elsewhere, it has created the AI-driven health technology company Anumana in partnership with nference to get it through regulatory approval, with the hope that it may become more widely available this year.
Friedman added that the Mayo clinic is conducting analysis and trials in conjunction with Norway and Russia, with discussions in South Africa, Nigeria and partnerships with Korea, and South America.
“We want to make sure these tests work across the diversity of humanity and if you make a test and only test it in one ethnic group or one sex or one cohort you can’t be confident it will be robust and reliable in different groups. Because there are actual ECG differences that are genetically encoded.”
Benjamin Glicksberg, an assistant professor of AI in Human Health and a member of the Hasso Plattner Institute for Digital Health at the Icahn School of Medicine at Mount Sinai maintains there is a bright future for AI in cardiology but adequate patient representation is key.
“Perhaps one of the biggest issues facing AI in healthcare is the huge potential for bias,” said Glicksberg, who directs a research team that combines health data science with AI to progress precision medicine.
Much of the data fed into these algorithms are not equally representative for all patient demographics, namely race and ethnicity, as they do not account for social determinants of health like socioeconomic status and access to healthcare, he maintains.
“Because these variables are often not taken into consideration for modeling, the algorithms can learn patterns that will not work for patients of all backgrounds.”
Dr. Fairbairn added that, while AI systems have the ability to free up clinician time and analyze nuanced data, there were limitations and “the computer is not always right”.
“They have to be equally accurate and precise but without those two things of course you’d be getting the same thing again and again and again very accurately but unfortunately it’s wrong,” he explained. “So that’s a major drawback. ‘
The ability of healthcare systems to afford this kind of technology also comes into play, as does patient confidentiality, which Dr. Fairbairn said patients particularly worry about.
AI requires big data to improve these models but this means getting a lot of patient data on a population level.
“How do we get around those stumbling blocks of being able to use large patient data in a confidential manner, without breaking patient confidentiality?” he asked. “Certainly, when you talk to patients, that’s a major concern.”
Anita Chakraverty is a UK-based journalist who has been writing about medicine and health across several international publications for more than 20 years. In her spare time, she enjoys reading, films and walks in the countryside.