Frequently characterized as either a panacea or a doomsday engine, artificial intelligence is certainly a powerful tool to improve patient outcomes, but its promise comes with some potential pitfalls
Jim Banks
Artificial intelligence (AI) is everywhere. From the algorithms that suggest what we might like to watch on streaming services, to the recommendations made by online retailers and search engines. Used in countless industries to analyze customer and market behavior in real time, AI is crunching an unimaginably large volume of data to identify patterns and inform decisions.
Although it is heavily regulated and change can be slow, healthcare is a hotbed of AI innovation. Take Novartis Foundation’s new AI4HealthyCities initiative, which seeks to analyze the key determinants of a healthy population by examining how the conditions in which we are born, grow, and work shape our cardiovascular health.
Such initiatives could not exist without AI, which can cope with many more datasets of far greater depth than any person could and can spot patterns infinitely quicker than the human brain. Unsurprisingly, AI is being directed at the world’s most pressing health issues, not least cancer, cardiovascular disease (CVD) and heart failure (HF).
It has been hard at work in oncology for many years. Sophisticated algorithms are studying the mechanisms of cancer initiation, progression, and metastasis. AI is simulating the atomic behavior of the RAS protein commonly associated with cancer, and the FDA has authorized the use of AI-enabled software to help pathologists analyze prostate biopsy images.
The use of AI is expected to continue expanding rapidly in cardiology, too, assisting with imaging analysis, supporting clinical decision-making, and improving patient outcomes in many other ways.
“As clinicians, we are all wondering about AI and its implications,” remarks Amber Johnson, MD, Assistant Professor of Medicine and Director of Cardiovascular Rehabilitation at the University of Chicago. “There is promise in many fields, and within medicine and clinical care this is not the first time we have talked about efforts to perfect the decision-making process with different decision tools that can be used to take better care of our patients. The concern is that we don’t really know the potential outcomes and implications.”
Primarily a health equity researcher and cardiologist, Johnson looks at evidence-based care and clinical decision-making to understand how they are applicable to patients. Her experience has made her keenly aware of both the benefits and the potential risks of using AI.
“It is scary in that it could be a small thing that aids us in what we are doing clinically, or it can completely take over like we have seen in so many science fiction movies,” she remarks. “But the potential for good is that AI and machine learning (ML) could help us remove some of the guesswork from what we do. Medicine is as much an art as a science.”
“With what you learn in medical school and all of the data you have to process, making a decision for the patient sitting in front of you is difficult,” she adds. “We are now focused more on patient autonomy and shared decision-making. So many different variables make each patient different, so can AI adjust for the social determinants of health, or difficulties of access to care, or understand whether the available data is applicable to this particular patient? If AI can handle those nuances it can help to shape better patient care, if not then it won’t be the tool that we want it to be.”
Healing the heart
HF is a huge public health problem worldwide. Coronary heart disease (CHD) is the leading cause, and reducing mortality is a high priority, so ML algorithms and vast datasets have been recruited to unearth relationships between patient attributes and outcomes. They are at work in diagnosis, outcome prediction, treatment, and medical imaging interpretation.
In the U.S., by 2030, 8 million people will suffer from HF, according to the American Heart Association, and associated costs could exceed $70billion [1]. The integration of AI tools is expected to help stratify patient risk, which could prevent the progression of HF, and maximize opportunities for early diagnosis, thereby opening up novel care pathways, reducing adverse outcomes, and optimizing the personalization of therapy.
In a recent study, the Mayo Clinic used AI to identify heart problems earlier by increasing the effectiveness of electrocardiogram (ECG) monitoring. By developing ECG-AI algorithms to predict a patient’s likelihood of developing heart conditions such as amyloidosis, aortic stenosis, atrial fibrillation, and more, it is now possible to calculate a patient’s biological age using single-lead ECGs from smartwatches and other portable monitoring devices.
A key application of this technology is in the detection of a weak heart pump—or low ejection fraction—and the FDA has cleared the 12-lead ECG algorithm for this use. Furthermore, an AI-assisted screening tool created by the Mayo Clinic has proven to be 93% effective in identifying people at risk of left ventricular dysfunction. In comparison, a typical mammogram for breast cancer is only 85% accurate.
Another recent breakthrough emerged from a pilot project supported by NHS England. Deployed in five U.K. hospital trusts, the technology identifies people at risk of a heart attack within the next 10 years. Hailed as a game-changer, it works by detecting inflammation in the heart that would not normally be evident from CT scans.
Developed by Caristo Diagnostics, an Oxford University spinout, it detects biological processes that precede the development of blockages within the heart, but which cannot be perceived by the human eye. Routine CT scans of patients presenting with chest pain are analyzed by the CaRi-Heart AI platform, which looks for the coronary inflammation and plaque that are associated with an elevated risk of CVD.
AI is not only becoming a powerful tool for predicting the risk of CVD and HF, but is also emerging as a key enabler of personalized treatment. In the Netherlands in 2023, the AI4HF project launched a new model for personalizing the management of HF patients that relies on a personalized risk calculator that links data on symptoms and lifestyle behaviors with blood tests, ECG readings, and cardiac imaging.
The initiative harnesses the largest-ever dataset of HF patients, using real-world data on hundreds of thousands of patients across Europe, South America and Africa through the BigData@Heart and FAIR4Health platforms, to develop the AI model.
“AI4HF promises to benefit patients with HF by adapting management to individual needs,” said project coordinator Professor Folkert Asselbergs of Amsterdam Heart Center at the launch of the initiative “In addition, a state-of-the-art AI passport will be introduced which uses new methods to continuously update the model following its deployment in real-world practice.”
Crucial caveats
AI-enabled healthcare technologies already complement the knowledge of doctors. Coupling direct patient care and data analysis enables doctors to spend more time with patients and improve decision-making processes. However, accountability is a thorny issue. As doctors rely more on AI, they cannot pass the buck for clinical decisions.
“We can use AI as a tool and have it as part of the conversation, but it requires a therapeutic relationship between a clinician and a patient,” says Johnson. “There must be some human interaction to use the information derived from AI. Without that, who is responsible for the decision? We can’t leave the system unchecked.”
For Dr. Khadijah Breathett of the Division of Cardiovascular Medicine at Sarver Heart Center, University of Arizona, the most important caveat with AI comes in the realm of health equity.
“AI has taken over all spaces at the moment and we want to make sure it is used properly,” she says. “It could perpetuate biases and we have seen decision support tools that inappropriately misuse race, resulting in poorer quality care for some groups. We must prevent algorithmic bias, so I would recommend that AI developers include equity scientists in the development process, and that they use a diverse data set.”
One example of bias raising has arisen is in the interpretation of ECG data. Some algorithms include age, sex and race to assist interpretation, and when race is a factor it may not recognize left ventricular hypertrophy (LVH)—thicker heart muscle. There has been a perception that black people tend to have thicker hearts, but LVH is not more common in black people, but rather in people with hypertension. The question then becomes why black people have more hypertension.
“If the algorithm is simply looking at pixels on a screen and making calculations, then that’s just math, which is fine,” says Johnson. “However, if decisions are derived from sensitive information, including demographic information, that is when it starts getting tricky.”
“In genetics, the data generally comes from a homogenous population—mostly white men,” adds Breathett. “There are plenty of opportunities in the advanced heart disease space to address bias and institutional racism. We could potentially leverage AI, which like any new technology brings an opportunity to advance the field, but it could make outcomes worse if done improperly.”
Future full of promise
In many ways, AI can make a clinician’s job easier by speeding up data interpretation. What a computer could calculate in seconds would take a doctor hours. Constant innovation provides exciting opportunities, but the industry must address concern about passing too much responsibility to an algorithm.
“AI is only as good as the person who created it,” says Breathett. “We have plenty of examples of erroneous recommendations that lead to poorer quality of care. You can’t run away from it, but you need to understand how to correctly use AI.”
“We must figure out how to make AI work to make care higher in quality, more equitable, more efficient,” she remarks. “I am cautiously optimistic about the use of AI. There is a lot of momentum behind it, and less momentum toward equity, so that makes me cautious, but I am an optimistic person who recognizes the power and vision of what we can do as scientists, leaders and communities to move the field forward in a way that provides optimal care.”
AI undoubtedly has tremendous potential in cardiology, but to fully unleash its power to improve patient outcomes it is not only technology that needs to advance. So, too, must attitudes toward ethics, equity, and accountability.
References
- P. A. Heidenreich et al., “Forecasting the impact of heart failure in the United States: A policy statement from the American Heart Association,” Heart Failure, vol. 6, no. 3, pp. 606–619, 2013. [Online]. Available: https://www.ahajournals.org/doi/10.1161/HHF.0b013e318291329a