Dr Paul Friedman, Professor of Medicine and Chair of the Department of Cardiovascular Medicine at Mayo Clinic, Rochester, Minnesota, delivered the lecture entitled: Innovation in Cardiovascular Medicine: Artificial Intelligence, the Next Frontier
Highlighting how Artificial Intelligence (AI) can be the ‘next frontier’ in medicine was the critical message from Dr Paul Friedman as he delivered the ESC 2020 Paul Hugenholtz lecture for innovation. Citing examples of where AI can outperform humans in terms of speed, workload, and capability, Dr Friedman also demonstrated its potential to pick up sensitive markers and predict cardiac events, potentially years before they may happen, allowing the opportunity to implement mitigating therapies.
He suggested that in medicine today, once a patient develops signs and symptoms, there is a ‘fundamental problem with the way we approach this’. ‘We begin ordering diagnostic testing and forming a treatment plan but in cardiovascular medicine it may be too late because the first event may be sudden cardiac death or a stroke or heart attack’, he said. ‘But we know that the metabolic and physiologic derangements that are causing that atherosclerotic lesion have been around for at least a decade before the event. If we have some way of detecting it, perhaps we could intervene before that clinical event happens’.
It is the potential for an AI role in this that Dr Friedman demonstrated throughout his lecture. Using the example of asymptomatic left ventricular dysfunction, which affects 2% of the global population, he pointed to professional society guidelines, multiple prospective randomized trials, and treatments that lower mortality and hospitalization. ‘The key is that we have to find out that the condition is there’, he added.
Commonly used tests such as electrocardiogram (ECG), computed tomography (CT), and magnetic resonance imaging (MRI) scans can find it, but his team at Mayo Clinic hypothesized that they could use a convolutional neural network, a form of AI, to read the ECG to see whether left ventricular dysfunction was present.
Feeding data from 100 000 ECGs into the programme, it was trained to identify the presence of a low ejection fraction, which cannot usually be determined from an ECG. The area under the receiver operator characteristic—a measure of the ability of the algorithm to read an ECG and see that a weak heart pump was present-was 0.93.
‘That was a very powerful test. But what was really striking was when we had a false positive’, he continued—when the computer said low ejection fraction, while the echocardiogram said normal ejection fraction. ‘If you follow those patients over the next six years, they have a five-fold increased risk of developing a weak heart pump. It is like it is predicting the future’, he added.
‘What is actually happening is that an early pathologic process is affecting those electrical currents and leading to subtle electrocardiographic changes. So, it appears to predict disease, or identify it more accurately before it becomes manifest and thus it may identify people at high risk for whom follow-up imaging studies may be appropriate’.
The computer algorithm also determines gender from the ECG, as well as a patient’s physiological age. People with hypertension, heart failure, or myocardial infarction were shown as older than their physiologic age meaning the algorithm could be a potential marker of physiologic age and a tool to make decisions over patients that are, or are not, suitable for certain therapies. ‘Maybe someone is too old to have that procedure irrespective of their chronologic age, or maybe that 90-year-old would be a good candidate for a transcatheter aortic valve replacement (TAVR) because physiologically they are only 70. Those are questions we are asking now’.
Another algorithm detected the presence of episodic atrial fibrillation from an ECG recorded during normal rhythm. Dr Friedman said: ‘It raises this concept of what if you knew you were going to get a disease, could you prevent it? And then, what are the social, insurance, and legal implications?’
The AI approach is massively scalable. It can be used in multiple formats, including smartphones, and further research at Mayo is under way to assess its application in primary care and integration into practice workflows. The potential extends to clinical trials—AI can dip into the Electronic Medical Record, screen patients, use natural language processing to identify individuals with specific phenotypes and rapidly identify candidates for research protocols and invite them into a study programme.
‘We have a rich pipeline of AI tools, but the important question is how do we bring this into practice, how can we make this impact patients’ lives for the better’, he said.
Dr Friedman said that AI requires a partnership of clinicians actively engaged with engineers to develop these tools and they must integrate into the workflow. Test, vet, validate, and educate are also key components of the process.
In terms of the historical arc of diagnosis, he concluded that AI is truly the next frontier, and a tool that will significantly advance our ability to understand and treat disease.
Conflict of interest: none declared.
