From Prevention to Prognosis: AI's Role in Heart Failure

This transcript has been edited for clarity. 

Hello. I'm Dr Michelle Kittleson, a heart failure transplant cardiologist at Cedars-Sinai in sunny Los Angeles, California. I'm delighted to have this opportunity to present a commentary on artificial intelligence (AI) in heart failure. I'll cover how AI is being used in clinical care, including to assess those at risk for heart failure and to catch early signs of illness and improve patient outcomes. 

Let's start with the basics. What is AI? AI is defined as the study of intelligent agents, which can perceive the environment and act as humans do, typically for narrowly defined tasks. AI often uses machine learning, whereby statistical and mathematical models and algorithms can progressively learn from data to achieve the desired performance on specific tasks. Machine learning works well where humans find it difficult to manually develop a set of rules for desired tasks, and the appeal of this approach is immediately obvious. 

As clinicians, we sift through complex data constantly to engage in pattern recognition that results in diagnostic and therapeutic decisions. How do we know how to weigh different data points or conflicting data points? Often this boils down to judgment, experience, style of practice — the art of medicine. Can we rely less on art and more on science?

Let's imagine how these approaches are being evaluated in the diagnosis and management of heart failure. We'll cover the role of AI in the prediction of risk for incident heart failure, heart failure diagnosis, and determination of prognosis.

Let's begin with AI in the prediction of incident heart failure. Can AI make us better at early detection? The data are promising. There are multiple studies, including some with around 500,000 patients examining over 300 predictor variables able to predict incident heart failure. 

In my opinion, the most exciting of these studies are those that cull data from the electronic health record (EHR). It will be incredible to receive an EHR alert advising a screening echocardiogram, or natriuretic peptide assessment for patients who screened as having a high risk for heart failure, especially because the screening tool offers no additional risk to the patient.

My take on AI in the detection of incident heart failure? The future is bright. 

One of the reasons I fell in love with cardiology as a trainee was that both patient history and the physical exam results matter. The diagnosis of heart failure, for example, relies on patient history and physical examination in conjunction with imaging and laboratory data. Clinicians have to be careful not to overtest or undertest. Can AI help with that? 

In fact, AI-based models, including some examining over 50,000 patients using electrocardiogram heart sounds and EHR data, have excellent performance metrics for the diagnosis of heart failure. For those patients who have already had an echocardiogram, AI can still help. AI-powered echocardiography may reduce interobserver variability. 

In fact, some of the algorithms based on deep learning have also been approved by the US Food and Drug Administration (FDA) for implementation in clinical practice to deliver automatic and repeatable analysis of cardiac images that are just as accurate as segmentations carried out manually by medical professionals. 

This is a particular diagnostic challenge in heart failure with preserved ejection fraction (HFpEF), given that it is a clinically heterogeneous syndrome lacking uniform pathophysiology upon which to target therapy. The promise of AI would be in designing future clinical trials and possibly guiding the selection of personalized therapy. 

In fact, using unsupervised clustering algorithms, a form of AI that takes into account clinical, cardiac, and imaging parameters, several studies have identified subgroups of patients with distinct phenotypes of HFpEF based on their profiles and outcomes. This is hugely important in a disease where we believe that our inability to have clear guideline-directed medical therapy that is truly effective is because it is so heterogeneous, as opposed to something like heart failure with reduced ejection fraction, where the echocardiogram often tells much of the story. 

Let's move on to determination of prognosis. So much of being a clinician is about predicting the future. Will my patient develop heart failure, and if they develop heart failure, what will their prognosis be? In my opinion, this is an area of great need where we rely mostly on judgment and experience, with variability between clinicians based on their interpretation of many complex data points. 

However, studies to date of AI in prognostication are limited owing to low reliability at individual patient–level risk, the variety of approaches to choose from, and the complexity of the statistical methodologies used. When it comes to prognostication, in a way, we have to rely on our clinical variables. 

One of the mnemonics most commonly used by advanced heart failure cardiologists is "I Need Help", focusing on those clinical markers that may be a predictor of poor outcomes, including the need for inotropic support, worsening symptoms, very low ejection fraction, end-organ damage such as kidney or liver dysfunction, defibrillator shocks, frequent hospitalizations, escalating need for diuretics, and inability to tolerate guideline-directed medical therapy. 

When those types of factors come up in clinical practice, our judgment and experience tell us the patient's prognosis is limited, and we need to start thinking about things like advanced therapies and prognosis discussions. 

We've discussed AI and the prediction of incident heart failure, diagnosis of heart failure, phenotyping of HFpEF, and determination of prognosis. When can we expect the incorporation of AI technologies into clinical practice? Before that, it's important to note the pitfalls, both scientific and practical. 

What are the scientific concerns? Finding the sweet spot of the number of variables to incorporate into the model is key. Too few variables may mean overfitting to the data set, with less widespread applicability. Too many variables means concerns for data quality and model performance. The key here is external perspective validation, where the model is then applied to a different population of the same type of patients. This is not currently implemented in most studies to date, but it is the key to getting it just right, with not too few variables and not too many variables, which leads to some practical concerns.

First, data privacy and data management principles are necessary to allow for training of algorithms on datasets of large numbers of patients, typically culled from an EHR, while also maintaining an individual's privacy. Second, because algorithms must be subjected to rigorous inspection before they can be implemented in clinical practice, particularly with prospective validation of algorithms, the regulatory licensing process will take time. 

The third challenge will undoubtedly be adoption and implementation by clinicians. The inertia will probably be due to some mistrust of AI models and "alerts fatigue" due to responding to automated queries in the EHR. Thus, education of the benefits and pitfalls of AI approaches will be essential, as will practical implementation strategies that help rather than hinder clinical workflows. 

Regulatory approval and certification will certainly help further establish clinicians' trust in the safety and efficacy of AI algorithms. Clinicians will be more likely to trust AI algorithms that have been approved by regulatory bodies, such as the FDA, and have undergone rigorous certification processes. 

Now that we've covered the challenges, let's end by looking forward to the future. What would a perfect world of AI in heart failure look like? I suppose it would look like a world where we get so good at prevention, there's no need for heart failure cardiologists like me. 

Short of that, in an ideal world AI would optimize enrollment in clinical trials; forecast the likelihood of adverse events occurring based on patient characteristics, such as age, gender, and medication history; and predict treatment response to individualized therapies. 

In short, if it works, it will allow us to use science to improve upon the art of medicine. 

I hope you've enjoyed this commentary on AI and heart failure. Thank you for listening.