Artificial intelligence Atrial Fibrillation Shaan Khurshid AF Symposium 2023
Artificial intelligence (AI) has shown promise in predicting atrial fibrillation (AF) risk by analyzing various patient data such as demographics, medical history, and electrocardiogram (ECG) recordings, which can potentially help identify high-risk patients and facilitate early interventions to prevent AF-related complications. Dr Shaan Khurshid (Massachusetts General Hospital, Boston, MA, USA) discusses recommended guidelines for AF screening, and current challenges in screening. He also goes on to highlight the types of AI that are being developed to improve AF risk estimation, the clinical trial findings from PULsE-AI (NCT04045639) and BEAGLE (NCT04208971) clinical trials to support the use of AI, and what he expects as future directions.
Full transcript available below
The abstract entitled ‘Use of Artificial Intelligence for Prediction of Incident Atrial Fibrillation‘ was presented at the Atrial Fibrillation Symposium, 02 – 04 February 2023.
Watch Dr Shaan Kurshid’s highlights on incident and risk of atrial fibrillation
Questions:
- What are the recommended guidelines for AF screening, and what are the current challenges? (0:27)
- What types of artificial intelligence (AI) are being developed to improve AF risk estimation? (1:54)
- Could you tell us a little about PULsE-AI and BEAGLE clinical trial findings and conclusions? (3:09)
- What do you consider the future directions in the development of AI for AF screening? (5:34)
Disclosures: Shaan Khurshid has nothing to disclose in relation to this interview.
Support: Interview and filming supported by Touch Medical Media. Interview conducted by Katey Gabrysch.
Transcript
I am Shaan Khurshid. I am a Cardiac Electrophysiology Fellow, Massachusetts General Hospital, Boston, MA, USA. I’ll be joining the electrophysiology staff this July, and I have a clinical and research interest in atrial fibrillation (AF) and in particular the prediction of disease using statistical methods as well as novel machine learning methods.
What are the recommended guidelines for AF screening, and what are the current challenges? (0:27)
So there have been multiple randomized trials of screening for AF. The reason why screening may make sense is that AF can be asymptomatic and episodic. Therefore, people can go in and out of AF without ever knowing. So it may make sense to look for AF even among people who are healthy or otherwise don’t feel any symptoms. That has been tested, as I mentioned, in multiple trials, but those trials are mixed. The earlier trials showed a benefit for AF screening using things like pulse palpation or 12 lead ECG. But more recent studies show that screening doesn’t really meaningfully increase AF yield as compared to usual care. It’s possible that usual care is more comprehensive now than in those earlier trials. Because of that, there’s controversy in the field as to whether AF screening should be performed based off of those earlier trials. The European guidelines from the European Society of Cardiology and others societies recommend AF screening for all individuals 65 and older using pulse palpation or ECG, with similar guidelines existing in New Zealand and Australia. However, in the United States Preventive Services Task Force has actually concluded that the evidence is insufficient to recommend for or against a screening. So there’s some controversy there.
What types of artificial intelligence (AI) are being developed to improve AF risk estimation? (1:54)
One exciting potential avenue is the use of AI (artificial intelligence) to predict AF risk and provide some potential advantages over other clinical approaches. Number one, these clinical scores are cumbersome to calculate and haven’t made their way into clinical practice. AI models have the potential for automated outputs that you could integrate with an EHR in the setting of routine clinical care. Another major advantage is that AI models can utilize raw data, things like the ECG itself that we routinely ascertain in a lot of patients which can provide additional predictive value or those clinical risk factors that I mentioned. One avenue that we’ve been working on as well as others, is using those raw modalities like the EKG and clinical risk factors and combining them into deep learning models to output an individualized AF risk estimate. One of the main highlights of the talk that I gave at the AF Symposium was reviewing some of these recent algorithms, including a model that we developed that combines clinical risk factors and the 12 lead ECG to provide a more accurate AF risk estimate than either an EKG alone or clinical risk factors alone.
Could you tell us a little about PULsE-AI and BEAGLE clinical trial findings and conclusions? (3:09)
It has been demonstrated that you can predict AF risk using AI. The next step is actually bringing those AI algorithms back to the bedside, and that’s the rationale for these studies. PULsE AI and BEAGLE clinical trials have the overall concept to use an AI algorithm, risk stratify individuals for AF, and then screen the people that are highest AF risk, so that you are more likely to have an efficient screening intervention, more likely to find AF cases in an efficient manner. That’s the rationale.
The PULsE AI study was a randomized trial in the United Kingdom using a neural network model for screening. Individuals randomized to intervention underwent risk stratification with a neural network, and those who were high risk were offered a screening intervention with ECG and service monitoring. Individuals who were randomized to control received usual care only. In PULsE AI there was no significant difference in newly diagnosed AF among screen versus control, but the screening intervention itself, the uptake was low. Only 30% of people who were offered screening actually did it, which is a major limitation among the people who received screening, there was a much greater odds of being diagnosed with new AF suggesting that the concept may be effective, but people need to actually participate in the intervention for it to work.
The BEAGLE study was a nonrandomized assessment of risk guided screening in the United States. People underwent risk stratification with an ECG based model, and again, those who are predicted to be at high risk were offered screening with service monitoring. In that study, there was a much there was a five-fold greater odds of being diagnosed with new AF amongst the high risk individuals who were offered screening. So it’s showing significant enrichment. Although that study was not randomized, they then compare their study population to matched controls based on clinical risk factors and showed that the increase in AF diagnosis was seen, particularly in the people who were high risk using the algorithm, demonstrating that concept that if you target high risk people with a screening, you’re more likely to have a more efficient, efficient screening approach.
What do you consider the future directions in the development of AI for AF screening? (5:34)
There has been a lot of encouraging work as of late in this field, but there’s a lot of work to do still. Some examples of where we need to go next could include models that have focused on the ECG because it’s available and it carries AF risk information. But there are other modalities that we can include to make the estimates even more accurate. We can include cardiac imaging, MRI, photoplethysmography from something like a watch, all to improve the AF risk estimation accuracy even further to early models that have tested model accuracy in similar populations as the ones where models were trained. More and more, we’re going to want to validate these models across large and distinct populations with varying patient characteristics, different countries, to make sure that the models are behaving in the way we think they are and that they’re robust and generalizable. Lastly, AF risk estimates in and of themselves is nice, but in order to realize the potential to improve outcomes, we’re going to want to integrate those AF risk estimates with the clinical action. So some of these studies that I just discussed are starting to do that, where we not only take a risk estimate, but then do something about it, like screening.
Subtitles and transcript are autogenerated
