Causal AI ‘translates’ CAD polygenic risk score into clinically actionable data
Click Here to Manage Email Alerts
NEW ORLEANS — Causal artificial intelligence used to augment a polygenic risk score for CAD helped estimate how much individual patients must lower their LDL, systolic BP or both to overcome inherited risk for a major coronary event.
A polygenic score combines genetic variants from all pathways leading to CAD; however, the risk score does not provide specific information about why a person is a risk, how to minimize that risk, or how much they will benefit from specific actions to reduce risk, Brian A. Ference, MD, MPhil, MSc, FACC, FESC, professor and director of research in translational therapeutics and executive director of the Centre for Naturally Randomized Trials at the University of Cambridge, said during a late-breaking clinical trials presentation at the American College of Cardiology Scientific Session. Because of these uncertainties, Ference said, researchers and clinicians struggle with how polygenic scores can be used to inform individual treatment decisions.
Image: Adobe Stock
“Importantly, persons who maintain low lifetime exposures to LDL and BP have a low lifetime risk for CV events, even at all levels genetic predisposition,” Ference said. “This finding implies that polygenic predisposition for CV events can be overcome by maintaining lower LDL and BP. However, how much a person needs to lower their LDL and BP, or both, to overcome their polygenic risk predisposition, is unknown.”
Predicting how low to go
Ference and colleagues used a polygenic risk score consisting of 4,051,820 variants to estimate lifetime risk for major coronary events, defined as fatal or nonfatal MI or coronary revascularization, among participants in each decile of the polygenic risk score. Researchers then used a causal AI algorithm that integrates information from Mendelian randomization studies and randomized trials to estimate the reduction in LDL or systolic BP needed to reduce risk among participants in each decile of polygenic risk score to the same level as participants with an average polygenic risk score. The researchers observed event curves in each group to validate the estimates.
“Causal AI encodes biological cause and effect to create algorithms that can both predict outcomes and prescribe specific actions to change outcomes, by accurately quantifying the effects of changes in exposure to the modifiable causes of disease,” Ference said. “The purpose of this study was to test the hypothesis that causal AI can translate polygenic risk scores into clinically useful information.”
Among 445,774 participants (54.1% women) followed to a median age of 71 years, 31,524 experienced a first major coronary event during 17,054,722 person-years of follow-up.
For most, polygenic risk for CAD was overcome with small lifetime reductions in LDL and systolic BP. However, the amount of LDL or systolic BP lowering needed to overcome polygenic risk increases the later LDL or systolic BP lowering is initiated, Ference said.
“Overall, causal AI estimated that the polygenic risk for CVD is a relatively weak risk factor for the vast majority of the study population that can be easily overcome with lower BP or a few mg/dL lower LDL,” Ference said.
In randomized tests, the causal AI was able to accurately estimate how much lower an LDL or BP was required to overcome every level of polygenic risk predisposition at all levels of polygenic risk score.
“For example, to overcome polygenic risk in the 80th percentile would require approximately 14 mg/dL lower LDL or a combination of 7 mg/dL lower LDL plus 2.5 mm Hg lower systolic BP,” Ference said.
The causal AI also estimated that, to overcome CV risk for people with a polygenic score in the 90th percentile would require a 20 mg/dL lower LDL or a combination of 10 mg/dL lower LDL plus a 3.5 mm Hg lower systolic BP.
Incorporating family history
The data also showed that family history is a strong risk factor for CV events, equivalent to a polygenic risk score in the 95th percentile or higher, Ference said, adding that polygenic risk and family history of CHD provide independent and additive information.
“Causal AI can substantially enhance the utility of polygenic scores for informing individual patient decisions,” Ference said. “In additional, causal AI can be used to accurately estimate how much each person needs to lower their LDL, BP or both, to overcome their overall inherited risk for CV events, including their family history and their polygenic predisposition, depending on the age at which LDK or BP lowering is started.”
Ference said an online app for translating polygenic risk score for CAD into clinically actionable information will be available at www.DeepCausalAI.org.
Perspective
Back to TopJennifer R. Dungan, PhD, RN
Dr. Brian Ference’s research team has placed real-world data on cardiac risk factors hand-in-hand with polygenic risk (PRS) scores and AI to produce a realistic approach to precision CV health that has potential beyond the hype of AI. From a CV genetics and precision health lens, it may be one of the most promising ways to revive PRS for actual clinical translation.
The inclusion of a PRS — particularly one with more than 4 million gene markers — is a strength of this work. Genetic variation and family history are nonmodifiable risk factors which explain about 50% of the variation in CVD. Yet, a simplistic family history indicator represents this complex risk factor in only some of the current CVD risk calculators used in health care. CVD is complex and multifactorial, which means that there are many genes, each with small effects, having gene-gene and gene-environment interactions that lead to a great deal of heterogeneity. Every person has an average of 1 gene variant per every 1,000 DNA base pairs, meaning that we all carry 3 to 4 million DNA variations. Some of those DNA changes that have a known inheritance pattern are rare (< 1% of population) and confer causal effects for CVD, such as with single-gene diseases like Marfan or Liddle’s syndrome, as compared with some DNA markers which are fairly common and their effects may (or may not) lead to increased risk of CVDs such as atherosclerosis and hypertension. It would be important for Ference’s PRS to contain the most validated DNA markers associated with strongly inherited CVDs and also the many variants associated with the more common, complex CVDs.
Is PRS plus AI plus risk factor data going to be the perfect solution for making a meaningful dent in the global CVD morbidity and mortality rates? More research and validation may offer answers. Are the AI algorithms and the data that drive them unbiased and generalizable to all people and contexts? Definitely not yet. Is this approach without limitations or flaws? Unlikely. But this work offers real hope for embracing the clinical utility of PRS, maximizing real-world clinical data towards personalized risk factor profiles, and has tremendous potential to meet people where they are with customizable, achievable risk reduction goals.
Perspective
Back to Top
Robert A. Harrington, MD, FAHA
It has always been uncertain what to do with polygenic risk scores if you are the average clinician or the average patient. They are hard to translate into action. But that is what was accomplished in this study. And not just general recommendations, but specific ones about, for example, how much a patient should lower their LDL or BP.
The most important findings were that this approach was actionable and personalized. It enables you to show the data to a patient and tell them it can be used to reduce their risk for heart disease or stroke if their cholesterol is lowered by a certain amount. It gives doctors and patients something actionable to work on; likely, multi-drug therapy will be involved. If we know how much remains toward the goal, we can determine what further measures to take, including which drugs to add.
One thing that worries clinicians about AI in health care is the “black box” nature; it bothers them if they don’t understand what is in it. We all have to help each other overcome it. What is nice about this system is that even though it uses advanced computation of AI, it makes sense to clinicians. It takes the clinician to the next step. If the first step is control of CV risk factors, the next step is by how much, and how much does one get by changing certain parameters of one’s risk. That makes a lot of sense to clinicians.
How we get blood chemistry results is also a “black box,” but clinicians are comfortable with that because they’ve been trained on it since day 1.
We are going to have to constantly test, update and reevaluate these models. A challenge for most AI systems is that the data comes from a small number of places. One paper showed that there are entire states that have not contributed any data to the AI algorithms being built. Most of it comes from California and Massachusetts. What about data from the people from Mississippi, Louisiana and other high-risk states? It’s an issue that has got to be addressed.
Collapse
Ference BA, et al. Late-Breaking Clinical Trials III. Presented at: American College of Cardiology Scientific Session; March 4-6, 2023; New Orleans (hybrid meeting).
Click Here to Manage Email Alerts