‘Causal AI’ May Help Guide Specific Actions to Reduce CVD Risk

Researchers have used artificial intelligence to estimate how much LDL and BP must decrease to overcome inherited risk.

‘Causal AI’ May Help Guide Specific Actions to Reduce CVD Risk

NEW ORLEANS, LA—An artificial intelligence (AI) system that incorporates concepts of biological cause and effect can, in combination with a polygenic score (PGS) for coronary artery disease, quantify how much individuals must lower their blood pressure and cholesterol levels in order to overcome their inherited risk.

“There’s a great deal of interest in trying to determine how we can use both polygenic scores and AI to directly inform individual patient care,” said Brian A. Ference, MD (University of Cambridge, England), when presenting the results this week in a late-breaking clinical trial session at the American College of Cardiology/World Congress of Cardiology (ACC/WCC) 2023 meeting.

One challenge is that PGS are based on a huge number of genetic variants, potentially millions, that predict lifetime risk of disease. “By combining all of these variants in all of the pathways leading to cardiovascular disease into a single metric, a polygenic score does not provide any specific information about why a person’s at risk, how to minimize that risk, and how much they’ll benefit from specific actions to reduce risk,” he said, adding that this means there’s much uncertainty over how to apply PGS in clinical practice.

“Causal AI” can help facilitate that application, Ference explained in his presentation. The technology “encodes biological cause and effect to produce algorithms that can both predict outcomes and prescribe specific actions to change those outcomes by reliably quantifying the effects of changes in exposure to the modifiable causes of disease,” he said.

As for how the technology will disseminate into real-world use, Ference told TCTMD: “I think that without question polygenic scores will become routine in clinical medicine in the next decade or so. It’s [a source of] a lot of information. It can be measured once and then iterated through and improved over time as the information becomes more actionable or useful.”

Calculating an individual’s PGS would require obtaining the patient’s genetic material, either through a blood sample or buccal swab. Causal AI could analyze the resulting PGS using a particular algorithm like the one tested here or be used in other contexts like electronic health records.

Just Small Changes in BP, LDL

“Importantly, persons who maintain low lifetime exposure to LDL and blood pressure have a low lifetime risk of cardiovascular events, including at all levels of genetic predisposition,” he explained, meaning that intervening to reduce LDL and blood pressure levels could prove useful at overcoming any inherited risk. Indeed, their study showed that, for the vast majority of the population, polygenic predisposition is a weak enough risk factor that it’s surmountable by small changes.

Ference and colleagues started out by training the causal AI system to estimate how much levels of systolic BP and LDL cholesterol impact MACE risk, based on data set of 1.8 million individuals from randomized clinical trials and genetic studies. They also developed a PGS for coronary artery disease using 4 million variants identified by genome-wide association sudies.

Then, they turned to the UK Biobank database and calculated PGS for 445,774 of that registry’s participants (54.1% women), who were followed up until the median age of 70.9 years. In all, 31,524 experienced at least one major coronary event (MCE; fatal or nonfatal MI, or coronary revascularization) during 17 million person-years of follow-up. The researchers applied their AI model to that UK Biobank population in order to estimate how much someone with above-average inherited polygenic risk would need to reduce their blood pressure, LDL cholesterol, or both to overcome that predisposition toward MCE.

And, as a final step, Ference et al used mendelian randomization to validate their system’s accuracy. What they found is lowering LDL and systolic BP does appear to overcome polygenic risk.

Individuals in the 80th percentile of PGS only needed to reduce their LDL by 14.6 mg/dL to match those in the 50th percentile. Equally effective would be to both lower LDL by 7.3 mg/dL and systolic BP by 2.5 mm Hg.

Individuals with PGS in the 90th percentile could overcome that inherited polygenic risk by lowering LDL by 20.3 mg/dL or achieving the combination of 10.1-mg/dL lower LDL and 3.6-mm Hg lower systolic BP.

The earlier the process started, the better. At age 35, someone in the 80th PGS percentile could overcome their risk by reducing their LDL by 14.8 mg/dL or systolic BP by 5.3 mm Hg, or with the combination of 7.5-mg/dL lower LDL and 2.5-mm Hg lower systolic BP. For someone who initiated treatment at age 65, the required reductions were greater: 42.6 mg/dL in LDL, 15.9 mm Hg in systolic BP, or the combination of 21.3 mg/dL and 8.0 mm Hg, respectively.

Ference specified that a person’s inherited risk is influenced both by polygenic predisposition and, to an even greater degree, by whether or not they have a parent with coronary heart disease [CHD]. “The distribution of polygenic score among persons with and without a family history of CHD is essentially the same—likely explaining why polygenic scores and family history have independent and additive effects on risk of major coronary events,” he said.

For someone who has one parent with CHD and an 80th percentile PGS, for example, LDL cholesterol would need to drop by 21 mg/dL and systolic BP by 5 mm Hg to overcome their total inherited risk. For someone who has a parent with CHD but 20th percentile PGS, their total inherited risk is canceled out by a 10 mg/dL reduction in LDL.

Actionable Information

Jagat Narula, MD, PhD (UTHealth Houston, TX), the discussant following the presentation, described the research as “absolutely outstanding.” Highlighting the American Heart Association’s 2022 scientific statement on polygenic risk scores for CVD, Narula praised Ference and colleagues for the steps they took to verify that “causal AI could translate a polygenic score for coronary disease into clinically actionable information to help inform personalized treatment decisions,” he said.

Narula called for a closer look at whether factors like smoking, pollution, obesity, metabolic disorders, mental health, and stress might modify their results.

Engaging the patients with agency, explaining to them that there’s something specific about themselves and that there is only this much that they need to do to make this better—that is a different way of talking to patients than we’ve ever talked to them before. Ami B. Bhatt

Ami B. Bhatt, MD (Harvard Medical School, Boston), chief innovation officer for the ACC, said that given the worldwide burden of cardiovascular disease, it’s important to address health at a population level. “What you managed to do is address population health with precision health, and that is challenging,” she told Ference during a press conference following his presentation.

Explaining the implications to the media, Bhatt said: “When we think about those patients who have this disease, first of all we have to find the disease. Family history is one way, but it is not the only way—the polygenic risk score helps.”

Once diagnosed, though, patients then need to get on guideline-directed medical therapy. “Engaging the patients with agency, explaining to them that there’s something specific about themselves and that there is only this much that they need to do to make this better—that is a different way of talking to patients than we’ve ever talked to them before,” she noted.

As of now, she continued, patients are told: “You need to be on a statin.” With causal AI, the message might be more along the lines of, “Let me tell you about yourself based on your score, your family history.” The latter approach, said Bhatt, “is really empowering of patient partnership and may be very successful.”

As for the AI tool itself, here, too, there are unique aspects, Bhatt noted. Rather than the usual AI approach of “I notice X happens and then Y; therefore, they must be related,” she explained, “you chose causal AI to say there is biology behind this that is [causing] this result.” This clarity about methods will enable clinicians to better understand—and explain to their patients—how making small changes can affect risk.

Next Steps on a Global Scale

Bhatt stressed that the majority of heart disease-related deaths occur in lower- and middle-income countries. As such, attention first should turn to specific populations in the United States and globally where this technology might prove beneficial, she said. “How can we democratize it for everybody?”

“That is a fantastic observation,” Ference agreed, noting that the 2022 World Heart Federation Roadmap for Cholesterol specifically mentions the potential role of causal AI as a tool to bolster prevention efforts. The idea is to “create equitable healthcare systems that can prevent disease more effectively rather than waiting for disease to happen, falling further and further behind.”

Next up for their project is a “400,000-person real-world, adaptive cluster-randomized trial,” taking place in Estonia, that will test their technology’s ability to prevent cardiovascular disease by reducing cumulative exposure to LDL, he said.

Speaking with TCTMD, Bhatt pointed out that the logistics of obtaining PGS for patients may be more difficult in certain settings with fewer resources. “In the low- and middle-income countries, how do we use causal AI? The first thing we have to do is study it there,” she said. Estonia is a good first step, but “the next step is actually to think about: how do we study this in a more-diverse population?”

Within the United States and around the world, structural racism continues to widen health disparities, Bhatt noted. “A lot of the Biobank and other data that we have may not be reflective of some of the places that have the greatest burdens of cardiovascular disease. So we have work to do in doing trials to ensure we have the right causal AI for the right populations,” she said, citing the example of South Asia, which makes up about 25% of the world’s population and 60% of the burden of cardiovascular disease.

That said, the study presented at ACC/WCC “is a fantastic start,” Bhatt emphasized.

Sources
  • Ference BA. Translating polygenic risk for coronary artery disease into clinically actionable information using causal AI. Presented at: ACC/WCC 2023. March 5, 2023. New Orleans, LA.

Disclosures
  • Ference reports consultant fees/honoraria from the American College of Cardiology, Amgen, Daiichi Sankyo, Eli Lilly and Company, the European Atherosclerosis Society, the European Society of Cardiology, Ionis Pharmaceuticals, KrKa Pharmaceuticals, Merck & Co, Mylan, Novo Nordisk, Pfizer, Regeneron, Sanofi-Aventis, and The Medicines Company, as well as research grants from Amgen, Esperion, Merck & Co, and Novartis Corporation.

Comments