NEW ORLEANS – Causal artificial intelligence (AI) can translate polygenic scores (PGS) and other genetic information into risk reduction strategies for coronary artery disease (CAD) that is tailored for each individual patient, according to an analysis presented at the joint scientific sessions of the American College of Cardiology and the World Heart Federation.
Tested for LDL cholesterol (LDL-C) and systolic blood pressure (SBP), causal AI explained how much each of these risk factors must improve at the level of each individual patient “to overcome overall inherited risk,” reported Brian Ference, MD, MPhil, director of translational therapeutics, University of Cambridge (England).
Ted Bosworth/MDedge News
Dr. Brian Ference
Unlike the “black box” risk assessments common to machine learning, which relies on disparate forms of information of often unknown relative significance, causal AI explains cause and effect. In the case of CAD, its ability to encode the biological causes means that it can “both predict outcomes and prescribe specific actions to change those outcomes,” Dr. Ference explained.
The concept is testable against observed biology using randomized evidence, which was the objective of the study Dr. Ference presented in the late-breaker session.
Causal AI trained on nearly 2 million patients
This study employed a causal AI platform trained on roughly 1.3 million participants in Mendelian randomization studies, as well as more than 500,000 participants in randomized clinical trials. The PGS estimate of inherited risk was constructed from almost 4.1 million variants from genomewide association studies.
To test the ability of causal AI to reveal how much LDL-C or SBP had to be reduced to overcome the inherited risk of CAD based on PGS, it was applied to 445,765 participants of European ancestry in the UK Biobank. The goal was to determine how much those with greater than average risk would need to lower their LDL-C or SBP to achieve average CAD risk.
When validated against observed rates of events, causal AI accurately characterized risk before estimating what reductions in LDL-C, SBP, or both would attenuate that risk.
Providing examples, Dr. Ference explained that a PGS in the 80th percentile can be overcome by lowering LDL-C by 14 mg/dL. Alternatively, the 80th percentile risk could also be overcome by simultaneously lowering LDL-C and SBP by 7 mg/dL and 2.5 mm Hg, respectively.
Required risk factor reductions increase with age because of the increased risk of the events. For example, while a 14.8 mg/dL reduction in LDL-C would be adequate to overcome risk defined by a PGS in the 80th percentile at age 35, reductions of 18.2 mg/dL, 28.9 mg/dL, and 42.6 mg/dL would be required, respectively, at ages 45, 55, and 65 years. The values climb similarly for SBP.
Family history of CAD adds an independent variable that further contributes to the ability of causal AI to estimate risk and the degree of risk factor attenuation to overcome the risk.
Even though family history is equivalent to having PGS above the 95th percentile, it is an independent and additive variable, according to Dr. Ference. As a result, inherited risk of CAD depends on both.
Still when family history is factored into the analysis, “causal AI accurately estimated the magnitude of lower LDL-C, SBP, or both needed to overcome overall inherited risk at all levels of higher or lower PGS,” he reported.
According to Dr. Ference, the value of causal AI is that it can generate very specific goals for each patient regarding modifiable risk factors. Causal effects of risk factors encoded in time units of exposure allow the patient and the clinician to understand the biology and the basis of the disease burden.