Typically, artificial intelligence (AI) is applied to analyze a complex set of variables to make correlations not readily made by unassisted observation. But an AI offshoot, sometimes referred to as causal AI, incorporates causation not just association, and it appears capable of changing the paradigm for preventing cardiovascular (CV) events.
“Causal AI is a new generation of AI algorithms that empowers AI to move beyond prediction to help guide clinical decision-making for each individual,” reported Brian A. Ference, MD, director of research in translational therapeutics, University of Cambridge (England).
In a novel study testing this premise, called CAUSAL AI, this approach was explored with two major risk factors, elevated LDL cholesterol (LDL-C) and elevated systolic BP (SBP). Based on a deep learning algorithm that studied the impact of these risk factors on the biology of atherosclerosis, causal effects of these risk factors were assessed and then embedded in risk estimation.
Causal AI Can Predict Treatment Effect
The study showed that the accuracy of risk prediction can be improved markedly with causal AI, but, more importantly, it suggests that causal AI can predict the impact of specific actions to reduce this risk in the context of the patient’s trajectory toward CV events.
“Risk-estimating algorithms are used to select patients at high risk who may benefit from interventions to reduce risk, but they do not include the causal effects of changes in LDL-C and SBP,” Dr. Ference explained.
As a result, they “may not accurately estimate the baseline risk of cardiovascular events caused by a person’s LDL-C or SBP level or the benefit of treating these risk factors,” he added.
In the CAUSAL AI study, presented at the annual congress of the European Society of Cardiology, risk prediction embedded with causal AI demonstrated the ability to match predicted events with actual events in several large sets of patient data.
“Embedding causal effects into risk-estimating algorithms accurately estimates baseline cardiovascular risk caused by LDL and SBP and the benefit of lowering LDL, SBP, or both beginning at any age and extending for any duration,” Dr. Ference said.
Deep-Learning AI Evaluated More Than 300 Gene Variants
The deep-learning AI was based on Mendelian randomization studies evaluating 140 gene variants associated with LDL-C and 202 variants associated with SBP.
In one test of the predictive impact of causal AI, risk prediction was first conducted in 445,771 participants in the UK Biobank with the Joint British Societies (JBS3) risk calculator. Relative to actual events in this population, the JBS3 alone “consistently underestimated the increased risk caused by elevated LDL, blood pressure, or both” over the lifetime of the patient, according to Dr. Ference.
It also systematically overestimated the risk of cardiovascular events among participants with lower LDL-C, blood pressure, or both.
However, after embedding the causal effect of LDL and blood pressure, “the same algorithm was able to precisely predict the risk of cardiovascular events,” Dr. Ference said. The improved accuracy resulted in “nearly superimposable observed and predicted event curves over time.”
Embedded Causal Effects Precisely Predicts Outcomes
Causal AI, embedded into risk analyses, was also able to correct for inaccurate risk benefit derived from short-term clinical trials. These also “systematically underestimate the benefit of lowering LDL, blood pressure, or both,” according to Dr. Ference.
“By contrast, after embedding causal effects of LDL and blood pressure into the algorithm, the same algorithm precisely predicted the benefit of lowering LDL, blood pressure, or both at every age, once again producing superimposable observed and predicted event curves.
In another evaluation conducted by Dr. Ference and coinvestigators, the JBS3 algorithm was applied to several major trials, such as the Heart Protection Trial and HOPE-3. By itself, the JBS3 algorithm predicted less benefit than actually observed.
“After embedding causal effects of LDL and blood pressure, the same algorithm was able to precisely predict the benefit of lowering LDL, blood pressure, or both observed in the trials after 3-5 years,” Dr. Ference reported.
In a sensitivity analysis, the accuracy of the prediction remained largely similar across stratifications by risk factors, such as male sex, presence of diabetes, family history of cardiovascular disease, and other variables. It was also similar across participant age prior to a cardiovascular event and all durations of follow-up.
The data presented by Dr. Ference provides compelling evidence that JBS3, which is widely used in the United Kingdom for risk estimates, does not accurately estimate the risk of cardiovascular disease caused by elevated LDL or SBP. It also fails to estimate the benefit of treating these risk factors.
“Therefore, they cannot be used to determine the optimal timing, intensity, and duration of therapies to prevent cardiovascular events,” Dr. Ference said.
By embedding the causal effects of LDL-C and blood pressure through an AI-based algorithm, the benefit of treatment can be estimated accurately “beginning at any age and lasting for any duration, thus providing the essential information to inform individual treatment decisions about ultimate timing, intensity, and duration,” according to Dr. Ference.
Routine Application Awaits Further Steps
Despite the promise of this concept, there are many steps to be taken before it is introduced into the clinic, asserted designated discussant Folkert Asselbergs, MD, PhD. In addition to testing the accuracy in multiple populations, “we have to do the trials as well,” meaning prospective evaluations to validate the concept is meaningful for improving outcomes.
However, he does not doubt that the concept of causal AI is promising and likely to have a meaningful impact on cardiology after further validation.
“Causal AI is a crucial step that we need to take for more efficient health care,” he said. One reason he expressed caution is that several risk scores enhanced by AI, although not necessarily causal AI, have shown only “modest predictive value” in several studies that he cited.
“Hopefully the data presented from the CAUSAL AI study will really help us take a step up in the discussion to see how we can really benefit by including genetic information in an AI framework to include causality in predicting risk and predicting benefit of treatment,” said Dr. Asselbergs, professor of precision medicine, University of Utrecht (the Netherlands) Medical Center.
Dr. Ference reported financial relationships with more than 15 pharmaceutical companies. Dr. Asselbergs reported no potential conflicts of interest.
This article originally appeared on MDedge.com, part of the Medscape Professional Network.