Key Points:
• Current risk estimating algorithms exclude causal effects and therefore do not accurately estimate baseline cardiovascular risk caused by LDL and systolic blood pressure (SBP)
• Using artificial intelligence to train algorithms on the causal effect of modifiable targets of disease can for the 1st time produce algorithms that accurately predict risk and benefit
• These updated algorithms could be used to prescribe specific actions to alter the trajectory of cardiovascular disease, including identifying the optimal timing, duration and intensity of LDL and SBP lowering.
The role of atherosclerosis as a chronic disease leading to cardiovascular issues including myocardial infarction (MI) and stroke is well established. Both systolic blood pressure (SBP) and low-density lipoprotein (LDL) cholesterol are modifiable risk factors which contribute to atherosclerotic cardiovascular disease. Though atherosclerosis begins early in life, primary prevention through modification of these and other risk factors may be able to prevent the poor clinical outcomes and downstream economic burden caused by atherosclerotic cardiovascular diseases.
This year’s 2022 European Society Congress (ESC) has included a fascinating focus on artificial intelligence (AI) and its projected use in Cardiology. A late breaking trial from Hot Line 6, “Causal AI substantially improves the validity of estimating cardiovascular risk and benefit”, fits in with this theme and teaches us new ways cardiologists may be incorporating AI into practice. Presented by Brian A. Ference, MD (University of Cambridge), the premise of this talk centers around the idea that current risk estimating algorithms which exclude causal effects are unable to accurately estimate baseline cardiovascular risk causes by lower SBP and LDL cholesterol. These models therefore do not afford clinicians the ability to predict when and how actions should be taken to lower a patients SBP and LDL to achieve the most optimal clinical outcome.
Researchers led by Dr. Ference used the Joint British Societies’ (JBS3) risk calculator as a risk estimating algorithm and estimated causal effects of LDL and SBP in discrete time-units of exposure conditional on prior exposure. The machine and deep learning algorithms were then trained on 1.8 million participants that were drawn from participants in randomized trials. To assess the accuracy of the JBS3 risk algorithm both with and without casual effects, the models were tested on an independent sample of more than 40,000 patients each from the UK biobank and from other randomized control trials (RCTS) that focused on modifiable risk factors (namely the Heart Protection Study [HPS], the Systolic Blood Pressure Intervention Trial [SPRINT] and the Heart Outcomes Prevention Evaluation [HOPE]). Primary outcomes were defined as major coronary events which included first occurrence of fatal or non-fatal MI or coronary revascularization. Participants were randomly divided into groups with higher or lower LDL, SBP or both in order to establish accuracy by ensuring the only difference between groups being compared was levels of the modifiable cause of disease which was the target of intervention. Notably there were no significant difference in groups that were randomized.
In order to perform the study, observed event curves from the UK biobank and the RCTs were compared to those predicted by the JBS3 algorithm, both before and after adding causal AI estimated effects of LDL and SBP lowering. This study found that the JBS3 risk estimating algorithm without causal effects systematically underestimated lifetime risk caused by increased LDL; systematically overestimated risk caused by lower LDL; systematically underestimated lifetime benefit of maintaining lower LDL; and systematically underestimated benefit of lowering LDL starting later in life as observed in randomized trials. After adding in causal effects to the AI, the same algorithm now accurately estimated benefit of maintaining lifelong lower LDL, SBP or both at ALL ages with almost identical observed and predicted event curves. The algorithm also accurately estimated both lifelong benefit of lower LDL, SBP or both at all ages and the benefit of lowering LDL, SBP or both starting later in life during every month of follow-up in randomized trials. Results were largely similar across sensitivity analyses, including gender, participants stratified by baseline risk including family history, diabetes and smoking status, and all durations of follow-up.
Dr. Ference concludes from this data that because current algorithms leave out the causal piece, he states, “current risk algorithms appear to suggest that LDL and BP don’t contribute to atherosclerotic disease”. He further notes, “we want to predict risk so we can take action to prescribe action – this has the potential to broaden the notion of how we use risk estimation by focusing on expected benefit”. He points out that any risk factor in the JBS3 algorithm could be used to further define risk, though only LDL and SBP were tested in this study. In discussion he also states that, “using causal AI we can determine how much to lower LDL to overcome the increased risk related to ones’ polygenic risk prediction”. Ultimately this novel evaluation of risk assessment has the potential to accurately estimate baseline cardiovascular risk caused by LDL and SBP. Furthermore, it can tell clinicians and patients the specific benefit of lowering LDL, BP or both beginning at any age and extending for any duration. In doing so, it allows for a more accurate risk assessment that can be used to individualize treatment decisions about the optimal timing, duration and intensity of risk factor modification.
Written By:
Jamie Diamond, MD, MPH