Two New Scores Help Hone 10-Year CV Death Risk in ASCVD Patients

Two newly developed tools for predicting 10-year risk of cardiovascular mortality in patients with established atherosclerotic cardiovascular disease (ASCVD) outperform a model commonly used to assess risk in secondary prevention, according a study published recently in JACC: Advances.

Equipped with the patient’s 10-year risk of CV death, clinicians are in a better position to set expectations and tailor care for those with ASCVD, said Samia Mora, MD (Mass General Brigham, Boston, MA), one of the paper’s two senior authors along with Olga V. Demler, PhD (Mass General Brigham and ETH Zurich, Switzerland).

In secondary prevention, “it’s not like everybody has the same risk,” Mora told TCTMD. “There’s a big variability in terms of risk for future events, whether that’s cardiovascular death or some other nonfatal event.”

For individuals at higher risk, it makes sense to more aggressively address risk factors and encourage optimal medical therapy, she noted. On the other hand, if the patient is at lower risk, there’s less need for intensive therapy, which can carry side effects.

Clinical practice guidelines are risk based, added Demler. “Accurate risk assessment is therefore fundamental to ensuring that patients receive appropriate preventive and therapeutic interventions.”

Whereas risk-prediction models in primary prevention “have a long and well-established history,” with continued evolution, the area of secondary prevention “has not enjoyed similarly rigorous model development,” Demler commented via email.

Importantly, the tools developed for primary prevention can’t work in this space, given how much “a prior cardiovascular event profoundly alters a patient’s risk profile,” she added. “Despite recent progress in machine learning, current secondary-prevention guidelines still rely on simple categorical rules or a table that assign patients to broad groups (“average,” “high,” or “very high” risk) rather than estimating continuous risk probabilities.”

UK Biobank and Mass General Brigham Datasets

Led by first author Olga Mineeva, PhD (ETH Zurich and the Max Planck Institute for Intelligent Systems, Tuebingen, Germany), the researchers developed and validated their models using data from 32,994 UK Biobank participants (mean age 61 years; 35.5% women) who enrolled between 2006 and 2010 and were followed through 2021. Additionally, they externally validated these models in a Mass General Brigham cohort of 54,969 patients (mean age 71 years; 41.4% women) who enrolled in 2007 and were followed through 2018. All had established ASCVD.

Two residual risk scores were created, with factors selected algorithmically using artificial intelligence:

• RRS16, based on 16 routinely available factors
• RRS24, based on 24 routinely available factors as well as patients’ self-reported health (ie, “In general, how would you rate your overall health?”) and additional biomarkers

Mineeva and colleagues compared these scores against the metric established by the 2018 cholesterol guidelines from the American Heart Association (AHA), the American College of Cardiology, and numerous other professional societies, which mainly focused on primary prevention. For secondary prevention, the document advises dividing patients into three risk categories (average, high, and very high risk) based on 10 factors—notably, it doesn’t provide personalized estimates of absolute risk for individual patients or take sex into account.

As a cardiologist, “if I have a patient in front of me, it’s really hard to think of the AHA algorithm, because you have to go and try to put them into [one] of these risk buckets,” said Mora. “That’s hard to do because it’s based on all these clinical variables that then you have to ask about. That takes time.” In contrast, their approach aims to leverage information commonly kept in electronic health records (EHRs) and automate the calculations.

[In secondary prevention], there’s a big variability in terms of risk for future events, whether that’s cardiovascular death or some other nonfatal event. Samia Mora

The researchers found that the simpler RRS16, which Mora said “could be basically applied from almost any EHR,” had C-statistics of 0.752 in the UK Biobank cohort and 0.750 in the Mass General Brigham cohort. It performed better than the 2018 guideline model, which resulted in C-statistics of 0.658 and 0.580, respectively.

The more comprehensive RRS24, on the other hand, achieved a C-statistic of 0.784 in the UK Biobank. This score could not be externally validated in the Mass General Brigham cohort, however, because it contained variables not available in the healthcare system’s electronic health records.

Both scores were well calibrated, with P values greater than 0.1.

With RRS16, the factors most tightly associated with higher risk were smoking (HR 2.21; 95% CI 1.98-2.47), history of congestive heart failure (HR 2.16; 95% CI 1.88-2.48), male sex (HR 1.70; 95% CI 1.51-1.91), history of atrial fibrillation (HR 1.61; 95% CI 1.42-2.84), and history of diabetes (HR 1.61; 95% CI 1.43-1.81). Albumin, urea, and C-reactive proteins also were predictive.

With RRS24, the strongest indicator of CV mortality was self-reported “poor” health (HR 1.90; 95% CI 1.67-2.17). Interestingly, vitamin D levels also were closely tied to risk—Mora said that their work, in this way, could help “guide us as to what factors may be related to risk that a clinician wouldn’t necessarily think about firsthand, especially for a patient with cardiovascular disease.”

The new scores’ “improved performance can be attributed to the inclusion of a broader range of clinical variables and advanced modeling techniques, including machine learning,” the investigators conclude.

They add: “Our findings also suggest that certain factors not currently used in ASCVD risk models for primary prevention, such as self-reported overall health ratings and specific biomarkers (ie, albumin, CRP, glycated hemoglobin, urea, and vitamin D), although being readily available, play an important role in secondary risk estimation.”

Accurate risk assessment is . . . fundamental to ensuring that patients receive appropriate preventive and therapeutic interventions. Olga V. Demler

Currently, the residual risk scores are intended for research settings, the investigators specify. That said, the externally validated RRS16 is available free online.

“While this tool is not ready yet for clinical decision-making,” said Demler, “we hope it will stimulate further evaluation and ultimately contribute to updating current clinical guidelines with a modern, validated, and data-driven risk model, ultimately improving care for patients most vulnerable to recurrent cardiovascular events.”

“We hope that people will find it useful,” Mora said. “The score was derived and validated in two large contemporary populations from the US and the UK. We tried to make it readily applicable by providing an online calculator, but future studies should also evaluate the accuracy of the risk scores in other more diverse populations.”

Looking ahead, ideally studies will be done to look at implementation and how the residual risk scores might improve medication adherence or clinical outcomes, she suggested. Other research could explore their performance in different geographic regions and in datasets capturing various degrees of risk.