Researchers employed cutting-edge machine learning techniques to enhance the precision of a national cardiovascular risk assessment tool, while maintaining its clarity and original risk correlations.
Risk assessment tools play a vital role in determining disease risk for millions of individuals, so their accuracy is essential. However, when these national models are adjusted for specific local populations, their effectiveness often declines, leading to a loss of accuracy and clarity. A team from Brigham and Women’s Hospital, part of the Mass General Brigham healthcare system, used sophisticated machine learning approaches to improve the accuracy of a national cardiovascular risk calculator while ensuring it remained interpretable and retained the original risk connections. Their findings revealed that, within an electronic health records group from Mass General Brigham, accuracy was enhanced overall, with approximately one in ten patients being placed into a different risk category to allow for more tailored treatment decisions. These findings are detailed in JAMA Cardiology.
“Risk calculators are extremely important as they form a key component of discussions between healthcare providers and patients concerning risk prevention,” stated the lead author Aniket Zinzuwadia, MD, a resident physician in Internal Medicine at Brigham and Women’s Hospital. “Nevertheless, when applying these global calculators to local communities, there can be variations due to the distinct characteristics of the area—such as different demographic factors, physician practices, or risk elements—so we aimed to customize the base cardiovascular disease risk model for local populations in a manner that enhances current methodologies.”
The American Heart Association introduced the Predicting Risk of Cardiovascular Disease Events (PREVENT) calculator in 2023 for adults aged 30-79. This updated tool assists in forecasting the chances of an individual experiencing a heart attack, stroke, or heart failure within 10 and 30 years. Although the PREVENT equations have performed well on a national scale, the researchers sought to determine if their method could better align risk assessments with local populations.
In the analysis, the researchers utilized electronic health record data from 95,326 Mass General Brigham patients aged 55 and above in 2007, who had at least one blood lipid or blood pressure measurement recorded between 1997 and 2006 and had interacted with the hospital system at least once between 2007 and 2016. The team employed XGBoost, a public machine learning framework, to adjust the equations of PREVENT while preserving the established relationships between known risk factors and outcomes from the original model. The results indicated improved accuracy and re-identified one out of ten patients in this group.
“This could potentially indicate a set of patients who, under the original model application, might not have been prescribed statins, yet could have benefitted from such treatment,” noted Zinzuwadia.
While further steps are necessary before this technique can be effectively integrated into patient care, the team aspires to see how it performs in other healthcare systems’ local populations and eventually empower clinicians and researchers to customize global risk models.
“A significant challenge in applying AI in medical research is to ensure that machine learning models not only exhibit flexibility but also are transparent, dependable, and anchored in domain knowledge,” explained co-senior author Olga Demler, PhD, an associate biostatistician at Brigham and Women’s Hospital’s Division of Preventive Medicine. “Our methodology illustrates that we can bypass the ‘black box’ aspect of AI applications and may pave the way for advanced algorithms that maintain their adaptability while providing assurances of their performance.”
Authorship: Other authors involved in the study include Olga Mineeva, Chunying Li, Zareen Farukhi, Franco Giulianini, Brian Cade, Lin Chen, Elizabeth Karlson, Nina Paynter, and Samia Mora.
Disclosures: Samia Mora has acted as a consultant for Pfizer on matters unrelated to the current research. Olga Demler and Nina Paynter have secured funding from Kowa Research Institute for unrelated work. Aniket Zinzuwadia has worked as an employee of Heartbeat Health on projects unrelated to this study.
Funding: The research efforts were backed by the National Heart, Lung, and Blood Institute (K24 HL136852, R21 HL156174, R21HL167173, K01HL135342, and R21125962), the American Heart Association (17IGMV33860009), the Swiss Federal Institute of Technology (ETH, Zurich, Switzerland), and Dataspectrum4CVD from the Swiss Data Science Center/Personalized Health & Related Technologies, Zurich, Switzerland, along with support from the National Human Genome Research Institute (U01HG008685).
