Ever since researchers with the legendary Framingham Heart Study created the first calculator to gauge the chances of having a heart attack, such tools have become a routine part of medicine. But the results aren’t as straightforward to interpret as the answers you used to get from your old high-school graphing calculator. The problem has to do with the challenge of interpreting the concept of risk.
Let’s use as an illustration the heart risk calculator designed by the American College of Cardiology and the American Heart Association to accompany their latest guidelines for the use of cholesterol-lowering statins.
It asks for 10 pieces of information, from age to cholesterol levels and smoking status, in order to estimate the chance of having a first “atherosclerotic cardiovascular disease event,” better known as heart attack or the most common type of stroke. Add your information and hit the “Calculate” button. The tool returns a number that represents your chance of having an event over the next 10 years. But what does this really mean? Experts disagree about the best interpretation.
article continues after advertisement
Say the number you get is 10 percent. One way to interpret it is like this: In a group of 100 people with the same risk factors as you, 10 will have an event over the course of the next decade. Experts call this the epidemiologic risk. While it can be helpful in planning an treatment and prevention efforts across an entire population, it probably doesn’t fully answer your questions about yourself. Am I going to be one of the 10 with an event or the 90 without an event? What about my other risk factors not included in the calculator? For example, you may have had one or two parents who died young from a heart attack or stroke, which would increase your risk, or you may exercise every day and be at the highest level of cardiorespiratory fitness, which would lower it.
Ideally, instead of knowing the risk of having heart attack in the next 10 years, we would rather know the definitive answer — am I going to have one or not?
Individual factors like smoking or high blood pressure or diet quality have been linked with disease risks for decades. Using statistical techniques, it is possible to capture the prognostic power of such factors into a risk estimate. If one indicator sketches an individual’s portrait, many of them working together (10 in the case of the heart risk calculator) carve a statue of the same subject, adding dimensions to the prediction.
As my colleagues Ralph B. D’Agostino, Allan D. Sniderman, and I wrote earlier this year in JAMA, if all past, present, and future predictors of a particular disease were known — and it was possible to quantify them — one could build an algorithm that would give a definitive answer about that disease occurring for each individual.
Until then, it’s important to make the best of the useful tools that Framingham and other reputable sources have created for us. To make that happen, physicians and other experts need to convey the result that emerges from a risk calculator in ways that people can easily grasp.
The Framingham Heart Study, for example, includes in one of its calculators a “heart/vascular age” in addition to the 10-year risk of cardiovascular disease. A heart/vascular age younger than your chronological age is good, one older than your chronological age isn’t.
An approach we are exploring at the Duke Clinical Research Institute capitalizes on human’s natural tendency to compare themselves with others. Telling an individual that she has a 15 percent chance of having a heart attack in the next 10 years may offer some motivation to adopt healthier habits, while telling her that 90 percent of women her age have a better risk profile than hers may be even more motivational.
The risk estimates offered by today’s models and algorithms are limited by the data on which they were developed and the information they include. With more and more data becoming available, predictive analytics become more powerful and more useful to precision medicine, which aims to tailor treatment to individual patients. Electronic health records may help us build calculators that can provide estimates in a doctor’s office or at a patient’s bedside.
It is important to keep in mind that health risk calculators don’t assess the benefit that may come from treatment. One that inadequately addresses the root causes of a disease may do little to bring down the risk, while an effective therapy may offer an incremental or long-term benefit among individuals at moderate or even low risk.
The quality of the data that go into building a risk calculator matters. Not every calculator can be trusted. Turning up high in a Google search doesn’t always mean that a calculator’s validity and performance are trustworthy. A few, like those produced by the Framingham Heart Study, have been thoroughly assessed and validated.
Risk calculators will likely become more common in everyday clinical settings, partially because more data are available to build them and partially because the medical community and the public find them to be useful. Innovations such as machine learning applied to the “big data” available in electronic health records and other sources will only serve to improve the performance of these calculators and increase their value.
But even with advances such as machine learning, we are unlikely to ever create a calculator that moves us from prognostication to certainty and delivers the correct answer for each individual. That shouldn’t dissuade us from using health risk estimates. But we cannot let the appeal of a number that appears easy to understand pass for real understanding of what that number means. If we settle for that, what will be at risk is the potential benefit of these calculators for patient health.
Michael J. Pencina, PhD, is director of biostatistics at the Duke Clinical Research Institute and professor of biostatistics and bioinformatics. He has worked on a number of large studies, including the Framingham Heart Study.