Why Cardiac Biomarkers Don't Help Predict Heart Disease

Perry Wilson delves into the reasons that adding cardiac biomarkers to conventional risk factors didn't improve the ability to predict cardiovascular disease.
https://www.staging.medscape.com/view...

-TRANSCRIPT-
Welcome to Impact Factor, your weekly dose of commentary on a new medical study. I'm Dr F. Perry Wilson of the Yale School of Medicine.

It's the counterintuitive stuff in epidemiology that always really interests me. One intuition many of us have is that if a risk factor is significantly associated with an outcome, knowledge of that risk factor would help to predict that outcome. Makes sense. Feels right.

But it's not right. Not always.
Here's a fake example to illustrate my point. Let's say we have 10,000 individuals who we follow for 10 years and 2000 of them die. (It's been a rough decade.) At baseline, I measured a novel biomarker, the Perry Factor, in everyone. To keep it simple, the Perry Factor has only two values: 0 or 1.

I then do a standard associational analysis and find that individuals who are positive for the Perry Factor have a 40-fold higher odds of death than those who are negative for it. I am beginning to reconsider ascribing my good name to this biomarker. This is a highly statistically significant result — a P value .001.

Clearly, knowledge of the Perry Factor should help me predict who will die in the cohort. I evaluate predictive power using a metric called the area under the receiver operating characteristic curve (AUC, referred to as the C-statistic in time-to-event studies). It tells you, given two people — one who dies and one who doesn't — how frequently you "pick" the right person, given the knowledge of their Perry Factor.

A C-statistic of 0.5, or 50%, would mean the Perry Factor gives you no better results than a coin flip; it's chance. A C-statistic of 1 is perfect prediction. So, what will the C-statistic be, given the incredibly strong association of the Perry Factor with outcomes? 0.9? 0.95?

0.5024. Almost useless.

photo of Perry Wilson Impact Factor-Cardiac Biomarkers
Let's figure out why strength of association and usefulness for prediction are not always the same thing.

I constructed my fake Perry Factor dataset quite carefully to illustrate this point. Let me show you what happened. What you see here is a breakdown of the patients in my fake study. You can see that just 11 of them were Perry Factor positive, but 10 of those 11 ended up dying.

photo of Perry Wilson Impact Factor-Cardiac Biomarkers
That's quite unlikely by chance alone. It really does appear that if you have Perry Factor, your risk for death is much higher. But the reason that Perry Factor is a bad predictor is because it is so rare in the population. Sure, you can use it to correctly predict the outcome of 10 of the 11 people who have it, but the vast majority of people don't have Perry Factor. It's useless to distinguish who will die vs who will live in that population.

Why have I spent so much time trying to reverse our intuition that strength of association and strength of predictive power must be related? Because it helps to explain this paper, "Prognostic Value of Cardiovascular Biomarkers in the Population," appearing this week in JAMA, which is a very nice piece of work trying to help us better predict cardiovascular disease.

I don't need to tell you that cardiovascular disease is the number-one killer in this country and most of the world. I don't need to tell you that we have really good preventive therapies and lifestyle interventions that can reduce the risk. But it would be nice to know in whom, specifically, we should use those interventions.

Cardiovascular risk scores, to date, are pretty simple. The most common one in use in the United States, the pooled cohort risk equation, has nine variables, two of which require a cholesterol panel and one a blood pressure test. It's easy and it's pretty accurate.

photo of Perry Wilson Impact Factor-Cardiac Biomarkers
Using the score from the pooled cohort risk calculator, you get a C-statistic as high as 0.82 when applied to Black women, a low of 0.71 when applied to Black men. Non-Black individuals are in the middle. Not bad. But, clearly, not perfect.

And aren't we in the era of big data, the era of personalized medicine? We have dozens, maybe hundreds, of quantifiable biomarkers that are associated with subsequent heart disease. Surely, by adding these biomarkers into the risk equation, we can improve prediction. Right?

Transcript in its entirety can be found by clicking here: https://www.staging.medscape.com/view...