Friday, July 18, 2008

The 2004 national cholesterol guidelines have no clothes

A pending board exam has forced me to review the 2003 NHLBI, ATP III Lipid management guidelines. I don't see patients, so I haven't had to really contemplate these before.

They're not a pretty sight.

Ok, so they're not quite as bad as a naked middle-aged emperor, but they still hurt the eyes.

The problem is they try to reconcile two different risk models. One risk model attempts to stratify people based on their similarity to a large population study - the Framingham model.

Another risk model is based on different research data sets, and tries to estimate risk based on a changing set of predictive "risk factors", such as Diabetes mellitus, and family history of heart disease.

Problem is, those two latter two big risk factors weren't a part of the Framingham model. In fact the Framingham model doesn't incorporate LDL cholesterol directly, it estimates it from Total and HDL cholesterol.

The two models look like this (table stolen from my obsolete online medical notes, this part was updated):

Item Risk calculation model Risk factor approach
Age x x
Gender x x
Total Cholesterol x  
HDL Cholesterol x x
LDL Cholesterol   x
Smoker x x
Hypertension x x
Family history   x
Diabetes   x

The guideline writers try to glue the two models together in a way that seems logical, but they really don't work that well. For example (LDL level in this table is the level where the statins start).

LDL Level Risk Factor Framingham 10 yr risk
> 100 CHD or "equivalent"* > 20%
> 130 2 + (ex. 46 yo male smoker) 10 - 20%
> 160 2 + (ex. 46 yo male smoker) < 10%
> 190 Treat based on LDL alone.

I played around with the online calculator, it wasn't hard to create a plausible patient with a Framingham risk of < 10% but a Risk Factor Model if CHD equivalent (basically a healthy diabetic patient, the right answer is clinical judgment with a bias towards treating if either of the risk models meet criteria. So treat if column A + either (B or C).

We really need a single integrated model of risk, not trying to juggle and compare two different models that can give contradictory answers.

Of course it may turn out that this single integrated model doesn't lend itself to memorization, but needs to be implemented as an electronic tool. Wouldn't be the first time that's happened.

No comments: