A committee of the American College of Physicians recently published recommendations for mammogram screening for breast cancer somewhat at odds with the current consensus promulgated by the ACS,ACOG and the USPSTF. They suggested individual risk assessment for women 40-49 rather than the blanket recommendation to get a mammogram every one or two years.Their discussion about individual risk assessment rekindled thoughts about that topic that have been percolating in my brain for some time.
Risk assessment for various medical conditions has become an everyday part of the activities of primary care physicians. Risk assessment involves the identification of something called risk factors, personal characteristics or test findings that are associated with increased incidence of a given disease. As best I can determine, this term was coined by the researchers in the Framingham study when they spoke of factors that were associated with an increased risk of coronary artery disease. As the "practice model" of internist practices changes from hospital based consultation type to office outpatient more attention is given to preventive medicine which is a world of risk factor identification and risk assessment exercises as well as recitation of various guidelines and targets.
In one of my several medical lives, I spent some time doing executive health exams -or "well baby" checkups on senior executives from several companies. Part of that drill was determining their heart attack risk using the Framingham Risk equation.
Here is an example using the equation from the National Cholesterol Panel's (NCEP) web site.A 67 year old non-smoking man, Mr. Jones,with a history of hypertension under control and systolic blood pressure of 120, with a total cholesterol of 170 and an HDL cholesterol of 75 would have a risk estimate of 9 % according to their risk equation.
This means that if we consider the 10 year health outcomes of 100 men from the Framingham data base with this particular set of characteristics, 9 would have a coronary event. (A so-called hard end point of either a myocardial infarction or coronary death.) Of course, we do not know who the 9 will be until the event occurs and we cannot tell Mr. Jones if he will be one of the nine or not.
What does this" risk "of 9% for Mr Jones mean? Maybe the following mind experiment will shed some light on that. Let's pretend we can clone Mr. Jones and we do so 100 times and consider the question of what will be the outcomes of these 100 Joneses. Will 9% have a coronary artery event or will it be the case that either all will be fine or all will have a coronary event. (My gratitude again to Dr. Goodman and his memorable article in the Annals of Internal Medicine for this line of though that I blogged about here.)
If we believe in medical determinism-clinical outcomes determined by a causal chain of events-we believe that either all will be fine or all will have a heart attack. They will all be fine if they and the original Mr. Jones do not posses the factor(s) that sum up or interact to bring on an event or all will have an event if Mr. Jones had-as will all his clones have-whatever factor(s) known and unknown which determine a coronary artery event. If we believe in a cosmic dice roll then some 9% will have an event and medical science will never know ahead of the event who will because it is simply random.
Another consideration is that while we have placed Mr. Jones in this set of men with these particular features , we could as easily-if we had the data available-place him into a different set or as a member of as many sets as the imagination allows. We could consider him as a member of a set defined by his age, his c-reactive protein value, his performance on a stress test, his calcium score on a heart CT scan and his triglyceride level and if we consider the event rate in a group of men with these features we may well arrive a different value which could be 22%. So what is his risk- 9% or 22% or any of the multitude of other numbers that we could construct in a similar manner.
That type of consideration led the imminent German statistical theorist, Richard Von Mises to say in his book " Probability,Statistics and Truth" that it is only possible to speak of probability in terms of a collective (or in more modern terms -a set or a group) and that to say, for example, that a given person has the probability o.10 of dying in the next year is nonsense. Yet, isn't this is exactly what we do when we we punch in a person's numbers into the Framingham equation and announce to the patient that his risk of a hard cardiac event in the next 10 years is 9%.
Could we not "fine tune" the risk assessment process by having a data base such as the Framingham data base but with more factors added on that seem to correlate with cardiac events, for example c-reactive protein, stress test results etc? It would seem so. However, even though Wang et al did that type of analysis the resultant increase in predictive value was not great and the flurry of letters to the editor with various statistical arguments might leave many of us with our heads spinning.The deterministic view would lead us to try and do this fine tuning to more accurately advise the patient on what to do in spite of the fact there will be residual uncertainty even as we add on other risk factor measurements. But we are still not measuring the individual's risk, we may have made the coarse grain data somewhat more fine, but the risk measurement is a conceptually different entity that the answer we obtain we measure some one's blood sugar or P-R interval.
We are still looking at the incidence or risk of an event in a group of patients sharing a certain set of characteristics as the patient of interest and calling the incidence or risk of that event in the group the individual patient's risk.
Probability theory is obviously a sine qua non of medical science and the evidence we use but we when we speak of a patient's probability or risk of developing a given disease as opposed the risk or attack rate of disease in a group we realize how-in Goodman's words- "ambiguous and elusive" that concept is.
So, it is with less that bubbling-over enthusiasm that I react to the ACP recommendation for individual risk assessment in regard to recommendations for mammogram screening, particularly when the risk equations based on the Gail Model seem to be little better than a coin flip according to their own review.
The ACP guidelines say :
"...the Gail model accurately predicts the number of cases of breast cancer that will develop in a given population but is much less accurate in determining which individual women will develop breast cancer"
My contention is that that the panel misunderstands the conceptual difference between the group prediction and the individual patient prediction.It is not just a matter of being "more accurate" in the former case, it is that in the individual case, the meaning of the term "risk" is unclear and that the activity only really makes sense in the case of the population. Goodman's article explains how at one time "chance" and "probability" were different concepts, the former referencing random events and the latter term used to denote a degree of belief applied to situations believe to be unique as in the expression of an opinion as to whether a claim was consistent with the supporting evidence.
When we see Mrs.Jones in the office, it is the case that she will or will not develop breast cancer over her lifetime. It may be that we are fooling ourselves as much as Mrs. Jones when we tell her what her "risk" is.