president of the american statistical association and he is an elected fellow in the american association for the advancement of science, american society for clinical investigation and the institute of medicine. he received many awards
including the public health service distinguished service medal and the nathan mantel lifetime ark achievement reward for epidemiology. in 1989, he described a statistical model that now bears his name, the gail model was the first (p'cer risk prediction
model applicable to the generp& population. it has been used very widely to council and educate women about their risk for breast cancer. the gail model provided theyã¡s methodological foundation for the development of subsequent models that address a variety of
other cancers as well as for various populations including minority populations. dr. gail is going to talk about his model for risk breast cancer and discuss both the implications and the potential for incorporating additional risk factors to improve
performance. please join me in welcoming the 2013 robert s gordon, jr. lecture award recipient, dr. mitchell gail. his presentation is entitled, using risk models for risk models #or breast cancer prevention.
dr. gail. [ applause ] >> thank you for the introduction and thank you for coming. it's not easy to get in here sometimes and certainly not easy to park. i really appreciate this
opportunity to speak today, especially because i had the privilege of knowing dr. gordon. dr. gordon was instrumental in establishing the society for clinical trials and in 1982 he was the president of the society for clinical tã±ials. in those days, i was a young
clinical trialist and there was no textbook for how to set up a trial, how to conduct a trial and analyze a trial and so it was very valuable to me to be able to go to the society meetings and learn from people who knew something about it. so it's a real pressure to give
this lecture in his onor and i'd like to acknowledge the intramural research program of epidemiology and genetics forr continuous support of this research and also for providing many collaborative opportunities that contributed to this work. so, i'll be talking today about
using risk models for cancer i'll talk about what i'm talking about, that is absolute risk, what is it? i'll mention in revrqjt the risk factors for breast cancer and then i'll talk about models that have been devel for projecting absolute risk and
their applications either in counseling individual patients, or in cancer prevention in the general population. and one reoccurring theme will be how much to snps add in these applications. now, here is a typical woman who might be coming for counseling.
age 40. she began menstruating at age 14, which is what i call a baseline level of risk. the lowest level of risk factor. she is nulliparous which increases hear risk, she had no biopsies so that is a baseline or lowest level of risk but her
mother had breast cancer. that increases hear risk. her risk. we could ask what is her risk relative to someone else, relative to a 40 yearly woman all of whose risk factors were at the baits line or lowest levels.
it is useful in deciding which risk factors which are important but it doesn't tell well woman what she needs to know about placing her risk in context. here is the very same woman but we asked a slightly different question or really a very different question.
we asked, what is the chance that this woman, she has all the same risk factors, what is the chance it woman will be diagnosed with breast cancer by age 70, over a 30 year period? according to this calculation, it is 11.6%. or .116.
and this is a number that is not relative to someone else. it is her number and she can use it to place that risk in context and begin to plan decisions about her life. on that basis. and i'd like to just give a little more background about
what the absolute risk we just quoted is. suppose a woman is well at age a and she has a bunch of risk factors like the ones on the previous slide, x. and suppose she wants to know what is the chance she'll be diagnosed with breast cancer in
a subsequent time period to age a plus t. a couple of things can happen. she might die of something else before she gets breast cancer. or, she might be diagnosed with breast cancer during that interval. an absolute risk is the
probability she will be that interval given that she is well at anal a and has risk factors, x. now, one thi'g that can decrees this absolute risk is an increased chance of dying of something else. so a woman at high risk of dying
of something else will have a lower absolute risk of breast cancer. certain things can increase the chance of being diagnosed with breast cancer. for examp&e, if the interval is long, there is a largq) chance of being diagnosed with breast
cancq). if she has a lot of adverse risk factors, there is a increased breast cancer and it tends to be the case that the older the woman is, the higher her chance of getting breast cancer is because the breast cancer cancer incidence rate increases very
rapidly with age. so that is the idea of absolute risk. and i'll talk about models that are used to estimate absolute risk but before i do, i'd like to remind you of some of the risk factors for breast cancer. and he are some i call strong
risk factors. age itself is very strong. 70-74-year-old woman has 56 times the risk of a 25-29-year-old woman. so age really has to be built in to any model. if you have a major mutation in a brca1 or brca2 gene, if you're
young, it can be maybe a 20-fold increase risk. if you're older, maybe a 2-4 increased relative risk. if you have been radiated for hodgkin lymphoma, let's say, in youth and received 40 gray, you 6-fold relative risk.havq if you had a cont lateral breast
cancer, that really increases your risk. being born in a western country is a major risk factor. compared to rural china, the risk is 5. mammographic density greater than 45% compared to less than 5%, relative risk of 4,
approximately. ã±ãºso these are some major risk factors. and there are also some moderately strong risk factors. a family history of moderately strong. if you have two or more affected first degree relatives depending
on the study, you can get relative risks 2-5. 1 is maybe 1.4 to 3. having a biopsy is a moderately strong risk risk factor if youhave atypical piperplasia. hormone replacemeã±t therapy for a substantial duration is a moderately strong risk factor
and a delayed age of first live birth a moderately strong risk factor. and then what i call weak risk factors consistently demonstrated but weak. age at menarchy. if you have an early age of menarchy, that increases your
delayed age of menopause increases your risk. bmi greater than 30 if you're postmenopausal increases your risk somewhat but if you're pre-men asal it decreases your drinkingethinal increases risks slightly and here is one. this is a snp, a single
nuclq cyberblast growth factor receptor 2. this is the strongest one found for breast cancer and one adverse allele increases risk by relative risk of 1.26. you have choices to make when you develop risk models.
one choice is, it primarily based on genetic theory or shouldvaã· you take amore empirical approach? and we'll see examples of those two approaches. of course, deciding what factors to go into the model is an important element.
should you have a very detailed family history or only some family history, how much of a reproductive history should you use? what about some of the medical history factors, not necessarily etiologic biprognostic like biopsies mammographic density?
in looking at the literature, it's important to keep in mind the data sources we are used to build to models and how therapiesed together. they have a single cohort and you can develop it that way but many models are pullr'g information from a variety of
sources. it's important also in looking at the various models to think about the target population that the investigators had in mind. it was a general population like the uk or the u.s.? because the rates from the u.k. or the u.s. might be embedded in
those models. or was the data primarily derived from a high-risk clinic? so the usefulness of these models will depend on all these here is some examples of genetically based models. some of the first models were based on the assumption that
breast cancer is an autosomal dominant disease. widely used, brca p)o, allows you to use extensive family history based on this theory and also to incorporate mutation data if you have it, to project and lyses beth developed a similar model before brca had
been cloned. the brca genes had been cloned. but elizabeth did a study herself that showed that really the familial aggregation patterns of breast cancer are not explained entirely by autosomal dominance. we know that the brca genes only
account for perhaps 20% of the familial aggregation. so, others have tried to expand on the autosomal dominant model by allowing for residual familial correlations above and beyond those ascribed to the dominant mutations. a nice model along this line is
boadicea and the other is the ibis model. these references describe how this residual familial (j account. one comment 6789 these brca pro now although not originally and the model is a model of absolute
risk because they account fo) the possibility of dying of something else, of something else before breast cancer. these two models are models of pure accumulative risk but they are larger than would be given if you accounted for the possibility of dying of some
other cause. now, this nci's breast cancer called the gail model, is the model that i was describing on the first couple of slides and just a little bit of history here. i started working in this area because i went to lunch with
john hill in dceg many years ago in the mid 1980s, and he was coupling women about their breast cancer risk and action that is they might be inclined to take and found that many women had considerably overestimated their risk of breast cancer and he wanted a
more objective estimate of the breast cancer risk for these women. they were probably thinking autosomeinal dominant. my sister has breast cancer therefore i have a 50% chance of having the gene. therefore i better have a
mastectomy. and so, that was kind of a background for trying to be a little bit more data-driven in developing a model. and we used data from the breast cancer detection demonstration project, a number of people in the division including louise,
kathy, and other people david and sylvia green, many people participated on this project. we derived the relative risk from the best chancer project and then reconstructed the bcddp cohort data to get information on incidence. and put it all together.
later, the incidence data from the bcddp was replaced by sear data and that is what is called the breast cancer risk assessment tool relative risk from the bcddp but sear incidence data incorporated to turn these introrisk factors that i had mentioned
so, that was that model. and i'll be referring to this model,bcrat in much of the talk. but no matter which model you come up with, it's important to try to validate the model before using it widely. and to validate a model, you should have independent cohort
data because you need air cohort to estimate absolute risk and to test absolute risk. and i'll talk about calibration basically does the model correctly predict the number of cancers that develop in the cohort over time? and in subgroups of the cohort?
later i'll talk about discriminatory accuracy but i'll concentrate on calibration right now because several of the applications that i'm about to talk about depend mainly on calibration, not on discriminatory accuracy. and here was one of the first
tests of the model. this model was used, bcrat was used to design the breast cancer prevention trial and joe constantineo and coworkers used the control arm of that trial to see whether the predictions of numbers of cancers that later occurred in the trial were in
agreement with what was expected based on the model. and consider these 2332 women who were aged 49 or less. for each woman a risk of calculated and all of those risks were summed up and they added up to 55.9. so that was the expected number
of cancers we anticipated in that group. actually 60 were observed during the course of the trial. so e/o was .9 and likewise it was 1.1 and all close to 1, indicating good calibration and over the entire population, the. /o ratio was 1.0.
and there were a number of other tests of the calibration of this model and it tends to do well. there was a recent paper in 2010 by sarah and coworkers showing that there was a period before 2004 where the model somewhat underpredicted. so it's important to keep
monitoring the models, monitoring national trends in order to make sure that the model stays in calibration. but it seems to be a well-calibrateed model at the present time. now, one thing you can do with a well-calibrated model, is try to
give a general perspective on risk when you're advising a woman about her options for disease prevention. and you can also use a well-calibrated model for a more formal weighing of risk and benefits and i'll try to describe that in the next few
slides. so a very contentious issue that came up not long ago. should women in their 40s have screening maaingography. the u.s. prevention service task force recommended against routine screening for women age 40-49.
they said no factors except age and deleterious mutations conveys a clinically important absolute increased risk for and the decision instead should be made based on patient context, including the patient's values regarding specific benefits and harms.
i agree with much of this statement except this part. because between age 40 and 50, the risk of breast cancer doubles. but flower many combinations of risk factors that more than out weigh a doubling of risk from the age difference between
40-50. so i suppose i have a 40-year-old woman uncertain about having scening mammograms but her mother and sister have and that is greater than that of a 50-year-old woman without the lowest level is .6%. according to the guidelines, the
50-year-old woman should get a mammography. and if you carry the argument further, there is every reason to believe that the 40-year-old woman is going to benefit more than a 50-year-old woman who has no risk factors. so, the point is that knowing
your absolute risk can give you a certain context and ability to think about problems, including perhaps whether you should have a mammogram. is this an unusual circumstance? well, lindsey woo and barry were working with me estimated that the proportion of women in their
40s who have breast cancer risks as high or greater than 50-year-old woman without risk let's look at the dark lines here, which are corresponding to non-hispanic white women. if that woman is 49 years old, over 90% of such women have a risk greater than that of the
50-year-old without risk and on average, 74% of non-hispanic white women have risks greater than the 50-year-old. of non-hispanic black women, the average is 31%. so there are actually between the 11 and 12 million women with
risks higher than a baseline risk 50-year-old woman and they might well consider taking mammograms on the same logic that says that the risk benefit ratio for mammography is good for the 50-year-old woman. that was the kind of informal weighing of management style
based on just know what your absolute risk is. can we make a more formal argument forp and this is work that was done on the basis of findings from the breast cans are prevention trial. 0ã·s )'fd73v(w
i'd like to show the weighing of risks and benefits depends on absolute risks. because absolute risk of each end point, each of the things effected by the intervention, is the common currency that can be used to equate the various outcomes and come to some
decision. so, basically, fishyet all published the key report on tamoxifen against placebo and it showed that it reduced the risk of invasive breast cancer by nearly half, reduced hip fractures by nearly half but it increased the risk of
endometrial cancer especially in older women. it increased the risk of stroke by 60%. it increased pulmonary embolism threefold. and it had effect on -- these were called life-threatening events and we called these
severe events. again, it reduced the risk of in situ breast cancer by 50% but increased the risk of deep vain thrombosis by 60%. so how can a woman decide whether or not to take tamoxifen? let's consider a 40-year-old
woman with a routeerous who has projected breast cancer risk of 2%. we ask her to consider what is going to happen to a population of 10,000 women just like her and in the absence of tamoxifen, l would happen ifand also what7 tamoxifen were given to that
now two had been of them, bauds of this -- 200 of them, we expected to develop invasive breast cancer, two hip fractures and 7 pulmonary embolism. that's if we give no tamoxifen and these are life-threatening events. if we give tamoxifen, we will
prevent nearly half of those. we'll prevent 97 of those 200 one of lh we will also cause 16 additional endometrial cancers. so instead of doing 10, it will be 26 endometrial cancers and we'll cause 13 additional strokes and 15 additional
pulmonary em limps. like wide, we can see what would happen for the severe events preventing 53 in situ breast cancers but cause 15 additional deep vein thrombosis. and some people think that all you should do is present data like this to the woman and let
her look at the various types of outcomes and come too her own conclusion about what to do. it might be helpful to sum she's up and find out there is a net 54 life-threatening events prevented and a net 38 severe events prevented in this 40-year-old woman in a similar
so it looks like it somebody something for her to strongly considered. if you wanted to make one number, you have to weight these differently and suppose we give a weight of one to life-threatening and a wait of a half to severe.
then 54 plus half of 38 is 73 life-threatening equivalent events prevented by taking tamoxifen t seems like she might want to consider that. if you're willing to adopt an index like that, there is that 73. you find that for young, white
women, there is a benefit of taking tamoxifen if their invasive breast cancer risk is 2%, 4% or 6%. the benefit increases with increasing breast cancer risk because you're preventing more breast cancer without en curbing additional adverse events.
notice what happens to 50-59-year-old white women. if the risk is 2%, you actually have a negative benefit because strokes and endometrial cancers are becoming more dominant. so, only the very high risk white women in her 50s should consider taking tamoxifen.
and you get a similar pattern for black women but the benefits are a little bit less because the background stroke rates are higher. so some conclusions for tamoxifen, young women at high risk stand to benefit the most. women without a uterus have a
more favorable risk benefit ratio. i didn't show that but i will in a minute. and there is no single risk level like 1.67% that applies to all women. the decision depends on the woman's age and her risk of
these other outcomes as well as her risk of blast cancer. so, you have to weigh all of these possible outcomes in deciding what is good for the woman. here is an extension of this work by andy friedman and colleagues in 2011 to older
women and using some slight changes in the background data updated information on the things that went into the same sort of calculation. the sort of yellow pattern here corresponds to moderately strong evidence for a benefit. the blue to very strong evidence
for a benefit and you can see for tamoxifen, only women with very high-five-year risks, women with a uterus, stand to benefit. the gray means they don't benefit. it's a negative. whereas forra locks fen, not only do you get a wider scope --
raloxifene -- young women in their 50s benefit but the benefits extend to older age groups too. and this is primarily becausera knocks 15 doesn't have adverse effects -- raloxifene -- on endometrial cancer risks or very well --
little. what if the woman doesn't have aute are us? well, then the risk benefit profiles for raloxifene and tamoxifen look similar. so there can be a role for tamoxifen especially for the woman who has a hysterectomy.
i have been trying to put your risk in context and making an actual treatment or intervention decision for the individual based on weighing her risks and benefits. now i'd like to try to see what the implications of this kind of thinking are at the
population-level. and there are four important applications of risk models that i'd like to consider. one is designing prevention trials. one is assessing population absolute risk reduction from prevention strategies and two
others, and these two don't require a very diminating model but these last two points, the high-risk strategy for interventions that have adverse side effects, and allocating prevention resources under cost constraints, they can benefit from this model and i'll explain
whatothat means in a little while. but before i get to designing prevention trials, i want to say a word about 㧠assessing population absolute risk reduction from prevention strategies. because i'm not going to present
slides on that. suppose your risk model modifiable risk factors such as alcohol consumption or level of exercise or body mass index? you could ask, what is going to happen to the absolute risk in the population if you make favorable changes to those risk
factor distributions? if you put everybody at the lowest level of risk factors for each of those? and i think this is worth very worth worth doing because you often hear the word, attributable risk, used for calculations of this type.
well and attribute age risk is valuable but it tells you what percentage of risk reduction there will be. this tells you how much will the absolute risk be reduced? you can talk about a 20% attributable risk reduction in risk by 20% that only amounts to
an absolute risk reduction of one or 2%. so it's important to take this perspective in addition to the attributable risk perspective when you are looking about the potential value of prevention elizabeth, who is working with in our group did this using
italian data published in jnci. let me turn now to these other three points. designing prevention trials first. the power of prevention trial to demonstrate that the intervention is really preventing disease, depends
primarily on the number of events observed. the number of breast cancers if you're talking about tamoxifen trial, let's say. and the number of events observed proportional to the average risk of trial participants.
so, using these absolute risk models, we are able to figure out how many events are likely to occur over the course of the trial, see how many women we need to study and how long we need to study them in order to have a well designed powerful study.
the absolute risk idea also has an impact on the eligibility criteria. the people designing the breast cancer prevention trial didn't want to include women in the study unless their breast cancer risk of high enough that they could probably benefit from
tamoxifen. because they knew tamoxifen had some side effects. so they chose the average 60-year-old woman or so a woman 60 years old or older would have a rye risk to war apt entry into the trial but they would allow younger women to enter the trial
follow they had the risk of an average 60-year-old woman, which turned out to be 1.67% in 5 years. so this is an ethical use of the risk model only to include people who stand to benefit from the intervention. another impact of this, of
thinking this way is the trial has increased sufficiency because by including somewhat higher risk women, you're increasing the number of events likely to arise. you can get by with a smaller study and a shorter study. so these models were used in the
design of the breast cancer prevention trial of raloxifene versus tamoxifen. now, i am going to turn to another topic discriminatory accuracy. and this is really the ability of the model and the risk factors in the model to be very
strong so that the people who actually get disease tended to have higher projected risks than the people who turn out not to get disease. and it's often measured by the area under the curve so-called the probability of a randomly selected case will have a larger
predicted risk than a randomly selected control. and here is an example. suppose this is the two risk distributions here. these are preak see and the red one is for cases and the -- frequency -- black is for non-cases.
if i take a random sample, one person from this population, and see whether that person is projected risk was higher than a randomly selected risk from a non-case, in other words, if i draw a random sample from this distribution see what is the probability this risk is bigger,
it turns out to be .8. and that is the idea of this area under the curve. we ideally are risk models discriminations enough that these curving would be separated and we would know exactly who is going to develop disease and who wouldn't.
but in fact, this would be considered pretty good this unfortunately is what the breast cancer risk assessment tool looks like. there is a lot of overlap between the distributions of risk in the people who turn out to develop disease and those who
don't develop disease. the auc only .6 and that means it has low discriminatory we'd like to improve that and get it up to.8 or 9. one way to get it up to .8 or 9 is to include other risk factors including snps and we'll lab look so see what they might be
able to do for us. a big impact on can snpsincrease dissemtory certain public applications? the basic idea is described in biostatistics that talks about the importance of risk distribution in the population. once you know the risk
distribution in the population and if you know the model is well calibrated, you can calculate all kinds of things like the area under the curve, expected losses in various public health applications. and i explained these ideas to corporate snps in papers in jnci
showing an empirical approach in a paper to evaluate the improvement in predictive ability or discriminatory accuracy from incorporating snps. and in work with park and others, we have been trying to see what might happen in the
future as we think about what snps might be discovered yet in the future. here we are 7 snps available in 2008. the strongest notice, the allele frequencies are pretty big because that is what can be detected by the usual designs
for gwas studies. the odds ratios per allele are pretty small. gee net rick mean was one.one 5 and the strongest was 1.1 26. so that means that if the person has two adverse alleles, it would be 1.2 6 squared or 1.59 is that person's risk compared
to somebody who had no adverse alleles at that locus. and by getting various combinations of these, you might wonder can we get discriminatory accuracy by incorporating snps in our models? i'd like to give you some feeling for that.
this is a very revealing slide. but this is 7 snps. if we have 11 snps which came along a little bit later, we got it up to .585. with 18 snps we got it up to .587. the improvement is diminishing because the snps that get added
later were harder to find and they tended to have smaller odds ratios per allele. but if we do a really huge job of compiling 3 times the number of cases and controls that have ever been compiled up to this present time for studying breast cancer, the work of dr. park and
others, suggests we'll get up to 70 snps and we'll be able to get a auc of .635, something of an improvement and begins to be an interesting number. tool might have 607. you have to ask four questions to get that. if you added 7 snps, you get it
to .632 and if you added 11.637, not much. we think you could get it up to .67 if you added all the foreseeable snps. they may not be quite adds far off as we had imagined when we did the calculations. assessment show plus
mammographic density will get you to .65, a nice strong riskrisk so we are starting to make progress but it's slow. now i'm going talk about two application that is reacquire high discriminatory accuracy. one is high-risk strategy for when the intervention has
adverse side effects. and later, i'll talk about allocation of preventive resources. both of these require good what do i mean by the population strategy and the high-risk strategy? a good place to find out about
this is the gook walled the strategy of preventive medicine -- a book called -- and he talks about the population strategy. suppose i have a intervention that is very safe and i can give it to everyone in the population without risk.
for example, suppose i could suggest to people to eat less salt every day. if erch in the population lowered their blood pressure by one or two millimeters of mercury, that would have more effect on the number of heart attacks on the population than
if i could find the few people in the population with very high blood pressure and go out and treat them. so if you could find a population-level strategy, it is very valuable to consider. two considerations. one might be that the
intervention has adverse effects so you don't want to give it to people unless they have a high enough risk of the disease of interest that your benefit is going to outweigh the losses there. another possible motivation for a high-risk strategy is
economic. you may not have enough intervention for everyone, why not give it to the people at highest risk who need it the most? let's go back to ta mock fen again. and whether this high-risk
strategy works. now i changed it talking about 100,000 women. they have a uterus. 50-15 is not such a good age range to be in for tamoxifen -- 50-59. we would expect 246 to develop invasive breast cancer and so
forth, summing up to 589.6 life-threatening events. iffy woo give tamoxifen to everyone, the population strategy, we are going to prevent a lot of breast cancer and hip fractureurs but because of the other serious events, we will end up much worse off.
so we can't use a general population strategy. we have to be more tailored in giving the drug. it turns out that it is only beneficial for women with very high breast cancer risk. 774 per 100 per year in the 50-59 year age range.
only about 1% of the population has a risk this size. so you're going to be focusing your intervention on only 1% of the population since a lot of the breast cancer will develop in the other 99% unless you have an extremely discriminating model.
you have limited potential for prevention here. and let me just show you how limited. here is again if none get tamoxifen, there will be 589.6 if you use the breast cancer risk assessment tool and just give tamoxifen with the high
enough breast cans tore justify it, you will prevent 1.4 events in 100,000 women in a year. not very impressive. if you add the snps, it's 1.8 per year. again, not very impressive. if we had a perfectly discriminating risk model, it
would just pick out the 246 women who are going to get the you would treat them and because you're treating so few, there would be very few adverse events and you would get a nice savings of about 119 or 120 so, if you had a perfectly discriminating model you could
do well. now, remember this is 1.4 we prevented so far. i'm going to make a comment on that on the next slide, this slide. how could we do better with a high-risk strategy? one thing to make the
intervention less toxic. because then you can give it to more people and you remember the pictures about raloxifene? a lot, it turns out thatra locks 17 a lot less toxic because it doesn't have any endometrial cancer risk. maybe 45% of the population
could get it. instead of preventing 1.4 events you might be able to prevent 20 make the intervention less toxic and more effective in preventing breast cancer and that's a possibility. and improve the discriminatory accuracy of the breast cancer
risk model. that is a possibility but it's a long slog to try do do that. but i think we are making some progress. and another thing one could do is have more refined modeling of the risks of the other outcomes. we have a good risk model for
stroke and a good risk modeled for breast cancer. we could be more accurate in deciding who should get the treatment because we know more about the stroke risk. and so there is some promise but basically, i think if we could move towards less toxic
agents, that would be the best way forward. now i'd like to talk about one other application that also requires good discriminatory accuracy of the model and that is allocating scarce preventive so, in fact, let's suppose that we only have enough money to
give mammograms to half the how would we do that? well, one thing we could do is give the mammograms at random and these women here are shown in conjunction with rectangles. the rectangles are the absolute risk of breast cancer of the but if we give them at random we
don't have to do a raske assessment. you see that some low-risk women get the mammograms and high risk women gets the mammograms and some intermediate. we expect to prevent about half of the deaths using this strategy as we would if we could
give the mammograms to everyone. now, suppose instead we do a risk assessment on everyone and then we only give the mammograms to those at highest risk. are we going to prevent more death? notice that in this picture, four people got the mammograms
and in this picture, five people in pink got the mammograms. and that is because we spent some money doing risk assessments on the entire population and finding the ones at highest risk to give it to. if it costs too much to dot risk assessment, you won't have any
money left for mammograms. you have to be careful there and that's an important ratio. the ratio of the cost of the risk assessment to the cost of the intervention and i assume that it was .02. because compared -- in other words, answering a questionnaire
versus getting a mammogram read. now, suppose everybody could get we could get 100% of the people think that you prevent about 15% of deaths whether in young women or older women by giving mammograms. so you get 100% of that benefit if you could give everybody a
mammogram. if you give the mammograms at random and only have enough to give half, well, you're going to only get 50% of that benefit. but if you rank them and use the breast cancer assessment tool, you will get 63%. that's an improvement.
and if you use the snps 2 and we can do better than this. you're up to 66%, 67%. you're seeing that you can get or save more lives if you use this ranking-type strategy. assuming that it is feasible to do. one thing about this, it may be
something of an however estimate because it costs money to draw the bloods -- over estimate -- to seas the snps and to transfer that information back. if this were all free, or if this cost just as much as this, 2% of the cost of the mammogram, these numbers would be valid.
but if the ranking procedure costs 20%, of the cost of the intervention, then it is useless. so i tried to talk a little bit about what absolute risks are and a little bit about the models that have been developed to compute absolute risk for
i vied to discuss counseling patients and how knowing an absolute risk can give a perspective used in thinking about management. and how it can be used more formerly to weigh favorable and unfavorable effects of a preventive intervention.
i tried to indicate that absolute risk is usesful in designing prevention trials and can give a useful perspective on the risk reduction that you're going to achieve potentially in the populations if you had certain preventive interventions.
all of these applications only require modest discriminatory but for implementing a high-risk prevention strategy or allocating scarce resources it's important to try to improve that discriminatory accuracy and we are making some progress if we get 70 good snps and
mammographic density incorporated into the model and if we find some other strong imaging or other factors we can put into the model, we can do better in these areas. and i have indicated on this slide some of the people who contributed to work directly
presented in this talk today. there are others that i have enjoyed working with that are not on the slide and are doing related things. so, with that, eye thank you and i'm happy to answer questions. >> [ off mic ] >> that's the instantaneous
it's basically the sear rate for the 70-74-year-old woman divided by the seer rate for the younger. in the next year, it's like a derivative, sort of. >> that was a wonderful talk, thank you. it was really showing us the
power of the approach. i wanted to ask you what you thought about two possibilities and ask you what you that about something like mammography, in theory you could be always updating the model, you could have a mammogram and then depending on the results you
could have a mammogram and then could -- so it will become less and less expensive presumably to be rapidly sort of updating these risks in realtime and maybe in order nary practice. so, i was curious whether you think that will be something that actually translates to
meaningful benefit? and then the other thing is you mentioned risk benefits and getting your bmi down. probably again, increasingly you're getting information so that you might be able to say here is your risk and we do know something about what happens
when weep attempt to, let's say lose weight or reduce alcohol or whatever you speak about, you could presumably feature that in too. so you could say, there is a broader set of predictions. i'll predict the likelihood that you are going to get breast
cancer in five years and then i'm going to tell you a bit about how likely you might be to succeed if you were to attempt to change x, y or z. so both of those assuming it is right, the computing power will just get more and more so. but it struck me they could be
distracting and yet potential extensions of the general method. so i was wondering whether you thought either of those would be kind of valuable directions to go or just sort of finessing and adding useless stuff. >> let me make sure i understand
the first part. it was about incorporating new risk factors like mammographic density as the data becomes available. is that right? >> in realtime. i get me pam grabbing gram, i come back and you say your
number just changed. >> i think that there have been enough studies on mammographic density we know it is a valuable risk factor. it should be incorporated into models. one nice thing about the breast cancer risk assessment tool, you
just ask four questions or so and you're done. as soon as you begin to use more informative data of the mammographic density type, it's a more elaborate process. but that is no reason why we shouldn't try to incorporate that information.
now, there is a long road between the idea and the quality controlled implementation. and, for example, how should we measure mammographic density? should we use bireds? should we use some automated procedure? we need to get agreement on that
and then we can recommend a model that uses an agreed upon way of measuring it. and then ideally before broadcasting it beyond the research setting, we should also try to do some independent validation and independent cohort.
so, it is ready. i think that it is ready to introduce, but there are some steps that would ideally be done before it is made widely for example, jim bow chen in our group, did work on incorporating mammographic density but the measurement used was a
hand-driven -- [ indiscernible ] and it is not something that could be widely and generally used. so, now bireds has shown a lot of promise and it is being widely used. i think it has good information wellinformation there.
maybe we should go with something like that. but i think there is more information that could be extracted in a standardized way. the behavioral one is -- i think that these models give you perspective on what would happen if.
and it's a big if. what would happen if the people actually did implement the intervention? what would happen if what we have observed in observational data on these associations translates into the effect of an because an observational
association is not the same as an intervention. and so, the issue is -- i'm not saying -- i think it can tell you what we potentially might hope to achieve. but i don't think i can tell you what we will achieve. >> thank you.
>> so with respect to the power of your model and how much it can go up. given that whole genome sequencing is becoming cheaper and cheaper, i was wondering if there might be a time that you do whole genome sequencing for a person for say 1000 dollars, and
the information would be useful not only for predicting breast cancer or everything else and that is what you do one time in your life. is it possible to predict, assuming we do a whole genome scan and that we find variants that are not common, but
increase the risk quite a bit, let's say high relative risk but uncommon mutations, can we predict using heritability how much your model can predict that one person will have the disease, assuming we have a whole genome sequencing from one person?
>> now, moi experience is really with snps and i think we have a pretty good grasp on where we are likely to be able to get with snps. but as to whole genome sequencing, i have read one paper by some coauthors of burt fogelstein based on twin studies
that try to say, since a twin is genome identical, we can try to use this information to figure out what we could possibly get if we were doing the whole genome sequencing. my recollection is that the results were not astounding. and this was published in
translational medicine or but i think that probably other people in the audience who might have more insight into what we might achieve. one limitation is, though, is clear. we have hard enough time trying to extract the information on
snps and we are using larger and larger samples to do this. the multiplicity problems of whole genome sequencing are going to be that much greater. so to pulley accident tract the information, especially if you have many small effects may take unimaginable or unattainable
sample sizes. >> and my question, assuming we use your model to offer let's say mammography to 63% of the population rather than 50%, if the full benefit of the model is used, how likely is it going to be to use or to do that in the united states where there are
many different insurance companies and lots of lawsuits if you tell them, this is not for you and then in one year, they get cancer. it might be very different in a country like u.k. where there is national health system that offers it and they can basically
triage people versus u.s. where different people -- >> you're talking about the resource allocation issue? >> exactly. >> there is some assumptions one of the assumptions is the people who come to be or have the risk assessed are random
sample of the general another assumption is that you can actually implement the risk-based allocation. and i'm just trying to show what could be done, not necessarily -- and i think that one has to think very carefully before trying to do something
like this in the particular context. >> so, talking about the chemo prevention trade off between risk and benefits like my understanding is we are considering all the life-threatening outcomes equal. so, has there been any attempts
in kind of incorporating more information like survival or how much -- like stroke, although stroke and breast cancer are both life trenting, they might not have the same weight in the cost they impose or survival. has there been any effort in this regard and do you think it
say worthwhile consideration, extension? >> we thought about doing that originally and we said it would be better to try to keep itism. mathematically there is no problem in assigning different positive weights to the outcome. and in fact, if you wanted to do
this in an interactive mode, it might be important to allow the woman to accipe her own weights, because she may have or may not care about some of the things that you mentioned but she may be deathly afraid of a stroke. and she would do anything to avoid a stroke.
so she might give that a weight of 10. and i think that things like that can be done by incorporating them in an interactive program.
No comments:
Post a Comment