Breast cancer risk prediction models improved by adding multiple biological markers of risk
Dr Xuehong Zhang - Dana-Farber Cancer Institute, Boston, USA
As you know, the goal of breast cancer risk prediction models is to identify women at particular high or low risk which can help at least planning the iteration trials as well as the clinical decision making like tamoxifen for prevention. We considered two models in this study, the first one is the National Cancer Institute’s Gail model which includes the factors of age, age at first birth, age at menarche, number of biopsies. The second model is the Rosner-Colditz model which was developed in the Nurses’ Health Study and was validated in another independent cohort. It includes the factors included in the Gail model plus the body mass index, aqua consumption and other well-known breast cancer risk factors. Those two models only included traditional breast cancer risk factors; the ideal predictive values of the following biological markers to the risk reproduction model are not well understood. As we know, a genetic risk score, percent mammographic density and circulating postmenopausal hormones are independent breast cancer risk factors. Several prior studies have evaluated the added predictive values of genetic risk score, percent mammographic density or both but none of the studies have yet evaluated the predicted added value for those three biological markers.
So the goal of this study is to evaluate the joint contribution of those three biological markers. We analysed both the overall invasive breast cancer as well as the oestrogen receptor positive, ER positive, tumours. Because of time limits I’ll only present the results on postmenopausal women not using hormones.
We used the data from the two prospective cohort studies. The first study is the Nurses’ Health Study which initiated in 1976, includes over 120,000 women. We sent out the biennial questionnaires to collect the information on lifestyles and breast cancer risk factors. Around 1990 over 32,000 women provided first blood, among whom 60% provided second blood around 2000. Over 29,000 others provided their cheek cells which were used, both the blood and cheek cells were used, for genetic analysis. The second cohort we used here is the Nurses’ Health Study 2 which initiated in 1989 and includes over 110,000 women. Similarly we collected the blood around 1996 and over 29,000 women provided cheek cells after 2000. The following slide gives you a sense of the number of cases and the controls in this study for the Gail model. Overall we have about 4,000 cases and 8,000 controls and we have both the questionnaire data and the genetic data. Among the women providing pre-diagnostic blood samples, we call it the blood sample cohort, we have measured the plasma hormones and also the mammographic density.
So the genetic risk score is a single number summary of the genetic risk of breast cancer and we derived this score based on 67 common genetic variants based on the meta-analysis of over 10,000 cases weighted by the strength of the association. We measured the mammographic density in the blood sub-cohort and also the three hormones in the blood sub-cohort.
This is the first important data. We showed an association between each of the biological markers with the breast cancer risk. I’ll show you in this slide genetic risk score and some mammographic densities are strong, independent breast cancer risk factors. Women at the top versus the bottom categories have about a 2.5-fold increased risk for both the genetic risk score and percent mammographic density. The associations were independent of each other and also other known breast cancer risk factors.
The next slide shows the main results on the association between circulating hormones and breast cancer risk. As you can see, the women in the high level group have about 1.5 to twofold increased risk of breast cancer, this is slide number 14. Also the association is independent of genetic risk score, mammographic testing or other breast cancer risk factors.
This is the main result for this study. As you can see, for the Gail score alone we have the AUC statistical analysis part. We calculated the AUC, the area under the curve, adjusting for age as a measure of discriminative accuracy for five year breast cancer risk. The AUC is the probability of randomly selected cases will have a higher risk than a randomly selected control. The AUC value varies from 50 to 100% with 50 equals the probability of flipping a coin and 100% is the perfect discrimination accuracy. So the higher the value the better the discrimination accuracy of the model.
This slide shows the main results. As you can see, the Gail model alone gave us the AUC of 55.2 which is very consistent with the majority of the previous studies. Adding each of the biological markers we have a significant improvement in the AUC changes ranging from 3 to 6 units. Adding three of them together gave us the best prediction, it’s about 10.8 significant improvement compared to the Gail model alone. We observed a similar pattern for the Rosner-Colditz model, as shown in the next slide. Adding those three biological markers together gave us about a six point unit increase, it reached 66.2.
Lastly we did a sub-analysis on the oestrogen receptor positive tumours, as shown in this slide, number 17. We observed a slightly stronger improvement in the AUC changes. Adding those three biological markers together for the ER positive tumours we observed an 11.7 significant increase. For the Rosner-Colditz model we observed a significant improvement, a 9.4 unit increase.
The conclusion of this study is that we found that incorporation of the multiple biological markers can improve the the Gail and the Rosner-Colditz models for both invasive and oestrogen receptor positive breast tumours. If validated in another independent population our findings could help identify women at high risk who are most likely to benefit from chemoprevention, other risk reducing regimens or screening.
Thank you. So how is your group going to take this forward? You are one of the most important international groups with the Nurses’ Health Study so how will you advance these findings now?
First of all, these initial findings need to be replicated in an independent population. We are in progress of collaborating to get funding to do that. We need to revise in the non-Caucasian population we need to predict absolute risk and assess the calibration.