Investigating the roles of personal characteristics in cancers of the breast and prostate
- Background: Breast cancer and prostate cancer have both high prevalence and high mortality burden on their respective sexes. Outcomes are highly variable and risk modeling is important to prevent costly and dangerous overtreatment. We investigate risk factors and build models of breast cancer and prostate cancer risk. Mammographic density is positively associated with breast cancer risk. The strength of this association has been well studied using film mammograms; however it is currently unknown how best to measure breast density from full field digital mammography (FFDM) images in order to predict future breast cancer risk. We compare three density metrics: BIRADS, Cumulus and Volpara. Prostate cancer prognosis is highly variable, and treatment decisions involve patients' competing risks of disease recurrence and death from other causes. However, recurrence predictions from available risk models do not consider such death, nor do they include body mass index (BMI) and race as potential predictors. To address these issues, we developed and cross-validated a personal prediction model for prostate cancer recurrence and recurrence-free death within ten years of radical prostatectomy (RP), and we compared the performance of its predictions with those of an existing risk model. Methods: We applied our modeling approach for risk stratification in two use cases. First was the breast study, which included 134 women with invasive breast cancer and 279 age- and race-matched controls who underwent screening mammography on a GE FFDM system during 2004-2013. We used the Cumulus 6 software for two-dimensional (2D) assessment of percent density (PD) and Volpara v150 software for three-dimensional (3D) assessment of Volpara percent density (VPD). We also extracted the BIRADS density recorded by the radiologist from the mammography report. We assessed intrareader reproducibility of PD by computing Pearson's correlation coefficient (R) on blinded repeat measurements. We estimated the odds ratios (ORs) for breast cancer associated with PD, VPD, and BIRADS density category using conditional logistic regression, stratified by age and race, and computed the area under the receiver operating characteristic curve (AUC) for each method. In the second case for the prostate study, we developed a model of prostate cancer recurrence by applying statistical learning methods to data from a cohort of 1276 patients who had undergone RP at the University of Pennsylvania. These methods provide internal cross-validation and can account for the competing risk of dying before recurrence. We evaluated model accuracy using standardized residuals that relate observed to predicted outcomes, and evaluated model discrimination using the area under the receiver operating characteristic curve (AUC). Results: For the breast study, the OR for women in the highest vs. lowest quartiles of PD was 3.07 (95% confidence interval 2.32-3.81), with AUC=0.66 (0.60-0.71), after adjusting for BMI, total breast area, parity, and menopausal status. The comparable OR for VPD was 2.11 (0.96-4.64), with AUC=0.63 (0.57-0.69). BIRADS density was strongly associated with breast cancer risk; the OR for women with extremely dense (category 4) relative to those with almost entirely fatty breasts (category 1) was 6.70 v (5.67-7.72), AUC=0.68 (0.63-0.74). Relative to women in BIRADS categories 1 or 2, the OR for women in BIRADS categories 3 or 4 was 2.81 (1.69-4.68), with AUC=0.66 (0.61-0.72). For the prostate study, cross-validation of the prostate cancer model showed greater accuracy for the new model compared to the existing one, which underestimated patients' risks. The accuracy gains also were noted when both models used the same set of covariates. The new model's discrimination (AUC=0.767 (0.711-0.823)) was similar to that of the existing model (AUC=0.735 (0.682-0.788)). The most important recurrence predictors were preoperative prostate-specific antigen (PSA) level and body mass index (BMI). Conclusions: We have demonstrated that statistical modeling approaches can be used to stratify disease risk in two disparate disorders. First, mammographic density assessed on FFDM images was significantly associated with breast cancer risk, confirming prior results with film screen mammography. Associations between breast cancer risk and density were strongest for radiologist-assessed BIRADS estimates and weakest for Volpara PD estimates. The new prostate cancer recurrence model showed better prediction accuracy than the existing one, and life expectancy and BMI were important determinants of recurrence risk. These findings, if validated in different settings with different patient populations, should facilitate decisions regarding post-operative treatment. We believe our approach is extensible to other disease scenarios.
|Type of resource
|electronic; electronic resource; remote
|1 online resource.
|Jeffers, Abra Marie
|Stanford University, Department of Management Science and Engineering.
|Weyant, John P. (John Peter)
|Weyant, John P. (John Peter)
|Rubin, Daniel (Daniel L.)
|Rubin, Daniel (Daniel L.)
|Statement of responsibility
|Abra Marie Jeffers.
|Submitted to the Department of Management Science and Engineering.
|Thesis (Ph.D.)--Stanford University, 2015.
- © 2015 by Abra Marie Jeffers
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...