Cancer risk prediction is a critical step towards the development of targeted cancer prevention and screening policies. There is a growing awareness that cancer risk prediction studies may be distorted by detection bias, particularly in screened populations. Detection bias occurs when screening and diagnostic patterns vary in association with potential risk factors. Detection bias can exaggerate or attenuate estimated disease-risk factor associations and may adversely affect our ability to develop sound prevention and screening policies. The objective of this application is to change the way that detection bias is assessed and addressed in cancer risk prediction. We will harness the technique of disease natural history modeling to decouple the underlying risk of disease from observed screening and diagnosis histories. We will rigorously investigate the performance of disease modeling to reduce detection bias and will apply our approach to assess and address detection bias that may already be impacting early detection guidelines in prostate and breast cancer. We will disseminate our models via an online user interface that will permit investigators conducting risk prediction studies in screened populations to assess their studies' susceptibility to detection bias. Finally, we will study the impact of detection bias on policy-relevant outcomes via a proof-of-concept study of prostate cancer screening.
Our specific aims are as follows:
Aim 1 [Methods development]: Develop and validate a cancer modeling method for assessing and reducing detection bias in risk prediction studies based on screened populations;
Aim 2 [Breast density application]: Apply the method developed in Aim 1 to assess and remediate any detection bias in published associations between breast density and breast cancer risk. Despite the major policy implications of findings that breast density leads to an elevated risk of breast cancer diagnosis, these findings have never been interrogated for detection bias;
Aim3 [Software dissemination]: Develop, test, and deploy an online user interface that will permit investigators conducting cancer risk prediction studies in screened populations to assess the potential detection bias;
Aim 4 [Policy impact]: Assess the impact of detection bias on harm-benefit tradeoffs of candidate prostate cancer screening policies as a proof of concept for the translation of detection bias to the policy setting. This application will pioneer the use of disease modeling as tool for addressing a source of bias that may be present across a wide range of policy-driving cancer risk predictions. The investigator team is comprised of leading investigators in the development of disease models for early detection. The proposed work will produce the most rigorous analysis to date of the way that detection bias works and how it may be addressed in practice.
As we learn more about cancer, we are able to better predict who is more or less likely to get it and we can begin to make recommendations that are tailored to an individual's disease risk. When people are screened for cancer, their risk of being diagnosed may go up, particularly if they are screened often. This application will develop methods that will help people understand how common risk factors like family history and breast density affect their risk of getting cancer rather than their risk of getting a cancer diagnosis.
