Grant Number:	1R21CA301112-01A1 Interpret this number
Primary Investigator:	Su, Yu-Ru
Organization:	Kaiser Foundation Research Institute
Project Title:	Predicting Surveillance Mammography Failure in Breast Cancer Survivors: a Statistical Framework Accounting for Irregular Surveillance Patterns
Fiscal Year:	2026

PROJECT SUMMARY Current clinical guidelines recommend annual breast surveillance examination with mammography for more than 4 million breast cancer survivors in the US to promote early detection of breast cancer recurrence and reduce breast cancer morbidity and mortality. Despite the potential for early cancer detection, surveillance mammography has sensitivity of only 70.4% in breast cancer survivors, indicating that nearly 30% of breast cancer recurrences are missed. Supplemental breast imaging modalities, such as magnetic resonance imaging and whole breast ultrasound, can improve the detection rate of second breast cancer when combined with annual surveillance mammography, but access to supplemental imaging is limited. An accurate risk prediction model for surveillance mammography outcomes (e.g., second breast cancer missed by surveillance mammography) is key for efficient resource allocation by identifying breast cancer survivors with greater need for supplemental imaging. Electronic health records (EHR) are excellent resources for clinical risk modeling but often feature irregular and outcome-dependent observation patterns which introduce analytical challenges. In breast cancer surveillance, such irregular observation patterns are the consequence of early and delayed return for the next surveillance mammogram. This results in statistical challenges such as differential truncation of follow-up time and collider- conditioning bias. It may be challenging to use existing methods including risk modeling with inverse probability weighting (IPW) to address these issues when modeling surveillance mammography outcomes. It is because the mammogram result, a component of surveillance performance outcomes, partially determine when to return for the next round of surveillance, violating the independence assumption between the outcome and the observation patterns given the covariates for IPW. Moreover, the performance of IPW-based risk models is sensitive to model misspecification of surveillance patterns, which is difficult to avoid in real-world applications. We aim to identify factors related to surveillance patterns in EHR and develop a new risk modeling method that accounts for irregular surveillance patterns. Aim 1 will identify factors associated with irregular surveillance patterns. Aim 2a will develop a novel and robust weighted risk modeling framework to account for irregular surveillance patterns. Aim 2b will evaluate how model misspecification of surveillance patterns impacts the predictive performance of existing and novel risk modeling approaches via plasmode simulations. We will leverage existing longitudinal EHR data from the Breast Cancer Surveillance Consortium and data from the US Census and American Community Survey for the proposed analyses. Findings from this study will provide valuable knowledge on factors impacting surveillance adherence. In addition, we will advance statistical methods for EHR-based risk modeling to generate reliable empirical evidence for accurate risk-based breast cancer surveillance, leading to improved early detection of cancer recurrence and cancer outcomes in the long-term.





None