Addressing Unmeasured Covariates in Source Cohorts with Transfer Learning for Survival Outcomes

About This Grant

PROJECT SUMMARY Accurate risk assessment is essential for guiding clinical decision-making in rare cancer cases, especially within a specific medical institution, due to variability and heterogeneity across institutions. However, the limited sample sizes typically available for rare cancers in a single institution present significant challenges for survival analysis. The primary objective of this research program is to advance statistical methods that enhance risk assessment for a target cohort by adaptively leveraging information transferred from external source cohorts. This research focuses on a common scenario where the target cohort from a single institution collects more detailed covariates—such as newly developed biomarkers and comprehensive genetic information—than the external cohorts sourced from cancer population registries or research consortiums. Conventional methods often assume that both cohorts share the same covariates, which limits their applicability when crucial covariates are missing in the source cohorts. To address these limitations, we propose two transfer-learning frameworks that adaptively borrow information from the source cohort while accounting for differences in covariates and time-dependent hazards. Our specific aims are: (1) develop a novel transfer-learning-based Cox model that accommodates the absence of key covariates in the source cohort, enabling effective information transfer; (2) create a group-specific transfer-learning-based Cox model that allows for flexible information borrowing at the subgroup level when heterogeneity between the target and source cohorts varies across subgroups; and (3) develop and disseminate publicly available, user-friendly software packages to ensure the reproducibility and application of our methods across multiple datasets. Although the proposed methodology is agnostic to disease type, we will demonstrate its utility in the context of inflammatory breast cancer (IBC) and myelodysplastic syndromes (MDS)—both of which are rare, aggressive cancers—making them ideal proof-of-concept cases for our methods. Overall, this project aims to advance statistical methods in personalized risk prediction and treatment strategies by facilitating adaptive knowledge transfer from external data sources, even when cohort discrepancies exist. More importantly, this work has the potential to significantly improve risk prediction and treatment selection for rare cancer types, ultimately helping clinicians develop optimal, patient-specific treatment strategies.