Mining minority enriched AllofUs data for innovative ethnic specific risk prediction modeling

PROJECT SUMMARY/ABSTRACT Advancement of health equity requires evidence and tools tailored for minority groups. The shift towards individualized precision medicine requires risk prediction tools to guide prevention and intervention. Due to the genetic heterogeneity and social economic disparity, risk factors may disproportionately impact race/ethnicity (R/E) groups. Overall risk prediction constructed from predominantly white populations can perform poorly on other ethnic groups, leading to mis-diagnosis, over-treatment and other adverse health consequences. Efforts on developing R/E-specific risk prediction at local healthcare systems are limited by the small sample size caused by inadequate representability of minority populations. To address the gap and to advance precision medicine for non-white patients, it is crucial to harness minority enriched clinical data and develop risk models transferable to point of care. The All of Us (AoU) program offers a wealth of comprehensive multi-modal data on whole genome sequencing (WGS), real-world electronic health records (EHR) and patient reported outcomes (PRO) with enhanced minority participation, providing the common evidence base for learning general R/E-specific risk patterns and training risk models for minority populations at local healthcare systems. In this proposal, we develop innovative methods for risk modeling in AoU data tailored for minority populations and its validation on external healthcare data. We will showcase the proposed methods in two use cases: 1) rheumatoid arthritis (RA) genome-wide association study (GWAS) at Mass General Brigham (MGB) focusing on the genetic risk factors; 2) cancer cardiotoxicity prediction study at M Health Fairview (MHF) focusing on clinical and social determinants of health (SDoH) risk factors. In Aim 1, we integrate risk factor and disease onset outcome data across WGS, EHR and PRO in AoU data to construct the risk prediction model that yields better risk prediction accuracy, risk factor identification and fairness across R/E groups. In Aim 2, we design privacy preserving algorithms to validate the generalizability risk modeling from AoU data on external healthcare data and establish the transfer learning strategy to adapt AoU risk models for local healthcare systems. We intend for the methods to facilitate development of risk modeling using AoU data with focus on minority populations, as well as toe demonstrate the potential impact of the AoU program on improving care at local healthcare.