Using Machine-Learning to Identify Differences in the Association between Blood-Based Biomarkers and Later Life Health Across Race and Ethnicity.

Abstract

Increasingly, biomarkers are used to understand health and health inequalities among older adults. Combined with advancements in machine-learning approaches, researchers are using predictive algorithms of later life health to identify biomarkers of interest and create biological risk scores. However, these algorithms may select biomarkers that are most important for majority populations, which, in most population-based samples, would reflect the health and aging of White older adults. Understanding how biomarker selection varies across race/ethnicity across different types of health outcomes is paramount to advancing GeroScience research. We used the 2016 Venous Blood Substudy (VBS) of the Health and Retirement Study (HRS). We fit race-stratified boosted decision tree models to predict all-cause mortality, multimorbidity, diabetes, and heart conditions from 54 biomarkers in the 2016 VBS that covered 11 biological systems. We, then, graphed biomarkers that had feature values above .01 for each algorithm to show racial/ethnic differences in biomarker selection. We found more variation in biomarker selection across racial/ethnic groups for all-cause mortality. We found little variation in biomarker selection for heart conditions and diabetes. There was some variation for multimorbidity but with substantial overlap across racial/ethnic groups. While machine-learning approaches for developing biological risk scores and identifying biomarkers linked to later life health will yield additional insight into aging processes in human populations, researchers must consider how these approaches may differ across race/ethnicity for different types of health conditions and its potential implications for Geroscience research.