Comparing the influence of social risk factors on machine learning model performance across racial and ethnic groups in home healthcare

Abstract

This study examined the impact of social risk factors on machine learning model performance for predicting hospitalization and emergency department visits in home healthcare. Using retrospective data from one U.S. home healthcare agency, four models were developed with unstructured social information documented in clinical notes. Performance was compared with and without social factors. A subgroup analyses was conducted by race and ethnicity to assess for fairness. LightGBM performed best overall. Social factors had a modest effect, but findings highlight the feasibility of integrating unstructured social information into machine learning models and the importance of fairness evaluation in home healthcare.