A transfer learning approach based on random forest with application to breast cancer prediction in underrepresented populations.

Despite the high-quality, data-rich samples collected by recent large-scale biobanks, the underrepresentation of participants from minority and disadvantaged groups has limited the use of biobank data for developing disease risk prediction models that can be generalized to diverse populations, which may exacerbate existing health disparities. This study addresses this critical challenge by proposing a transfer learning framework based on random forest models (TransRF). TransRF can incorporate risk prediction models trained in a source population to improve the prediction performance in a target underrepresented population with limited sample size. TransRF is based on an ensemble of multiple transfer learning approaches, each covering a particular type of similarity between the source and the target populations, which is shown to be robust and applicable in a broad spectrum of scenarios. Using extensive simulation studies, we demonstrate the superior performance of TransRF compared with several benchmark approaches across different data generating mechanisms. We illustrate the feasibility of TransRF by applying it to build breast cancer risk assessment models for African-ancestry women and South Asian women, respectively, with UK biobank data.