Targeting Underrepresented Populations in Precision Medicine: A Federated Transfer Learning Approach

Abstract

The limited representation of minorities and disadvantaged populations in largescale clinical and genomics research has become a barrier to translating precision medicine research into practice. Due to heterogeneity across populations, risk prediction models are often found to be underperformed in these underrepresented populations, and therefore may further exacerbate known health disparities. In this paper, we propose a two-way data integration strategy that integrates heterogeneous data from diverse populations and from multiple healthcare institutions via a federated transfer learning approach. The proposed method can handle the challenging setting where sample sizes from different populations are highly unbalanced. With only a small number of communications across participating sites, the proposed method can achieve performance comparable to the pooled analysis where individual-level data are directly pooled together. We show that the proposed method improves the estimation and prediction accuracy in underrepresented populations, and reduces the gap of model performance across populations. Our theoretical analysis reveals how estimation accuracy is influenced by communication budgets, privacy restrictions, and heterogeneity across populations. We demonstrate the feasibility and validity of our methods through numerical experiments and a real application to a multi-center study, in which we construct polygenic risk prediction models for Type II diabetes in AA population.

Algorithm Fairness in AI for Medicine and Healthcare

“…With respect to fairness, federated learning paradigms for decentralized AI-SaMD development has been demonstrated to have a directly mitigate disparate impact via model development on larger and more diverse patient populations 58 . For instance, In the case of population shift as a result of genetic variation, decentralized information infrastructure have been previously proposed to harmonize biobank protocols and developed tangible material transfer agreements amongst three hospitals, which demonstrates the potential applicability of federated learning paradigms in developing large and diverse biobank data for diverse populations 189 .…”

“…In developing polygenic risk scores, federated learning has been used as an integration strategy in merging heterogeneous population data from multiple healthcare institutions, with subsequent validation of federated models on underrepresented populations 58,190 . In the previous examples of site-specific staining variability across different hospital sites, federated learning can be used to train decentralized models that are invariant to stain via domain generalization, as well as domain adaptation in refining AI-SaMDs locally to the test data distribution with minimal updates 29 .…”

COMMUTE: communication-efficient transfer learning for multi-site risk prediction

“…The key of transfer learning is to characterize and leverage the similarities between the source and the target populations. Previous work such as Li et al 29 , 30 , Tian and Feng 47 , Xu and Bastani 55 characterize the similarity between the target model and the source models by a 0 distance measure…”

“…Methods that combine transfer learning and federated learning can be highly relevant in the application of multi-biobank risk prediction. For example, Li et al 29 proposed a federated transfer learning approach to leverage source data to the target population. Cai et al 7 proposed an integrative estimation procedure for multi-task high-dimensional regression.…”

“…The most significant drawback of these methods is that they all require each site to share both the first and second-order derivatives of the loss function with respect to the model parameters. In a moderate or high dimensional setting, when the number of parameters ( p ) can be comparable or even greater than the sample size of a population ( n ), sharing second-order information ( 0 ( p 2 ) numbers) can be challenging regarding privacy, data storage and communication cost 29 .…”