A0632
Title: A robust angle-based transfer learning
Authors: Tian Gu - Columbia University (United States) [presenting]
Abstract: The increasing numbers of large-scale biobanks and institutional data networks have brought unique opportunities to link patient genomics, electronic health records, and survey data for studying complex human diseases, primarily to address the diminished model performance in minority and disadvantaged groups due to their low representation in biomedical studies. A novel angle-based transfer learning (angleTL) method is proposed to improve risk prediction in underrepresented populations by integrating data from multiple biobanks, different ancestries, and related outcomes. It protects data privacy by learning from pre-trained models in external data sources without sharing patient-level data and accounts for potential data heterogeneity. Theoretical guarantees are provided for the model performance and insights regarding when the external model can be helpful to the target model. It is shown that angleTL unifies several benchmark methods by construction, with examples using data from the UK biobank and the electronic medical records and genomics (eMERGE) network.
