Kara Liu

Machine learning (ML) models trained on sensitive personal information have the potential to drive innovation in many domains, including fairer prediction methods and novel clinical insights. However, privacy regulations heavily restrict the sharing and subsequent research of valuable healthcare data. To address this, I propose a novel generative model for synthetic electronic health records (EHRs) that ensures patient privacy while accurately representing complex medical conditions. My work will leverage biological knowledge graphs to generate nuanced disease representations. By conditioning on these disease representations using a generative diffusion model (DDPM), I will generate realistic EHR data that captures complex patient phenotypes. Finally, my project will train the generative method to enforce patient privacy. By integrating advanced techniques from both medicine and computer science, my research holds promise in democratizing access to EHR data, fostering widespread medical research and enabling the development of improved healthcare solutions.