Program - AI Reliable

Zhiyu Wan, PhD, is a postdoctoral research fellow in the Department of Biomedical Informatics, Vanderbilt University Medical Center. He is mentored by Dr. Bradley A. Malin, and is currently working in the Health Information Privacy Laboratory (HIPLAB), and the Center for Genetic Privacy and Identity in Community Settings (GetPreCiSe). His research interests include: optimizing privacy-preserving data sharing, with an emphasis on health and genomic data; leveraging game theory, artificial intelligence, and machine learning techniques to improve social good; and analyzing social networks and network security. His latest research project is using game theory to protect genomic data against privacy attacks. He finished his PhD in Computer Science at the Department of Electrical Engineering and Computer Science, School of Engineering, Vanderbilt University. He obtained his bachelor’s degree in Automation (minor in Business Administration), as an outstanding graduate, from Xi’an Jiaotong University, which he entered through the Special Class for the Gifted Young of China. 

Title: Promoting responsive real world data sharing in medicine using AI agents 

Abstract: In the latest AI boom, effective deep learning and large language models that rely upon gigantic amounts of training data are accelerating healthcare and medicine research. An insufficient amount of real-world data can lead to unreliable and unfair machine-learning models. However, sharing health data has privacy concerns. Years ago, I started working on optimizing the way to share real-world data responsively, with a focus on promoting privacy and fairness. Using a game theoretic lens, I found the best way to share genomic data, electronic health records (EHR), images, and voice recordings from patients to third parties or the public. In most experiments, I analyzed privacy risk and data utility using real-world data from EHR along with external resources such as voter registration lists and recreational genomic databases. This is a fundamentally different approach because it accounts for adversarial behavior and capabilities and tailors protections to anticipated recipients with reasonable resources, not adversaries with unlimited means. This multi-agent-based approach can be integrated with existing data protection models and can be applied to assess risk and suggest an optimal data-sharing strategy based on a privacy-utility tradeoff. The results show that most data can be effectively shared with low privacy risk in a game theoretic model. Additionally, I used these risk metrics to evaluate and guide the development of data protection methods such as synthetic data generation which generates fake health records based on real world data and is expected to bring real-world evidence in medicine.