BE Seminar – Hanwen Xu, “Advancing Precision Medicine with Precision Journey”

Traditional medicine failed because it is designed for the “average patient” as a one-size-fits-all solution. The generative AI revolution promises the moonshot of conditioning each treatment on the right patient at the right time, named as precision medicine. But knowing who the right patient is and when the right time is is far from trivial. Many current biomedical research projects limit their scope to a static snapshot of the patient journey. However, our research vision is that precision medicine should go beyond and operate across the entire patient journey. In this talk, we present Precision Journey, a general framework for delivering the right treatment to the right patient at the right time by jointly modeling patient state (who the patient is) and patient journey (how the patient evolves). First, to decode patient state from existing large medical data, we developed GigaPath with efficient long-context modelling to handle a billion-pixel image. Aproposed GigaTIME, a cross-modal translation framework that imputes the expensive medical signals from routine and standard clinical data. Finally, we simulate the patient journey to follow how the patient changes over time, to determine the right time to use a treatment. In addition to contributing to the moonshot of precision medicine, our work is also characterized by the real-world clinical applications. The models we developed have been used by radiologists and pathologists, accelerating their clinical practice for better patient care.

Hanwen Xu is a Ph.D. candidate in Computer Science at the University of Washington. His research focuses on gigascale, multimodal AI for modeling the patient journey, with the goal of moving precision medicine beyond static prediction toward individualized, time-aware decision making. His research has made important contributions in computer science and AI for medicine, which merit first-authored publications in top scientific journals (Nature, Cell, NEJM AI, Nature Communications, Cell Genomics) and top CS/CompBio venues (AAAI, RECOMB, CVPR workshop), and fellowships (Weil Family Endowed Fellowship in Computer Science & Engineering), with more than 6.5 million HuggingFace downloads and 700 GitHub Stars. His work also received media coverage from 50+ news outlets worldwide, including MIT Technology Review, Forbes, and Yahoo. The models he developed have been widely used by pathologists, radiologists, oncologists, and researchers from Providence Genomics, UW Medicine Hospital, and MSKCC.