Prashnna Gyawali

Over the last decade, there has been phenomenal success in AI algorithms, driven mainly by
supervised deep learning, across various domains, including computer vision and NLP. However, similar
performance in health and medicine remains challenging due to current models’ ineptness to generalize across
unseen data (e.g., new patient cohort). First, unlike other domains, it’s expensive to curate training datasets with
well-annotated labels for the healthcare domain. Second, depending upon data modalities, different data
attributes such as patient’s age, sex, race, and anatomical variations often confound with the predictive task,
exhibiting disparities in performance across various patient subgroups. Furthermore, adopting AI models in a
real-world clinical setting would require robust explanation and interpretation. In this talk, I will present work
on these critical aspects of healthcare AI. First, I will present my work on learning unbiased deep learning models
by disentangling (or separating) representations. Second, I will discuss the importance of interpretability in
healthcare and the challenges current deep learning models face. I will then present my recent work to address
some of the difficulties around interpretability. Finally, I will envision future opportunities such as foundation
models and federated multimodal learning.

Biography:

Prashnna K Gyawali is an Assistant Professor in the Lane Department of Computer Science and
Electrical Engineering at West Virginia University. Previously, he was at the School of Medicine at Stanford
University as a postdoctoral scholar. He did his Ph.D. Ph.D. in Computing and Information Sciences from the
Rochester Institute of Technology (RIT). His research lies at the intersection of machine learning and healthcare.
He is interested in developing robust and fair AI models for real-world problems in health and medicine –
primarily to improve models’ generalizability to diverse patient groups. During his Ph.D., he interned at Google
Health and Verisk Analytics. He has published over 20 peer-reviewed research articles, including at premier
conferences such as ICLR, ICDM, MICCAI, and IPMI, and in impactful journals like Nature Medicine, Nature
Communications, IEEE Transactions of Biomedical Engineering, and Computers in Biology and Medicine.

Address:West Virginia University, , Morgantown, West Virginia, United States, 26505