TOWARDS ROBUST AND UNBAISED AI FOR HEALTHCARE

About The Speaker

Dr. Prashnna K Gyawali serves as an Assistant Professor in the Lane Department of Computer Science and Electrical Engineering at West Virginia University. Prior to this, he distinguished himself as a postdoctoral scholar at the School of Medicine at Stanford University. Dr. Gyawali earned his Ph.D. in Computing and Information Sciences from the Rochester Institute of Technology (RIT) and completed his Bachelor's in Engineering at IOE Pulchowk Campus in Nepal. His research operates at the critical junction of machine learning and healthcare. He is dedicated to developing robust and fair AI models that address real-world health and medical challenges, with a particular emphasis on enhancing the generalizability of these models across diverse patient populations. During his Ph.D., he gained practical experience through internships at Google Health and Verisk Analytics. Dr. Gyawali's scholarly contributions are notable, including authoring over 20 peer-reviewed research articles. His work has been featured at premier conferences such as ICLR, ICDM, MICCAI, and IPMI, and published in prestigious journals like Nature Medicine, Nature Communications, and Nature Communication Biology. His research endeavors have been supported by funding from prestigious sources including WVHPEC, DARPA, and NSF CITeR.


Lecture Abstract 

Over the past decade, AI algorithms, primarily driven by supervised deep learning, have achieved remarkable success across various domains such as computer vision and natural language processing (NLP). However, replicating this success in the health and medicine sector is more challenging, primarily due to the existing models' limited ability to generalize across unseen data, like new patient cohorts. The first challenge in healthcare AI is the high cost and complexity of curating well-annotated training datasets. Additionally, in healthcare data, varying attributes such as a patient’s age, sex, race, and anatomical differences often intersect with the predictive task, leading to disparities in model performance across patient subgroups. Moreover, the application of AI models in real-world clinical settings necessitates robust methods for explanation and interpretation. In my talk, I will delve into these vital aspects of healthcare AI. Initially, I will present my research on developing unbiased deep learning models through the disentanglement of representations. This will be followed by a discussion on the critical role of interpretability in healthcare and the challenges faced by current deep learning models in this context. I will then introduce my recent work aimed at overcoming some of these interpretability challenges. Lastly, I will discuss my group’s ongoing projects in multimodal learning, highlighting the potential advancements and future opportunities in this area, including the exploration of foundation models and federated learning. This presentation aims to shed light on the intricacies of healthcare AI and the innovative approaches we are pursuing to address its unique challenges. 


Details of the Program: 

Date: Jan 4, 2024, Thursday 

Time: 3:30-5:00 pm 

Venue: Alliance Francaise, Dhobighat Lalitpur 


Participants Demography: 

Total Present: 30 

Students: 13

Professionals: 8 

NAAMII Alumni: 3

Facebook Live views during the event: 993 

Volunteers present: 10 


Flow of the session:  

Registration began: 3:15pm 

Lecture: 3:30-4:15 

Q&A: 4:15-4:30 

Tea break and Networking: 4:40-5 


Presentation slide: NAAMI_seminar_Gyawali.pptx 

Youtube Video:  Coming soon …