INFORMS-Pittsburgh Beyond Accuracy: Reliable AI for Healthcare Across Shifts and Silos
Who: Prashnna K Gyawali, Ph.D.Assistant Professor Computer Science and Electrical Engineering, West Virginia University
Abstract: Artificial intelligence has achieved strong predictive performance across healthcare applications, yet real-world deployment remains difficult because healthcare data rarely match the assumptions of conventional machine learning. Models encounter distribution shifts across devices, institutions, patient populations, and acquisition conditions, often leading to failures that conventional confidence scores do not reveal. At the same time, relevant data remain distributed across institutional silos and cannot always be centrally pooled. This talk will present our work toward reliable healthcare AI under these two interconnected challenges. I will first discuss methods for detecting out-of-distribution cases by examining the structure of learned representations, followed by explainable AI approaches that help identify the concepts and evidence driving model predictions and failures. I will then discuss why addressing these challenges requires learning across decentralized data and present our work on federated learning under institutional, label, and multimodal heterogeneity. Together, these efforts illustrate a path toward healthcare AI that can recognize unfamiliar conditions, explain its decisions, and learn collaboratively without requiring centralized access to sensitive data.
Speaker bio: Prashnna Gyawali is an Assistant Professor in the Lane Department of Computer Science and Electrical Engineering at West Virginia University (WVU). Prior to joining WVU, he completed his postdoctoral training at Stanford University and earned his Ph.D. in Computer Science from the Rochester Institute of Technology (RIT). His research focuses on developing reliable and trustworthy AI systems, with an emphasis on improving the generalization, robustness, and interpretability of deep learning models. His work spans self-supervised learning and foundation models, out-of-distribution detection, and explainable AI, with a particular focus on high-impact applications in healthcare and medical imaging.
INFORMS-Pittsburgh is the Pittsburgh Chapter of the Institute for Operations Research and Management Sciences (http://www.informs.org) INFORMS advances and promotes the science and technology of decision making through an array of highly cited publications, conferences, competitions, networking communities, and professional development services. For those in the Pittsburgh region, INFORMS-Pittsburgh desires to be your professional home and a source of recourses for professional development and growth. Join us on our LinkedIn Page (https://www.linkedin.com/groups/8901129/) or on INFORMS Connect Pittsburgh Chapter page (https://connect.informs.org/pittsburgh/home)