When
Tuesday, February 24, 2026 at 11:00 a.m.
Chongle Pan
Professor of Computer Science and Biomedical Engineering
University of Oklahoma
"Trustworthy AI for Health: Reference-Anchored Reasoning, Uncertainty-Quantified Benchmarking and Interpretable Genomics"
ECE 530 | Zoom link
Abstract: This seminar will present a trustworthiness-focused research program in AI for health, spanning reasoning LLM training, uncertainty-quantified evaluation for medical imaging and interpretable predictive genomics. In the first section, we strengthened the trustworthiness and accuracy of reasoning-LLM predictions by training chain-of-thought to align with trusted reference traces. The ARIX algorithm anchored model reasoning to reference traces by combining reinforcement learning with contrastive learning. On a clinical-trial eligibility assessment task, ARIX-trained reasoning traces provided an auditable form of explainability and improved assessment accuracy. In the second section, we improved evaluation reliability through the NACHOS framework, which integrated nested cross-validation, automated hyperparameter optimization and high-performance computing to reduce leakage and quantify variance in test performance estimation. This framework supported our development of Octascope, a lightweight OCT foundation model pre-trained via curriculum learning to improve transfer across tissue types and downstream tasks. Finally, we advanced predictive genomics with multi-task learning for polygenic risk estimation of many cancer types. Model interpretation identified salient genetic variants and uncovered shared genetic structure across cancers. Collectively, these studies highlight the need for coordinated investments in data collection, scalable compute and transdisciplinary collaboration to translate trustworthy AI into solutions for grand challenges in health.
Bio: Chongle Pan is a professor of computer science and biomedical engineering at the University of Oklahoma (OU), founding director of OU’s AI for Health Center, and associate director of the Data Institute for Societal Challenges (DISC). His lab develops trustworthy AI and high-performance computing (HPC) methods for health and biomedical discovery, spanning reasoning LLM for EHR data, computer vision for medical imaging and multi-omics bioinformatics. Pan has authored 100+ peer-reviewed papers (H-index 44) and holds eight patents/applications and he has led or co-led NIH, DoD and DOE projects totaling ~$18M overall (~$6M to his lab) since 2018. Pan has mentored 12 PhD students and 8 postdocs and teaches courses in parallel programming and data science.