Poster #106 - Pankhuri Singhal

Leveraging Open-Source Large Language Models to Identify Undiagnosed Patients with Rare Genetic Aortopathies

Pankhuri Singhal, PhD1*, Zilinghan Li, MS2*, Ze Yang, MS3*, Tarak Nandi, PhD2, Colleen Morse, PT1, Zachary Rodriguez, PhD1, Alex Rodriguez, PhD2, Volodymyr Kindratenko, DSc3,4, Giorgio Sirugo, MD, PhD1, Reed E. Pyeritz, MD, PhD1, Theodore Drivas, MD1, Ravi Madduri, MS2,5+, and Anurag Verma, PhD1+ 1Department of Medicine, Division of Translational Medicine and Human Genetics, University of Pennsylvania Perelman School of Medicine, Philadelphia, PA, 19104, USA 2Data Science and Learning Division, Argonne National Laboratory, Lemont, IL, 60439, USA 3The Grainger College of Engineering, Department of Electrical and Computer Engineering, University of Illinois Urbana-Champaign, Urbana, IL, 61801, USA 4National Center for Supercomputing Applications, University of Illinois Urbana-Champaign, Urbana, IL, 61801, USA 5Consortium for Advanced Science and Engineering, The University of Chicago, Chicago, IL, 60637, USA *these authors contributed equally to this work.

Rare genetic aortopathies are chronically underdiagnosed due to phenotypic heterogeneity, leading to delayed recognition and potentially fatal outcomes. Although genetic testing enables proactive management, it typically depends on primary care physicians identifying a genetic basis for symptoms, a step often missed in routine practice. Broad-scale, automated methods are needed to flag patients who may benefit from genetic evaluation. Clinical notes, which contain detailed narratives of patient history and presentation, represent a promising substrate for such approaches. Here, we present an open-source, Large Language Model (LLM)-enabled pipeline for genetic testing recommendation that integrates retrieval-augmented generation (RAG) over curated aortopathy corpora to contextualize patient notes against domain-specific knowledge. We validated the pipeline using 22,510 progress notes from 500 individuals (250 confirmed cases and 250 matched controls) in the Penn Medicine BioBank. The system achieved an accuracy of 0.852 and F1-score of 0.844 in categorizing patients as candidates for genetic testing. Unlike traditional workflows that depend on physician-initiated referrals, our framework provides a dynamic, automated recommendation layer that updates as new clinical documentation becomes available. Importantly, the architecture is portable and adaptable: by substituting the domain corpus and prompts, it can be extended to other diagnostic challenges such as early detection of neurodegenerative disease or risk stratification for cardiac adverse events. This study demonstrates the feasibility of transparent, domain-adaptable, and privacy-compliant integration of LLMs into clinical decision support. By enabling earlier identification of at-risk individuals, our framework lays the groundwork for scalable AI-assisted precision medicine and improved patient outcomes in rare disease contexts.