Biomedical and Health Informatics PhD candidate Pedram Golnari, MD, is presenting.
Enhancing Diagnosis Capabilities with Large Language Models
Clinical reasoning is a cornerstone of medical care that uses aniterative process to integrate heterogeneous and potentially limitedinformation to evaluate multiple diagnoses while eliminating incorrect choices.In addition to identifying correct diagnosis, physicians performmulti-dimensional comprehension tasks, such as therapeutics and prognosis, asessential components of medical care. Therole of artificial intelligence (AI) systems has been explored in clinicaldecision making, assisting in clinical data analytics, generating diagnosis,and more recently in differential diagnosis. Generative AI (Gen AI),specifically large language models (LLMs), have demonstrated strongcapabilities in understanding, reasoning, and generating natural languageconstructs in medical context using general purpose models and fine-tuned LLMs.However, the capabilities of LLMs in performing multi-stage diagnostic reasoningacross clinical diagnosis and mechanistic distinction has not been studiedbefore. LLM contextualized using both fine-tuning and prompt tuning havedemonstrated high accuracy over standardized, multiple-choice questions (MCQs)medical benchmark. Onepromising approach to inject domain knowledge into Gen AI models is usingontologies and knowledge graphs (KGs). In biomedicine, ontologies providestandardized vocabularies and relationships for a given domain, while knowledgegraphs represent facts as nodes (entities) and edges (relationships) to createa network of interconnected information. Knowledge graphs have beensuccessfully used to integrate heterogeneous biomedical data and enableadvanced analytics such as knowledge retrieval, reasoning, and hypothesisgeneration. This marriage of ontology-driven structure with content fromliterature yields a knowledge graph that is both comprehensive (covering keydomain concepts) and context-rich (grounded in published evidence). Weintroduce Knowledge Graph-augmented Medical Agent (KroMA), a framework toground generically pre-trained LLMs to perform specialized clinical reasoning.
If unable to attend in person in Biomedical Research Building room 105, you may join via Zoom at
Meeting ID: 958 2937 2435
Passcode: 087450