Research Focus June 30 2026
Research Focus: Parameter-Efficient Fine-Tuning of MedGemma-27B-text-it for Structured Variable Extraction from Surgical Pathology Reports
Dr. Jihad Obeid (Left) is a Professor and SmartState Endowed Chair in Biomedical Informatics in the Department of Public Health Sciences at the Medical University of South Carolina (MUSC). He serves as the site PI for ADAPT in SC at MUSC. Dr. Chad Arledge (Middle) is a post-doctoral scholar on ADAPT in SC at MUSC. Dr. Paul Heider (Right) is an Assistant Professor in Biomedical Informatics at MUSC. The team’s research focuses on navigating the complex landscape of clinical data within electronic health records and leveraging AI to characterize patients and improve downstream processes such as research recruitment, quality improvement and public health surveillance.
Surgical pathology reports contain rich, clinically meaningful information that is critical for cancer research, registry reporting, and downstream decision support. However, these reports are typically composed of unstructured free text and writing styles vary widely across pathologists, making structured and scalable variable extraction challenging. As a result, clinical teams and cancer registries rely on extensive manual review of pathology reports to identify, extract, and encode clinically meaningful variables.
In this work, the use of MedGemma-27B-text-it, an openly available specialized large language model (LLM) from Google optimized for medical text comprehension, is investigated for automated structured variable extraction from surgical pathology reports.MedGemma is then fine-tuned using quantized low-rank adaptation (QLoRA) on a curated dataset of lung and breast cancer patients at the Medical University of South Carolina annotated with cancer registry variables, including primary site, laterality, behavior, grade, histology type, and staging.
Following fine-tuning, MedGemma demonstrates consistent improvements in extraction accuracy across all evaluated cancer registry variables compared to the baseline MedGemma model. These findings highlight the utility of parameter-efficient fine-tuning for clinical information extraction from heterogeneous, unstructured pathology narratives.
More broadly, this work illustrates how domain-adapted clinical language models can reduce the burden of manual abstraction and support more timely, consistent capture of high-value oncology data. By enabling scalable extraction from routine pathology narratives, this approach has the potential to accelerate cancer research, strengthen registry reporting, and advance learning health system infrastructure.
