Fuse Presents Poster at ACRO Examining AI’s Impact on Documentation

At the Radiation Oncology Summit ACRO 2025 In Las Vegas, Fuse Oncology experts shared a poster presentation on the impact of automation using artificial intelligence (AI) on note taking in radiation oncology.

Radiation oncology clinicians are subject to significant administrative burden, greatly impacting workflow in clinics nationwide. By automating the history of present illness (HPI) section of consultation notes, the researchers hypothesized that AI could improve efficiency while maintaining—or even enhancing—documentation quality.

Fuse’s Yusuf Elnady, Christel Smith, PhD, Lauren Mancuso, BS, RT(R)(T), Matthew Terry, MBA, BSRT (R)(T), and Christopher D. Jahraus, MD from Generations Radiotherapy and Oncology in Alabaster, Ala., utilized a commercial and HIPAA-compliant large language model (LLM) to write HPI with structured data extracted from source documentation in a commercial radiation oncology AI documentation system. The system examines batches of intake documentation from initial consults or scanned document images—such as referral documents, and pathology, lab and imaging reports—and they applied customized optical character recognition (OCR), then extracted relevant structured data to inform the HPI.

To validate accuracy, the researchers evaluated two separate aspects of the workflow for 13 prostate patients: Quality of the data extracted via the OCR system, and quality of the AI-generated HPI. After comparing the AI-generated HPIs to the patient record of the previously treated prostate patients, the researchers found that of 1472 data elements extracted, the system was accurate 97.2% of the time.

To ensure quality of the AI-generated HPI, they used six common LLM metrics, determining the following with results: faithfulness (strong factual consistency), hallucination (<1%, ensuring minimal misinformation), bias (0%, confirming neutrality), tone alignment (100% compliance), keyword presence (84% match rate, subject to source material variability), and answer relevance (97.3%, ensuring precise and relevant documentation).

The AI-driven workflow proved to excel in extracting structured data from diverse document types, including scanned and faxed images. Even when scanned faxed images of the documents were of poor quality, the OCR was effective in identifying the important data elements. The poster demonstrated that the HPI generator consistently achieved industry-standard quality across the above six key metrics.

These results provide an opportunity to streamline HPI and consult preparation using AI to reduce documentation time and uphold quality. Doing so frees up clinicians to spend less time on administrative duties and more time on patient care, a finding that was exciting to share with ACRO attendees. The team from Fuse hopes to expand the study to include more disease sites and test cases, incorporating human oversight to refine and verify extracted data for future clinical use.