Leveraging Large Language Models to Generate Clinical Histories for Oncologic Imaging Requisitions.

Area of Science:

• Artificial Intelligence in Radiology
• Clinical Informatics
• Oncologic Imaging

Background:

• Physician-provided histories on oncologic imaging requisitions often lack crucial details, hindering accurate interpretation.
• Clinical information is vital for improving the interpretation of oncologic imaging studies.

Purpose of the Study:

• To evaluate the efficacy of large language models (LLMs) in automatically generating clinical histories for oncologic imaging requisitions.
• To compare LLM-generated histories against original requisition histories for completeness and quality.

Main Methods:

• A multidisciplinary team identified 10 key parameters for oncologic imaging history.
• GPT-4 was used to generate structured clinical histories from electronic health record clinical notes for 200 patients.
• LLM performance was assessed using recall, precision, and F1 scores; histories were compared for completeness and radiologist preference.

Main Results:

• GPT-4 demonstrated high performance (F1 = 0.983) in extracting oncologic parameters from clinical notes.
• LLM-generated histories were significantly more complete, including primary oncologic diagnosis (99.5% vs 89%) and acute symptoms (15% vs 4%).
• Radiologists preferred LLM-generated histories (89%) for interpretation, citing improved completeness and lower perceived harm.

Conclusions:

• Large language models can accurately automate the generation of clinical histories for oncologic imaging from clinical notes.
• LLM-generated histories offer superior completeness and are preferred by radiologists, enhancing imaging interpretation and patient safety.
• This technology has the potential to streamline the requisition process and improve the quality of oncologic imaging.