Accurate clinical histories are essential for oncologic imaging, providing radiologists with the necessary context to interpret scans effectively. The details included in imaging requisitions influence study triage, protocol selection and diagnostic accuracy. However, despite their importance, requisitions often lack key information, such as the primary oncologic diagnosis, prior treatments and acute symptoms. Incomplete histories can lead to suboptimal imaging protocols, missed findings and increased diagnostic uncertainty.
At the same time, electronic health records (EHRs) contain comprehensive clinical notes, but these documents are often too lengthy or poorly structured for radiologists to review efficiently. The disconnect between the wealth of available data and the limited information included in requisitions represents a significant challenge in oncology care. Large language models (LLMs) offer a promising solution by extracting and structuring clinical histories from EHRs, making the most relevant details accessible to radiologists. Recent research evaluating GPT-4’s performance in generating oncologic imaging histories has demonstrated that AI-generated summaries are significantly more complete than those provided by ordering physicians. LLM-generated histories include essential diagnostic and treatment information more frequently, are preferred by radiologists and may improve patient safety by reducing the risk of misinterpretation.
The Role of Clinical Histories in Imaging Accuracy
Clinical histories provide radiologists with critical background information, allowing them to tailor imaging protocols and refine diagnostic assessments. Essential details include the patient’s primary oncologic diagnosis, prior treatments such as chemotherapy or radiation, recent surgical history and any acute or worsening symptoms. The presence of these details can improve sensitivity in tumour detection, reduce missed findings and enhance overall imaging accuracy. Without a clear and structured history, radiologists may have to rely on assumptions, increasing the risk of diagnostic errors.
Despite their recognised importance, requisitions often fail to include sufficient clinical information. This issue has been widely documented in the literature, with radiologists referring to the lack of detailed histories as a persistent challenge in their field. Studies have shown that inadequate requisition histories contribute to suboptimal imaging protocol selection and may impact patient safety. Conversely, when clinical information is complete and structured, radiologists are better equipped to interpret imaging findings accurately.
One of the major obstacles to improving requisition histories is the practical challenge of accessing and summarising relevant details from lengthy EHR notes. Oncology patients often have extensive medical records containing valuable but unstructured information. Reviewing these records manually is time-consuming and impractical for radiologists, particularly in high-volume clinical settings. The gap between available data and usable information necessitates a more efficient approach to clinical documentation.
Large Language Models as a Solution
LLMs have demonstrated strong capabilities in information extraction and summarisation, making them well-suited for generating structured clinical histories from unstructured EHR notes. In a recent study, GPT-4 was evaluated for its ability to extract key oncologic details and generate concise but comprehensive requisition histories. The results showed that AI-generated histories included essential diagnostic and treatment parameters more frequently than those provided by physicians.
Specifically, LLM-generated histories more often contained information about the primary oncologic diagnosis, acute symptoms, prior surgeries and active cancer treatments. The accuracy of AI-generated histories was validated using recall, precision and F1 scores, demonstrating a high level of performance. Importantly, radiologists overwhelmingly preferred LLM-generated histories over traditional requisitions, citing their completeness, clarity and potential to reduce misinterpretation risks.
The study also assessed whether AI-generated histories could improve patient safety. Radiologists indicated that original physician-provided requisitions had a significantly higher perceived risk of harm due to missing or incomplete information. In contrast, LLM-generated histories were rated as more reliable for guiding imaging interpretation and protocol selection. These findings suggest that AI-assisted documentation could enhance radiology workflows by ensuring that key details are consistently included in requisition histories.
Despite these promising results, there are challenges to implementing LLM-generated histories in clinical practice. The use of AI in medical documentation requires careful validation to ensure accuracy and reliability. Additionally, integrating LLMs with existing health IT systems will require technical and regulatory considerations. While AI-generated histories can enhance efficiency, oversight from medical professionals remains essential to verify outputs and mitigate potential errors.
Implications for Clinical Practice
The integration of LLMs into oncologic imaging workflows could lead to significant improvements in radiology practice. By automating the extraction of clinical details from EHRs, LLMs provide radiologists with more structured and complete histories, reducing the likelihood of missed information. Enhanced requisition histories may improve imaging protocol selection, increase diagnostic accuracy and ultimately contribute to better patient outcomes.
One of the key advantages of AI-generated histories is their potential to standardise documentation practices. Variability in physician-provided histories can lead to inconsistencies in imaging requisitions, whereas LLMs can ensure that all relevant parameters are included consistently. Standardised clinical histories may also support quality improvement initiatives by reducing errors related to incomplete documentation.
However, clinical implementation requires further evaluation in real-world settings. Prospective studies are needed to assess the impact of AI-generated histories on clinical decision-making and patient outcomes. Additionally, considerations around physician workflows, AI oversight and regulatory compliance must be addressed before widespread adoption. While LLM-generated histories offer a compelling solution to long-standing challenges in oncologic imaging, their integration should be approached with careful planning to maximise benefits while mitigating risks.
LLMs represent a transformative innovation in oncologic imaging, addressing the long-standing issue of incomplete requisition histories. By leveraging AI-generated summaries, radiologists gain access to more complete and structured clinical information, improving diagnostic accuracy and patient safety. Research findings indicate that LLM-generated histories are significantly more detailed than traditional requisitions, preferred by radiologists and perceived as less likely to contribute to diagnostic errors.
Despite challenges related to implementation and oversight, the potential benefits of AI-assisted documentation are substantial. The ability of LLMs to extract and summarise relevant details from clinical notes provides a scalable solution to enhance communication between referring physicians and radiologists. Further studies will be essential to refine AI’s application and assess its impact in real-world clinical settings. If successfully integrated, LLM-generated clinical histories could become a standard tool in oncologic imaging, improving efficiency, accuracy and overall patient care.
Source: Radiology
Image Credit: iStock
