5403. Visual Communication of Traumatic Injuries by Automatic Rendering of a Graphic Summary Based on Natural Language Processing of CT Reports
Authors* Denotes Presenting Author
Nitai Bar *;
Rambam Healthcare Campus
Anat Illivitzki;
Rambam Healthcare Campus
Eyal Bercovich;
Rambam Healthcare Campus
Objective:
Initial medical response in traumatic injury scenarios requires a multidisciplinary approach involving radiologists, as well as trauma surgeons, emergency medicine physicians, and lead administrators. In such settings, immediate and efficient data flow is essential, but nevertheless remains a major challenge. This is particularly pronounced in mass casualty incidents, where triage, patient navigation, and team coordination become critical. We aim to develop a tool leveraging an advanced language model to automatically extract and classify findings from CT reports, and channel data to achieve real-time rendering of a graphical summary of traumatic findings. Graphical illustrations are an efficient means of communicating complex data, which is required for prioritization of patients and supports life-saving decisions taken under time pressure in the trauma room.
Materials and Methods:
Our dataset comprised all chest CTA trauma protocol scans performed between 1.1.2012 – 1.4.2022, at a tertiary trauma center in the north of Israel. Traumatic injuries were manually annotated to produce “gold-standard” labels. GPT models were subsequently used to extract meaningful data from the reports and classify them into predefined categories, enabling a rule-based automatic rendering of a graphical summary illustration using open-source graphical tools. Performance of the language model was assessed with mean AUC, F1, and exact match scores as compared to the gold standard labels.
Results:
Preliminary findings suggest exact match scores > 80% and F1 scores > 90%, demonstrating the model's ability to accurately predict and classify traumatic injury labels, including rare ones. The NLP model thus allows real-time generation of structured data, which can be utilized to streamline communication between teams, promote notification of acute findings and ensure adequate patient prioritization.
Conclusion:
We introduce a pipeline harnessing natural language models and graphic tools to allow real-time data analysis. Graphical illustrations of medical data based on automatically generated classifiers from medical reports are a promising novel tool contributing to optimize the trauma workflow.
