In an ischemic stroke, an artery in the brain is blocked by a blood clot, which deprives brain cells of blood supply. Therefore, doctors need to act quickly and unblock the artery with the help of a catheter. So-called mechanical thrombectomy requires a large amount of data to be recorded and transferred to various registers. In this study, Dr. Nils Lehenen, chief physician at the Diagnostic and Interventional Neuroradiology and Pediatric Neuroradiology Clinic at the University Hospital Bonn (UKB), discovered that ChatGPT could greatly aid in this data transfer. did. This result was published in the specialized journal “RADIOLOGY.''
Mechanical thrombectomy requires a variety of data: when the patient arrived, when the CT scan was performed, when was the first puncture, when blood flow was restored, etc. Must be recorded in the patient report and manually transferred to the patient report. Various registries for clinical outcomes and prospective studies.
“This is a labor-intensive task and is prone to transcription errors,” said Dr. Nils Lehnen, who also conducts research at the University of Bonn. “So we asked ourselves whether an AI like ChatGPT could make this transfer happen faster, and perhaps even more reliably.”
In radiology, ChatGPT has already been tested in a variety of procedures. For example, to simplify reports or answer patient questions during breast cancer screening. However, the ability of ChatGPT to successfully extract data for a database from free-text reports of mechanical thrombectomy and generate clinical data at the same time has been previously unexplored and was the research objective of this new study.
Dr. Lehnen's research group first created German prompts for ChatGPT and tested them on 20 reports to identify errors and adapt the prompts. After modification, data extraction using ChatGPT was tested against his 100 internal reports from UKB. In order to make the best comparison, an experienced neuroradiologist also summarized the results without looking at his ChatGPT assessment. The researchers compared the results and found that ChatGPT correctly extracted 94 percent of the data entries and required no post-processing. Researchers considered only his ChatGPT data entries that exactly matched the expert's data as correct. Deviations such as additional symbols, punctuation, and synonyms were classified as inaccurate.
To verify these results, the researchers tested 30 additional external reports using the same prompts. ChatGPT achieved 90% correct data entry.
This suggests that ChatGPT could be an alternative to manually retrieving this data. However, the report and prompts were created only in German, so the findings may need to be reviewed for other languages. Additionally, poor results are still observed for certain data points, indicating that human oversight is still needed. However, we expect that further optimization of the prompts will further improve the results, and that ChatGPT will make this area easier in the future. ”
Dr. Nils Lehnen, Senior Physician in the Diagnostic and Interventional Neuroradiology and Pediatric Neuroradiology Clinic at Bonn University Hospital (UK)
sauce:
Bonn University Hospital (UKB)
Reference magazines:
Lehnen, NC; other. (2024). Data extraction from free-text reports on mechanical thrombectomy in acute ischemic stroke using ChatGPT: A retrospective analysis. Radiology. doi.org/10.1148/radiol.232741.