FEASIBILITY OF AUTOMATED REGISTRATION TO THE JAPAN TRAUMA DATA BANK USING GENERATIVE AI

Abstract

　　This descriptive study aimed to develop and evaluate a system automating data registration into the Japan Trauma Data Bank (JTDB) using generative AI (GPT-4). The primary objective was to streamline the labor-intensive process of manual data entry for severe trauma cases. We targeted patients admitted to Jichi Medical University Hospital between January and November 2024. Our method involved extracting unstructured text from electronic emergency medical records. This text was then entered into the GPT-4 model via a QR code interface, which converted it into structured data suitable for the JTDB registry. The system's accuracy was evaluated against physician-entered data, serving as the gold standard, across 148 of the 198 official JTDB items. These items were categorized into seven distinct groups for detailed analysis. A pilot evaluation involving 55 cases revealed a high overall accuracy rate of 85.0% (±5.9%). The system performed exceptionally well in categories such as “Demographics (Age/Sex) ” (100%) and “Type of Injury” (98.5%±2.8%). However, it faced challenges with more dynamic data points, with “Vital Signs” showing a lower accuracy of 75.7% (±25.6%). Future work will focus on refining prompt engineering and implementing a human verification step to further improve system reliability.