This study utilized the upmedic platform, a tool specifically designed for creating structured radiological reports (Fig 1). During the preparatory phase, triggers - pre-defined phrases that describe abnormal findings and can be raised by voice command or mouse click - were automatically generated using large language models (LLMs). These models were trained on datasets containing previously completed radiology reports, received from each radiologist. A set of individual triggers was automatically generated for each radiologist. Subsequently, radiologists reviewed the triggers, with the option to accept, edit, or create new suggestions (Fig. 2). This iterative process ensured that the final set of triggers reflected both the standardized language of LLMs and the practical expertise of radiologists. In the final phase, 10,154 reports prepared by 22 radiologists who actively incorporated the triggers into their routine reporting workflows were analyzed.

The use of triggers was compared against two conventional methods: manual keyboard entry (67,360 reports) and voice dictation (21,664 reports). Metrics for comparison included: mean words per minute and mean characters per minute entered into the report with a specific method.