Tumor documentation for example for the Bavarian cancer registry to date is still done by typists, manually filling the information from the medical report to the cancer registry form. In this work, we explored the potential of a large language model based system (LLM) for automated tumor documentation.

We developed an LLM based method, to read, output and fill in 10 typical tumor documentation related items from medical reports. We anonymized and modified potential identifying information of 10 physician letters. All patients had stage II or III colorectal cancer. We then measured the time it takes a typist to read and copy/paste the information from the medical report to the tumor documentation form. We compared and measured the time it takes for the LLM model alone. Finally, we measured the time it takes a typist to check for errors when the form was already prefilled by the LLM.

The typists manual fill in time was 14:45 minutes, the LLMs time was 0:45 minutes (-94.9%). The typist error check time with LLM prefill was 06:28 minutes (-56.2%). The LLM correctly extracted 100/100 items.

State of the art large language models offer huge potential time saving benefits for automated tumor documentation. Larger studies are needed for evaluation.