Data Extraction from Unstructured Medical Records with the GPT-4 Chatbot

Background: This study presents a novel approach for using GPT-4 (generative pre-trained transformer 4) as a tool for data extraction from unstructured medical records and assesses its performance on this task.

Methods: Fifty fictitious patient medical records were drafted in German language and GPT-4 was provided with instructions, on how to process each one. Data extraction involved text-mining and classification tasks for nine variables. The accuracy, recall, precision, and F1-Score of the GPT-4 model were assessed for each requested variable.

Results: The accuracy of GPT-4 in data-mining tasks was 100% for all requested variables (patient identifier, EuroSCORE 2, admission and exit date). GPT-4 exhibited the following performance in classification tasks for each requested variable: stroke (accuracy 96%, F1-Score 95.80%), cardiac tamponade (accuracy 100%, F1-Score 100%), pacemaker implantation (accuracy 100%, F1-Score 100%), atrial fibrillation (accuracy 100%, F1-Score 100%) and death (accuracy 100%, F1-Score 100%).

Conclusion: GPT-4 exhibited excellent performance in both text-mining and classification tasks. Reliable data extraction from unstructured medical records seems possible with the GPT-4 chatbot.

 

Scroll to top

Die Autoren geben an, dass kein Interessenkonflikt besteht.

Publikationsverlauf

Artikel online veröffentlicht:
13. Februar 2024