Editing Physicians' Responses Using GPT-4 for Academic Research

Abstract

The integration of Artificial Intelligence (AI) into digital healthcare, particularly in the anonymisation and processing of health information, holds considerable potential.

Objectives: To develop a methodology using Generative Pre-trained Transformer (GPT) models to preserve the essence of medical advice in doctors' responses, while editing them for use in scientific studies.

Methods: German and English responses from EXABO, a rare respiratory disease platform, were processed using iterative refinement and other prompt engineering techniques, with a focus on removing identifiable and irrelevant content.

Results: Of 40 responses tested, 31 were accurately modified according to the developed guidelines. Challenges included misclassification and incomplete removal, with incremental prompting proving more accurate than combined prompting.

Conclusion: GPT-4 models show promise in medical response editing, but face challenges in accuracy and consistency. Precision in prompt engineering is essential in medical contexts to minimise bias and retain relevant information.

Keywords: Artificial Intelligence; Data Anonymisation; Medical Informatics; Natural Language Processing.