Artificial intelligence meets medical expertise: evaluating GPT-4's proficiency in generating medical article abstracts

Abstract

Purpose: The advent of large language models like GPT-4 has opened new possibilities in natural language processing, with potential applications in medical literature. This study assesses GPT-4's ability to generate medical abstracts. It compares their quality to original abstracts written by human authors, aiming to understand the effectiveness of artificial intelligence in replicating complex, professional writing tasks. Materials and Methods: A total of 250 original research articles from five prominent radiology journals published between 2021 and 2023 were selected. The body of these articles, excluding the abstracts, was fed into GPT-4, which then generated new abstracts. Three experienced radiologists blindly and independently evaluated all 500 abstracts using a five-point Likert scale for quality and understandability. Statistical analysis included mean score comparison inter-rater reliability using Fleiss' Kappa and Bland-Altman plots to assess agreement levels between raters. Results: Analysis revealed no significant difference in the mean scores between original and GPT-4 generated abstracts. The inter-rater reliability yielded kappa values indicating moderate to substantial agreement: 0.497 between Observers 1 and 2, 0.753 between Observers 1 and 3, and 0.645 between Observers 2 and 3. Bland-Altman analysis showed a slight systematic bias but was within acceptable limits of agreement. Conclusion: The study demonstrates that GPT-4 can generate medical abstracts with a quality comparable to those written by human experts. This suggests a promising role for artificial intelligence in facilitating the abstract writing process and improving its quality.