Comparison of an AI-Generated Case Report With a Human-Written Case Report: Practical Considerations for AI-Assisted Medical Writing

Denver S Pinto; Sharon M Noronha; Gaurav Saigal; Robert M Quencer

doi:10.7759/cureus.60461

Comparison of an AI-Generated Case Report With a Human-Written Case Report: Practical Considerations for AI-Assisted Medical Writing

Cureus. 2024 May 16;16(5):e60461. doi: 10.7759/cureus.60461. eCollection 2024 May.

Authors

Denver S Pinto¹, Sharon M Noronha², Gaurav Saigal³, Robert M Quencer³

Affiliations

¹ Radiology, Jackson Memorial Hospital, Miami, USA.
² Anesthesiology, St. John's Medical College, Bangalore, IND.
³ Radiology, University of Miami Miller School of Medicine, Miami, USA.

Abstract

Introduction: The utility of ChatGPT has recently caused consternation in the medical world. While it has been utilized to write manuscripts, only a few studies have evaluated the quality of manuscripts generated by AI (artificial intelligence).

Objective: We evaluate the ability of ChatGPT to write a case report when provided with a framework. We also provide practical considerations for manuscript writing using AI.

Methods: We compared a manuscript written by a blinded human author (10 years of medical experience) with a manuscript written by ChatGPT on a rare presentation of a common disease. We used multiple iterations of the manuscript generation request to derive the best ChatGPT output. Participants, outcomes, and measures: 22 human reviewers compared the manuscripts using parameters that characterize human writing and relevant standard manuscript assessment criteria, viz., scholarly impact quotient (SIQ). We also compared the manuscripts using the "average perplexity score" (APS), "burstiness score" (BS), and "highest perplexity of a sentence" (GPTZero parameters to detect AI-generated content).

Results: The human manuscript had a significantly higher quality of presentation and nuanced writing (p<0.05). Both manuscripts had a logical flow. 12/22 reviewers were able to identify the AI-generated manuscript (p<0.05), but 4/22 reviewers wrongly identified the human-written manuscript as AI-generated. GPTZero software erroneously identified four sentences of the human-written manuscript to be AI-generated.

Conclusion: Though AI showed an ability to highlight the novelty of the case report and project a logical flow comparable to the human manuscript, it could not outperform the human writer on all parameters. The human manuscript showed a better quality of presentation and more nuanced writing. The practical considerations we provide for AI-assisted medical writing will help to better utilize AI in manuscript writing.

Keywords: average perplexity score; burstiness score; chatgpt; comparison with human writing; hypothyroidism.