Assessing AI Simplification of Medical Texts: Readability and Content Fidelity

Abstract

Introduction

The escalating complexity of medical literature necessitates tools to enhance readability for patients. This study aimed to evaluate the efficacy of ChatGPT-4 in simplifying neurology and neurosurgical abstracts and patient education materials (PEMs) while assessing content preservation using Latent Semantic Analysis (LSA).

Methods

A total of 100 abstracts (25 each from Neurosurgery, Journal of Neurosurgery, Lancet Neurology, and JAMA Neurology) and 340 PEMs (66 from the American Association of Neurological Surgeons, 274 from the American Academy of Neurology) were transformed by a GPT-4.0 prompt requesting a 5th grade reading level. Flesch-Kincaid Grade Level (FKGL) and Flesch Reading Ease (FKRE) scores were used before/after transformation. Content fidelity was validated via LSA (ranging 0–1, 1 meaning identical topics) and by expert assessment (0–1) for a subset (n = 40). Pearson correlation coefficient compared assessments.

Results

FKGL decreased from 12th to 5th grade for abstracts and 13th to 5th for PEMs (p < 0.001). FKRE scores showed similar improvement (p < 0.001). LSA confirmed high content similarity for abstracts (mean cosine similarity 0.746) and PEMs (mean 0.953). Expert assessment indicated a mean topic similarity of 0.775 for abstracts and 0.715 for PEMs. The Pearson coefficient between LSA and expert assessment of textual similarity was 0.598 for abstracts and −0.167 for PEMs. Segmented analysis of similarity correlations revealed a correlation of 0.48 (p = 0.02) below 450 words and a −0.20 (p = 0.43) correlation above 450 words.

Conclusion

GPT-4.0 markedly improved the readability of medical texts, predominantly maintaining content integrity as substantiated by LSA and expert evaluations. LSA emerged as a reliable tool for assessing content fidelity within moderate-length texts, but its utility diminished for longer documents, overestimating similarity. These findings support the potential of AI in combating low health literacy, however, the similarity scores indicate expert validation is crucial. Future research must strive to improve transformation precision and develop validation methodologies.