Large Language Model Simplification of Open Access Pediatric Strabismus Literature: Cross-Sectional Validation of Readability and Clinical Fidelity.

Background: Peer-reviewed medical literature consistently violates established health literacy readability targets, creating a gap that effectively excludes patients and caregivers from accessing evidence-based information. Objective: This study aimed to evaluate whether a large language model (LLM) can generate plain-language summaries of pediatric strabismus literature while preserving clinical fidelity and meeting established health literacy readability targets. Methods: This cross-sectional study analyzed 85 open access, peer-reviewed pediatric strabismus articles published between 2022 and 2025, stratified by strabismus subtype, surgical relevance, and publication type. Full-text articles were processed using DeepSeek-V3 (DeepSeek) via a structured prompt, which instructed the model to provide a simplified summary meeting the following requirements for each article: a seventh-grade or lower reading level, a maximum length of 800 words, and strict preservation of medically significant data. Primary outcomes were readability scores measured by the Flesch-Kincaid Grade Level (FKGL) and Simple Measure of Gobbledygook (SMOG) indices. Secondary outcomes included clinical fidelity, which was independently assessed by 2 fellowship-trained pediatric strabismus specialists. Results: Baseline articles demonstrated a mean FKGL score of 15.79 (SD 1.53) and a mean SMOG score of 14.41 (SD 1.09). Following LLM simplification, the mean FKGL score significantly decreased from 15.79 (SD 1.53) to 7.84 (SD 1.30), representing a mean difference of 7.95 (95% CI 7.52-8.38; P<.001). Similarly, the mean SMOG score decreased from 14.41 (SD 1.09) to 7.68 (SD 0.94), representing a mean difference of 6.73 (95% CI 6.42-7.04; P<.001). Postsimplification readability did not differ significantly by strabismus subtype or surgical relevance (all adjusted P>.05). However, case reports retained slightly higher FKGL scores (mean 8.35, SD 0.89) compared to reviews (mean 7.89, SD 1.37) and original research (mean 7.48, SD 1.40) (adjusted P=.003). Out of the 85 summaries, clinical fidelity was rated good in 81 (95.29%), moderate in 4 (4.71%; these were exclusively summaries of review articles), and poor in 0 (0%). Conclusions: DeepSeek-V3 effectively reduced the reading level of complex pediatric strabismus literature by approximately 8 grade levels, achieving National Institutes of Health-recommended eighth-grade or lower targets without compromising clinical accuracy. When integrated with clinician oversight, LLM-generated summaries offer a scalable, equitable tool to enhance health literacy and support shared decision-making for patients and caregivers.