作者
Jose Arriola‐Montenegro,Andres Chapoñan-Lavalle,Charat Thongprayoon,Jing Miao,Karina Ordaya-Gonzales,Noppawit Aiumtrakul,Natalia Nombera-Aznaran,Iasmina M. Craici,Wisit Cheungpasitporn
摘要
Objectives: Graphical abstracts are increasingly used in nephrology, but their structural organization varies, including the extent to which they exhibit Tree-of-Thought (ToT)- and Graph-of-Thought (GoT)-inspired features. We examined whether greater ToT/GoT-inspired visual alignment was associated with improved information extraction and comprehension by a vision-capable large language model (LLM). Methods: We evaluated 100 graphical abstracts from nephrology journals. Two nephrologists independently scored each using a prespecified six-domain structural rubric (multistep reasoning, branching/exploration, graph-like structure, output clarity, visual reasoning flow, and overall ToT/GoT-inspired alignment; total 0-12). Abstracts were categorized a priori as low (≤3), intermediate (4-6), or high (≥7) structural alignment. A vision-capable LLM (ChatGPT 5.4) analyzed each image using a standardized prompt. Responses were compared with manuscript-derived reference answers and independently graded by two nephrologists who were blinded to structural scores and structural-category assignments during comprehension scoring using a five-domain comprehension rubric (total 0-10). Group differences were assessed with Kruskal-Wallis tests, and the structural-comprehension association with Spearman correlation. Results: Of 100 graphical abstracts, 23 had low, 28 intermediate, and 49 high structural alignment. Median [IQR] total structural scores were 3.0 [3.0-3.0], 5.0 [5.0-6.0], and 9.0 [7.0-10.0], respectively. Median total LLM comprehension scores increased across low-, intermediate-, and high-alignment groups from 4.5 [3.5-5.0] to 6.0 [5.5-7.0] and 10.0 [9.5-10.0], respectively (Kruskal-Wallis H=79.10, P<0.001). Similar graded differences were observed across objective identification, methodology, key findings, interpretation/conclusion, and summary quality (all P<0.001). Total structural score was strongly correlated with total LLM comprehension score (Spearman ρ=0.91, 95% CI 0.87-0.93, P<0.001). Conclusions: Greater alignment with ToT/GoT-inspired visual features was associated with better LLM comprehension of nephrology graphical abstracts. Structured visual organization may facilitate AI-assisted extraction and synthesis of scientific information, but validation across additional multimodal models and medical specialties is needed.