Kidney Week

Abstract: FR-PO0025

Tree-of-Thought and Graph-of-Thought Structure Improves Large Language Model Comprehension of Nephrology Graphical Abstracts

Session Information

• Artificial Intelligence: Clinical Prediction, Decision Support, and Digital Kidney Care
  October 23, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Arriola Montenegro, Jose J., Mayo Clinic Division of Nephrology and Hypertension, Rochester, Minnesota, United States

Jose J. Arriola Montenegro, MD

• Ordaya Gonzales, Karina M., Mayo Clinic Division of Nephrology and Hypertension, Rochester, Minnesota, United States

Karina M. Ordaya Gonzales, MD

• Aiumtrakul, Noppawit, Mayo Clinic Division of Nephrology and Hypertension, Rochester, Minnesota, United States

Noppawit Aiumtrakul, MD

• Nombera-Aznaran, Natalia R., Department of Nephrology, University of Alabama at Birmingham, Birmingham, Alabama, United States

Natalia R. Nombera-Aznaran, MD

• Thongprayoon, Charat, Mayo Clinic Division of Nephrology and Hypertension, Rochester, Minnesota, United States

Charat Thongprayoon, MD, FASN

• Miao, Jing, Mayo Clinic Division of Nephrology and Hypertension, Rochester, Minnesota, United States

Jing Miao, MD, PhD, FASN

• Craici, Iasmina, Mayo Clinic Division of Nephrology and Hypertension, Rochester, Minnesota, United States

Iasmina Craici, MD

• Cheungpasitporn, Wisit, Mayo Clinic Division of Nephrology and Hypertension, Rochester, Minnesota, United States

Wisit Cheungpasitporn, MD, FASN

Background

Graphical abstracts are increasingly used in nephrology to convey complex clinical and research information visually. However, their design varies, particularly in the use of structured reasoning frameworks such as Tree-of-Thought (ToT) and Graph-of-Thought (GoT). Their impact on interpretation by vision-capable LLMs is unclear.

Methods

We did an experimental AI study using 100 graphical abstracts from nephrology journals. Each was independently annotated by 2 nephrologists using a rubric across 6 domains (score 0–12):multistep reasoning, branching, graph structure, output clarity, visual flow, and ToT/GoT alignment. Abstracts were categorized as Control (≤3), Ambiguous (4–6), or ToT/GoT-High (≥7). A vision-capable LLM (ChatGPT, 5.4) analyzed each abstract using standardized prompts to identify objective, methods, findings, and conclusions.Responses were compared with gold-standard answers and graded by two blinded nephrologists.Data were summarized with medians and compared using Kruskal–Wallis.

Results

Of 100 abstracts, 23 were Control, 28 Ambiguous, and 49 ToT/GoT-High. Scores differed (median[IQR]:3 [3–3], 5 [5–6], 9 [8–10]; p<0.001), showing increasing complexity. ToT/GoT-High abstracts showed greater branching, graph organization, and visual flow. Control abstracts lacked these features. With ChatGPT, higher ToT/GoT alignment was associated with improved comprehension,accurate identification of objectives, methods, and conclusions. Lower-alignment abstracts produced incomplete responses.

Conclusion

Graphical abstracts incorporating structured reasoning (ToT/GoT) are associated with improved interpretability by vision-capable LLMs. Branching and graph organization enhance comprehension.These findings support integrating structured reasoning into graphical abstract design to optimize human and AI understanding.