Application of Large Language Models in Automated Interpretation of Urodynamic Parameters

Area of Science:

• Urology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Urodynamic studies (UDS) are crucial in urology but require expertise and face interobserver variability.
• Large language models (LLMs) show potential in medical diagnostics, but their use in UDS interpretation is unexplored.

Purpose of the Study:

• To assess the diagnostic performance of LLMs in interpreting urodynamic parameters.
• To compare LLM performance against urologists with varying experience levels.

Main Methods:

• Analysis of 320 urodynamic studies using Deepseek-R1 and GPT-4 LLMs.
• Comparison of LLM diagnostic accuracy with junior and senior urologists.
• Evaluation using ROC curves, AUC, diagnostic accuracy, and the QUEST framework, adhering to TRIPOD+AI guidelines.

Main Results:

• Deepseek-R1 achieved 92.50% accuracy, outperforming GPT-4 (85.94%) and junior urologists (83.75%), nearing senior urologists (95.94%).
• Deepseek-R1 excelled in QUEST framework criteria and showed high clinical utility, particularly in decision support and educational value.
• ROC analysis confirmed strong performance for Deepseek-R1 (AUC 0.89-0.92) and GPT-4 (AUC 0.84-0.88).

Conclusions:

• LLMs, especially Deepseek-R1, show significant potential for automated urodynamic parameter interpretation.
• These AI tools could enhance clinical decision support, training, and quality assurance in urodynamics.
• LLM application may improve diagnostic consistency and access to expert-level interpretation in urology.