This multicenter retrospective study aims to evaluate the diagnostic and therapeutic performance of three large language models-ChatGPT, Gemini and Deepseek-using 800 archived inpatient medical records from urology departments across four tertiary hospitals. The study will focus on the accuracy and applicability of these models in disease recognition, preliminary diagnosis and treatment recommendation generation, in order to explore their potential value and limitations in supporting clinical decision-making in real-world settings.
The case data is sourced from the four hospitals involved in the study, with complete and authentic diagnosis and treatment records.
Patients must be 18 years or older, with no gender restrictions.
Complete medical records, including the following core information: patient' s basic information, present illness history, past medical history, physical examination, and auxiliary examinations (including laboratory and imaging tests).
A clear discharge diagnosis and treatment plan (including therapeutic measures and follow-up arrangements).
Medical records with significant missing information, such as key clinical details (present illness history, diagnostic or treatment records, etc.).
Cases where the diagnosis or treatment plan is unclear, or where treatment has not been fully completed for an initial diagnosis.
Cases where the primary diagnosis is not urological.
Cases with major errors or inconsistencies in the records that could affect further assessment.
The study population was drawn from the following institutions: The First Affiliated Hospital of Fujian Medical University, The Second Affiliated Hospital of Fujian Medical University,Shishi City Hospital and Shaowu City Hospital
Cohort
Retrospective
Large Language Model Assessment (ChatGPT, Gemini, DeepSeek)
Other interventionDe-identified inpatient medical records were retrospectively collected from the urology departments of four tertiary hospitals (200 cases per site, 800 in total). Each case included standardized clinical information such as demographics, chief complaint, history of present illness, past medical history, physical examination, laboratory and imaging findings, discharge diagnosis and treatment plan. To simulate the role of an AI system in a "first-visit physician" scenario, all diagnostic conclusions, differential diagnoses and treatment plans were removed before being input into the models. Three large language models (ChatGPT, Gemini and DeepSeek) were prompted with a standardized instruction: "Based on the above clinical information, provide your preliminary diagnosis, differential diagnoses and treatment recommendations." Each model generated outputs including (i) primary and secondary diagnoses, (ii) differential diagnosis lists with reasoning and (iii) preliminary treatment suggesti
Diagnostic Accuracy: Assessed by Top-1 accuracy
Top-1: Proportion of cases where the model's first diagnosis matches the true primary diagnosis.
Diagnostic Accuracy: Assessed by Top-3 accuracy
Top-3: Proportion of cases where the true diagnosis appears in the model's top 3.
Diagnostic Completeness
Proportion of the model's diagnoses that overlap with all diagnoses (primary and secondary) in the case.
Differential Diagnosis Quality
Evaluated by experts using a Likert 5-point scale, considering factors like common disease coverage, logical clarity, and specificity
Click on the lead sponsor to view all of their trials.
