Pre-diagnosis Diagnosis Treatment (PDT): A Novel Architecture for Diagnosis and Treatment-option Generation in Rare Castleman Diseases based on GPT4o

Background The application of machine learning techniques in disease diagnosis is advancing rapidly, especially for common diseases common diseases with large datasets. However, these techniques often struggle with rare diseases due to data scarcity. With recent technology advancements, large language models (LLMs) can provide higher diagnostic accuracy and personalized treatment recommendations without extensive training. However, due to limited publicly available data for models such as GPT and Llama, these models often lack information on rare diseases. Methods A Pre-diagnosis Diagnosis Treatment (PDT) architecture was proposed specifically for Castleman disease (CD), a rare lymphoproliferative disorder. For PDT architecture, a pre-diagnosis was machine first employed to identify potential CD patients in different departments. Subsequently, diagnostic and treatment agents enriched with specialized knowledge of CD, generate accurate diagnoses and personalized-treatment plans. PDT was tested on 10 CD patients and 10 non-CD patients, and the results were presented to five clinicians for evaluation. Results Interpretation PDT employs the RAG mechanism with LLMs for accurate CD diagnosis and personalized treatment, integrating the latest clinical guidelines and research. Clinician ratings confirmed the effectiveness of PDT in the management of patients with CD.