Domain-specific large language model for predicting prostate cancer treatment plan.

• DOI: 10.1200/jco.2025.43.5_suppl.428
• Publication Date: 2025-02-18T14:30:51Z
• Authors: Umar Ghaffar; Amara Tariq; Mouneeb Choudry; Logan Briggs; Aneeta Channar; Imon Banerjee; Man Luo; Irbaz Bin Riaz; Haidar Abdul-Muhsin

428 Background: Prostate cancer management presents a significant healthcare burden, with the need to efficiently triage patients for treatment. Our objective is leverage large language models predict physician-recommended treatment plans from unstructured clinical notes. By accurately predicting plans, we aim risk stratify and effectively, thereby optimizing allocation of physician resources. Methods: 448 initial urology consultation patient notes following first positive prostate biopsy were identified. The recommended final treatments received manually annotated establish ground truth labels (Table 1). dataset was split 80:20 training testing, preprocessed remove plan sections formatted into question-answer (QA) format. A domain-specific model (LLM) inspired by GPT specialized tokenizer (PCa- LLM) terminology developed. QA built using PCa-LLM compared those GPT-2 as backbone treatments. Results: For our LLM (PCa-LLM) showed superior performance higher AUROC scores curative vs. non-curative (0.78 0.65), chemo-hormonal other (0.89 surveillance all (0.72 0.70), while both achieved same high 0.99 treatments, demonstrated better (0.77 0.74) (0.71 0.66), GPT2 outperformed 0.70). Both an Conclusions: predicted most categories than GPT2, scores, can be utilized Task Physician -Recommended Treatment Plan Final Received Curative Prostatectomy/Radiation 228 230 Non-Curative Focal Therapy 22 15 Active Surveillance 40 45 Chemo-hormonal 30 Model Predictions (AUROC) 0.65 0.78 0.74 0.77 Chemohormonal 0.89 0.66 0.71 0.64 0.60 0.67 0.59 0.70 0.72