AI for Evidence-based Treatment Recommendation in Oncology: A Blinded Evaluation of Large Language Models and Agentic Workflows

Guannan Zhai¹Merav Bar²Andrew Cowan^3,4Samuel Rubinstein⁵Qian Shi⁶Ningjie Zhang⁷En Xie⁸Will Ma^8*

• ¹The George Washington University Department of Statistics, Washington, United States
• ²Bristol Myers Squibb, New Jersey, United States
• ³Fred Hutchinson Cancer Center Clinical Research Division, Seattle, United States
• ⁴The University of Alabama at Birmingham Division of Hematology and Oncology, Birmingham, United States
• ⁵The University of North Carolina at Chapel Hill Division of Hematology, Chapel Hill, United States
• ⁶Mayo Clinic Department of Quantitative Health Sciences, Rochester, United States
• ⁷Shanghai Jiao Tong University Department of Bioinformatics and Biostatistics, Shanghai, China
• ⁸Hope AI, Inc, Princeton, NJ, United States

Background: Evidence-based medicine is crucial for clinical decision-making, yet studies suggest that a significant proportion of treatment decisions do not fully incorporate the latest evidence. Large Language Models (LLMs) show promise in bridging this gap, but their reliability for medical recommendations remains uncertain. Method: We conducted an evaluation study comparing five LLMs' recommendations across 50 clinical scenarios related to multiple myeloma diagnosis, staging, treatment, and management, using a unified evidence cutoff of June 2024. The evaluation included three general-purpose LLMs (OpenAI o1-preview, Claude 3.5 Sonnet, Gemini 1.5 Pro), one retrieval-augmented generation (RAG) system (Myelo), and one agentic workflow-based system (HopeAI). General-purpose LLMs generated responses based solely on their internal knowledge, while the RAG system enhanced these capabilities by incorporating external knowledge retrieval. The agentic workflow system extended the RAG approach by implementing multi-step reasoning and coordinating with multiple tools and external systems for complex task execution. Three independent hematologist-oncologists evaluated the LLM-generated responses using standardized scoring criteria developed specifically for this study. Performance assessment encompassed five dimensions: accuracy, relevance, comprehensiveness, hallucination rate, and clinical use readiness. Results: HopeAI demonstrated superior performance across accuracy (82.0%), relevance (85.3%), and comprehensiveness (74.0%), compared to OpenAI o1-preview (64.7%, 57.3%, 36.0%), Claude 3.5 Sonnet (50.0%, 51.3%, 29.3%), Gemini 1.5 Pro (48.0%, 46.0%, 30.0%), and Myelo (58.7%, 56%, 32.7%). Hallucination rates were consistently low across all systems: HopeAI (5.3%), OpenAI o1-preview (3.3%), Claude 3.5 Sonnet (10.0%), Gemini 1.5 Pro (8.0%), and Myelo (5.3%). Clinical use readiness scores were relatively low for all systems: HopeAI (25.3%), OpenAI o1-preview (6.0%), Claude 3.5 Sonnet (2.7%), Gemini 1.5 Pro (4.0%), and Myelo (4.0%). Conclusion: This study demonstrates that while current LLMs show promise in medical decision support, their recommendations require careful clinical supervision to ensure patient safety and optimal care. Further research is needed to improve their clinical use readiness before integration into oncology workflows. These findings provide valuable insights into the capabilities and limitations of LLMs in oncology, guiding future research and development efforts toward integrating AI into clinical workflows.