Yan Leyfman, Medical Oncologist, Co-Founder and Executive Director of MedNews Week, shared a post on LinkedIn:
“Thank you so much Dr. Arturo LoAIza-Bonilla for your exceptional mentorship. Grateful as well for the opportunity to work with Dr. Connor Yost and your incredible team. Your guidance made this project possible.
Major Gap in Lymphoma Care:
NCCN guidelines are foundational in lymphoma management, yet real-world adherence is far from consistent—65–80% overall and as low as 11–46% in several lymphoma subtypes. Even in Hodgkin lymphoma, where 88% of clinicians report relying on NCCN, PET-adapted approaches remain inconsistently applied. This variability underscores a critical unmet need: scalable tools that can support reliable, guideline-concordant treatment decisions.
Our Approach:
We developed a RAG-enhanced GPT-4 agent by indexing the 2025 NCCN Lymphoma Guidelines and tested it against 50 diverse, real-world clinical vignettes.
Key Results:
- Significantly higher guideline concordance: RAG-GPT achieved a mean mG-PS of 0.32 vs 0.10 with baseline GPT-4 (Δ=0.22; p=0.038).
- Zero severe hallucinations, increasing trustworthiness and clinical safety.
- More ‘gold-standard’ treatment plans: 46% with RAG-GPT vs 22% with baseline GPT-4.
- Improved clarity and readability, though not statistically significant.
The Bigger Picture:
Lymphoma is uniquely challenging for AI-driven decision support due to its heterogeneity, bimodal age distribution, and lack of standardized regimens—especially in relapsed/refractory, transformed, or rare subtypes. In many cases, clinicians must go beyond NCCN and rely on judgment informed by limited case series. These complexities constrain LLM performance more than in algorithmically clearer cancers like breast or CNS tumors.
Yet despite these challenges, linking GPT-4 directly to NCCN guidelines via RAG meaningfully improved both accuracy and safety, emphasizing the value of guideline-anchored architectures in complex oncology domains.
Takeaway:
RAG-enhanced LLMs are emerging as powerful tools to narrow the guideline-adherence gap in lymphoma. Continued refinement and prospective validation will be essential to translate these advances into trustworthy, real-world clinical support.”
Title: Retrieval-Augmented GPT-4 Improves NCCN-Concordant Lymphoma Treatment Recommendations
Authors: Connor Yost, Sarah Monick, Arturo Loaiza-Bonilla, Nikita Tripathi, Peter Palumbo, Yan Leyfman, Yash Kumar, Natalie Ertz-Archambault, Robert Galamaga
You can read the full article in ASH Blood.
More posts featuring Yan Leyfman.