2026 ASH®
Annual Meeting & Exposition December 12-15, 2026 ▪ New Orleans, Louisiana

Paper Details

Abstracts, authors, affiliations, and disclosures.

Saturday, December 12, 2026
05:30 p.m. - 07:30 p.m.
MCCNO - Hall F
Technology and Procedures, Molecular Testing, Clinical Procedures, Treatment Considerations, Supportive Care, Biological Therapies, Transplantation, Bispecific Antibody Therapy, Chimeric Antigen Receptor, (CAR)-T Cell Therapies, Adverse Events, Artificial Intelligence (AI), Natural Language Processing, Biological Processes, Microbiome, Clinical Practice (Health Services and Quality), Practice Models, Quality Improvement, Quality of Care, Diseases, Bone Marrow Failure Syndromes, Aplastic Anemia, Hemoglobinopathies, Sickle Cell Disease, Thalassemia, Immunology, GVHD, Infectious Diseases, Viral, Bacterial, Fungal, Lymphoid Malignancies, Lymphoid Leukemias, Myeloid Malignancies, Chronic Myeloid Malignancies, Measurable Residual Disease, ALL, Research, MPN, Clinical Research, Clinical Trials, Real-World Evidence, Registries, Study Population, Human, Adult
PAPER

2450 A large language model agent for allogeneic transplant conditioning regimen selection: A blinded validation study

Ron Ram I, Head, BMT Unit, Odelia Amit, MBBS, BSc2, Ronit Gold, Rn3, Muhamad Ganaim, MD4, Gil Fridberg5
1 Tel Aviv Sourasky Medical Center, Tel Aviv, Israel
2 Tel Aviv Sourasky Medical Center, Tel Aviv, Israel
3 Tel Aviv Sourasky Medical Center, Tel Aviv, Israel
4 Tel Aviv Sourasky Medical Center, Tel Aviv, Israel
5 Tel Aviv Sourasky Medical Center, Tel Aviv, Israel

Introduction: Selection of a conditioning regimen before allogeneic hematopoietic cell transplantation is a complex, multifactorial decision made by a multidisciplinary team (MDT), integrating patient, disease, and donor baseline and dynamic characteristics. Published guidelines alone cannot fully capture this complexity, as each center maintains its own evolving institutional protocol, and the reasoning behind individual MDT decisions is often too intricate and context-dependent to be fully codified in writing. We evaluated whether a large language model (LLM) agent could reproduce real MDT conditioning decisions, and separately quantified the contributions of exposure to historical MDT decisions ("learning") versus a written institutional standard operating procedure (SOP).

Methods: Using Claude Sonnet 4.6 (Anthropic) via independent, stateless API calls, we first exposed the model to real MDT conditioning decisions from 2023-2024 (n=50), constituting its learning corpus. We then compared five agent configurations, blinded, on documented allogeneic conditioning decisions from 2025 (n=32): (i) general knowledge alone; (ii) learning alone; (iii) institutional SOP alone; (iv) learning+SOP (equal weight); and (v) learning+SOP with an explicit conflict-resolution rule plus clinical-reasoning discipline principles. Each evaluation used a stateless call with no conversation history, so the model received only de-identified clinical input and never any information about case outcomes or prior model outputs, eliminating self-evaluation bias. The 2 best-performing configurations (ii and v) were rebuilt with an expanded learning corpus of 82 decisions (2023-2025) and prospectively tested on conditioning decisions from 2026 (n=16), unseen in any prior step. Success was determined based on concordance for conditioning intensity and specific regimen (exact vs. same-backbone match), scored against the actual MDT decision.

Results: On the 2025 retrospective set, learning alone (ii) outperformed general knowledge alone (i) on intensity concordance (87% vs 77%) and exact regimen concordance (30% vs 23%). Adding the institutional SOP without discipline principles (iii, iv) reduced intensity concordance (50-67%) despite improving exact regimen concordance (27-37%), apparently due to rigid application of numeric cutoffs over holistic judgment; combining SOP with explicit discipline principles (v) recovered intensity concordance to 77%. Systematic review of discrepancies between agent recommendations and actual MDT decisions across these configurations identified 13 concrete gaps in the written SOP (undocumented decision criteria and contradictions), which were incorporated into a revised institutional protocol. The two best-performing configurations were rebuilt with an expanded 82-case corpus, one incorporating the revised SOP, and prospectively evaluated on 16 conditioning decisions from 2026. The learning-only agent achieved 50% intensity concordance, 12.5% exact regimen concordance, and 19% exact+same-backbone concordance. The learning+SOP+discipline agent, using the revised SOP, achieved 86.7% intensity concordance (Cohen's kappa 0.79, a chance-corrected agreement measure; 95% confidence interval [CI] 0.48-1.00), 33.3% exact regimen concordance, and 86.7% exact+same-backbone concordance. As an exploratory signal, cross-configuration agreement on the 2025 set correlated with per-case accuracy (r=0.62), though in 2 of 30 cases all five configurations concurred on an incorrect answer, indicating agreement can also reflect shared bias rather than genuine certainty.

Conclusion: An LLM agent trained on institutional MDT decision history achieved clinically meaningful concordance with conditioning decisions in a genuinely blinded, prospective evaluation. The agent combined institutional precedent with a written protocol and explicit rules for resolving conflicts between them. Beyond conditioning selection, this framework offers a generalizable approach to auditing clinical decisions, refining institutional protocols, and preserving expert reasoning for future practice.

Ron Ram I, Head, BMT Unit: BMS - Consultancy (Includes expert testimony); BMS - Consultancy (Includes expert testimony); Gilead - Consultancy (Includes expert testimony); Sanofi - Consultancy (Includes expert testimony); GILEAD - Consultancy (Includes expert testimony); Novartis - Consultancy (Includes expert testimony); Sanofi - Consultancy (Includes expert testimony); Pfiser - Consultancy (Includes expert testimony)