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

Paper Details

Abstracts, authors, affiliations, and disclosures.

Monday, December 14, 2026
06:00 p.m. - 08:00 p.m.
MCCNO - Hall F
Technology and Procedures, Assays, Imaging, Molecular Testing, Newborn Screening, Omics Technologies, Pathology, Serologic Tests, Bioinformatics, Computational Biology, Emerging Technologies, Profiling, Treatment Considerations, Combination Therapy, Therapy Sequence, Biological Therapies, Antibody Therapy, Transplantation, Bispecific Antibody Therapy, (CAR)-T Cell Therapies, Monoclonal Antibody Therapy, Non-Biological Therapies, Chemotherapy, Surgical, Adverse Events, Artificial Intelligence (AI), Computer Vision, Deep Learning, Generative AI, Machine Learning, Natural Language Processing, Biological Processes, Pathogenesis, CHIP, Genomics, Hematopoiesis, Immune Mechanism, Metabolism, Molecular Biology, Multi-Systemic Interactions, Clinical Practice (Health Services and Quality), Clinical Guidelines, Diseases, Hemoglobinopathies, Sickle Cell Disease, Immunology, GVHD, Lymphoid Malignancies, Lymphoid Leukemias, Lymphomas, Plasma Cell Disorders, Myeloid Malignancies, Acute Myeloid Malignancies, Chronic Myeloid Malignancies, Measurable Residual Disease, ALL, CLL, Non-Hodgkin Lymphoma, B Cell Lymphoma, Indolent Lymphoma, Aggressive Lymphoma, Research, Fundamental Science, AML, APL, Translational Research, MDS, MPN, CMML, Clinical Research, Clinical Trials, Health Outcomes Research, Patient-Reported Outcomes, Real-World Evidence, Registries, Survivorship, Study Population, Human, Adult, Elderly, Pediatric
PAPER

6118 Local large language models enable validated day-resolved reconstruction of graft-versus-host disease trajectories from clinical free text

Anna Katharina Foerster, MD, Miriam Mozaffari Jovein, MD2, Irmgard Matt3, Hauke Wilcken, MD4, Kristina Maas-Bauer5, Tobias Wertheimer6, Justus Duyster7, Jesus Duque Afonso, MD8, Philipp Arnold9, Ralph Wäsch, MD10, Jürgen Finke, MD11, Robert Zeiser, MD, PhD12, Claudia Wehr13, Thomas Meyer, MD14
1 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
2 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
3 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
4 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
5 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
6 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
7 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
8 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
9 Department of Radiology, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
10 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
11 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
12 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
13 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany
14 Department of Internal Medicine I, Hematology, Oncology and Stem Cell Transplantation, Medical Center - University of Freiburg, Faculty of Medicine, Freiburg, Germany

Background

Graft-versus-host disease (GvHD) is a major cause of non-relapse mortality (NRM) after allogeneic hematopoietic cell transplantation. Registries capture GvHD as interval-based summaries, including occurrence, maximum grade, and organ involvement, rather than day-resolved trajectories. The limited temporal resolution may obscure information on disease duration, recency, and recurrence. The manual extraction of these static summary fields is resource-intensive. Reconstructing longitudinal GvHD trajectories requires greater effort and is hard to scale. Large language models (LLMs) could automate this process, but sensitive data favor local, privacy-preserving deployment, and outputs require validation against human annotation.

Methods

We developed a privacy-preserving pipeline in which three local LLMs from the Qwen and Gemma families extracted GvHD events from German clinical notes using up to 64GB of VRAM. The dataset comprised 16,998 clinical PDF documents totaling 4.2GB, with a mean of 30 documents per patient and a 02/2026 cut-off. Per-patient, per-date majority consensus defined events, followed by LLM-based grading to generate dated a/cGvHD trajectories with organ involvement.
Static endpoint validation: In 574 transplants performed between 01/2020 and 01/2026, LLM trajectories were collapsed to transplant-level endpoints and compared with the routine database, maintained by a medical documentation specialist, and with a human review of that database.
Temporal trajectory validation: Two physicians independently performed longitudinal annotation in GvHD-enriched cohorts of 31 (physician A) and 39 (physician B) transplants; 31 overlapped and were used for inter-rater and event-level consensus analyses.
Temporal outcome analysis: An exploratory landmark analysis evaluated a temporal GvHD score comprising recency-weighted burden, active-day proportion, and flares in Cox models for NRM and overall survival (OS), testing its added value beyond static GvHD flags.

Results

Static endpoint validation: Across 574 transplants, agreement for clinically relevant GvHD (aGvHD ≥II or cGvHD ≥moderate) was κ=0.889 for LLM versus review, κ=0.810 for routine database versus review, and κ=0.770 for LLM versus routine database. Weighted agreement for maximum aGvHD/cGvHD grade was κw=0.842/0.904 for LLM versus review, κw=0.804/0.626 for routine database versus review, and κw=0.707/0.605 for LLM versus routine database.
Temporal trajectory validation: Physicians A and B longitudinally graded 1,239 and 1,312 dated events, respectively. Manual annotation of the full cohort was estimated at 231 physician-hours. In the physician-annotated cohorts, trajectories were collapsed to a binary endpoint of clinically relevant a/cGvHD. Against physician A, both LLM and routine database achieved κ=0.919; against physician B, both achieved κ=0.658. Inter-physician agreement was κ=0.644. To assess the temporal detail uniquely provided by the pipeline, dated LLM events type-matched to physician consensus were compared on exact dates in the 31 transplants. Conditional event-level F1 scores were 0.995 for clinically relevant aGvHD and 0.973 for clinically relevant cGvHD. Accuracy was 0.985 for acute overall grade and ranged from 0.847 to 1.000 for organ-level classifications, lowest for pulmonary involvement.
Temporal outcome analysis: In exploratory landmark models, the temporal GvHD score provided information beyond static database flags. For NRM, the static GvHD flag was associated with outcome alone (HR 2.68, p=0.013) but not after inclusion of the temporal score (HR 1.47, p=0.44). The temporal score remained associated with NRM (p=0.037) and improved model fit and discrimination. Similar findings were observed for OS.

Conclusions

A local, privacy-preserving LLM pipeline reconstructed day-resolved a/cGvHD from clinical free text across a large, validated transplant cohort, with agreement that matched or exceeded routine structured annotation and fell within human-to-human variability. By converting documentation into time-resolved trajectories, the pipeline supports analyses unavailable from static registry fields, including time-dependent prognostic modeling, treatment-response evaluation, and characterization of disease onset, persistence, flare, and resolution. Beyond GvHD, this study demonstrates a scalable framework for transforming sensitive clinical narratives into structured longitudinal time-series data using local, privacy-preserving LLMs.

Anna Katharina Foerster, MD: No relevant conflicts of interest to declare.