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

Paper Details

Abstracts, authors, affiliations, and disclosures.

Monday, December 14, 2026
02:45 p.m. - 04:15 p.m.
MCCNO - 393-396
This session highlights the transition of AI from diagnostic innovation to clinical decision support and personalized patient management. Presentations demonstrate how AI can predict treatment-related toxicities and complications, identify cardiovascular and bleeding risks, enhance MRD assessment, automate diagnostic triage, and improve disease risk stratification. Together, these studies illustrate the emerging role of AI in anticipating clinical outcomes, guiding therapeutic decisions, and enabling more precise, individualized care in hematologic malignancies.
Technology and Procedures, Assays, Imaging, Molecular Testing, Pathology, Bioinformatics, Emerging Technologies, Treatment Considerations, Biological Therapies, (CAR)-T Cell Therapies, Adverse Events, Artificial Intelligence (AI), Computer Vision, Deep Learning, Machine Learning, Natural Language Processing, Diseases, Bleeding and Clotting, Thromboembolism, Thrombotic Disorders, Lymphoid Malignancies, Lymphoid Leukemias, Lymphomas, Plasma Cell Disorders, Measurable Residual Disease, ALL, Non-Hodgkin Lymphoma, B Cell Lymphoma, Aggressive Lymphoma, Research, Translational Research, Clinical Research, Real-World Evidence
PAPER

960 Integration of a bone marrow pathology foundation model with the Μayo 20/2/20 model improves risk stratification in smoldering multiple myeloma

Jansen N Seheult, MBBCh, BAO, MSc, MS, MD, Steven Huang2, Reyhan Keser, PhD3, Prasad Bandodkar, Ph.D.4, Yurim Lee, PhD5, Zachary Larson6, Yanglan Ou, D.Sc., M.E.7, Dong Chen, MD, PhD8, Min Shi, M.D.9, Horatiu Olteanu, M.D., Ph.D.10, Linda Baughn, PhD11, Gunveen Sachdeva12, Richard Kandasamy, PhD13, Nicholas Jeffrey Braun, PhD14, Moritz Binder, MD, MPH15, Eric Hsi, M.D.16, Wilson Gonsalves, MD17, Dragan Jevremovic, MD PhD18, S. Vincent Rajkumar, MD19, Shaji Kumar, MD20, Wenchao Han, PhD21, Saurabh Zanwar, MBBS, MD22
1 Mayo Clinic, Rochester, Minnesota, United States
2 Mayo Clinic, Rochester, Minnesota, United States
3 Mayo Clinic, Rochester, Minnesota, United States
4 Mayo Clinic, Rochester, Minnesota, United States
5 Mayo Clinic, Rochester, Minnesota, United States
6 Mayo Clinic, Rochester, Minnesota, United States
7 Mayo Clinic, Rochester, Minnesota, United States
8 Mayo Clinic, Rochester, Minnesota, United States
9 Mayo Clinic, Rochester, Minnesota, United States
10 Mayo Clinic, Rochester, Minnesota, United States
11 Mayo Clinic, Rochester, Minnesota, United States
12 Mayo Clinic, Rochester, Minnesota, United States
13 Mayo Clinic, Rochester, Minnesota, United States
14 Mayo Clinic, Rochester, Minnesota, United States
15 Mayo Clinic, Rochester, Minnesota, United States
16 Mayo Clinic, Rochester, Minnesota, United States
17 Mayo Clinic, Rochester, Minnesota, United States
18 Mayo Clinic, Rochester, Minnesota, United States
19 Mayo Clinic, Rochester, Minnesota, United States
20 Mayo Clinic, Rochester, Minnesota, United States
21 Mayo Clinic, Rochester, Minnesota, United States
22 Mayo Clinic, Rochester, Minnesota, United States

Background: Smoldering multiple myeloma (SMM) is an asymptomatic plasma cell disorder with a highly heterogeneous risk of progression to multiple myeloma. The Mayo/IMWG 20/2/20 model is the current standard for risk stratification but does not fully capture this heterogeneity, as some patients classified as low or intermediate risk progress rapidly while others classified as high risk remain stable. With early treatment now available for high-risk SMM, improving risk stratification has become increasingly important. Although bone marrow biopsy is routinely performed at diagnosis, the 20/2/20 model incorporates only plasma cell percent, potentially overlooking other prognostic information in the biopsy. We evaluated whether BloodBytes, a bone marrow pathology foundation model, provides prognostic information beyond the Mayo/IMWG 20/2/20 model.

Methods: We included patients with SMM diagnosed at a single institution between September 2000 and February 2022, each with a digitized H&E-stained bone marrow core biopsy and a 20/2/20 score. The endpoint was progression within 2 years, the current clinical benchmark used for determining high-risk status. Representations were extracted from 20x tiles sampled (median 425 per slide) from H&E whole-slide images of the biopsy core using BloodBytes, a bone marrow-specific pathology foundation model trained on 11 million multi-resolution patches from 9,457 H&E-stained bone marrow core biopsies. Tile representations were pooled into a slide representation (6144-dimensional) followed by principal component reduction (20 components) and logistic regression. Attention-based multiple-instance learning (MIL, 512-dimensional) was evaluated in parallel with the same downstream classifier. Four-fold cross-validation used disjoint test folds grouped by patient and stratified on progression, 20/2/20 category, and diagnostic era, yielding one out-of-fold prediction per patient. Incremental value over 20/2/20 was assessed by out-of-sample incremental log-likelihood, permutation testing with within-stratum shuffling, and within-stratum discrimination.

Results: Of the 203 patients, with a median follow-up of 11.8 years (reverse Kaplan-Meier, 95% CI 9.4–14.0), 62 (30.5%) progressed within 2 years from diagnosis of SMM. The 2-year progression rates rose across 20/2/20 categories: 11/77 (14%) in low-, 15/61 (25%) in intermediate-, 36/65 (55%) in high-risk 20/2/20 categories. As single predictors, 20/2/20 achieved an AUC of 0.715 and BloodBytes achieved an AUC of 0.723; combining the two improved the AUC to 0.774. Adding the BloodBytes score improved prediction in held-out patients, with a positive incremental log-likelihood in all four folds (total +6.34) and a stable image coefficient across folds (+0.33 to +0.38; OR 1.42 per logit). Permutation testing with within-stratum shuffling confirmed a conditional association independent of 20/2/20 (p=0.001).

Discrimination was concentrated among patients the clinical model designated lower risk (BloodBytes x 20/2/20 interaction p=0.037). The BloodBytes score achieved an AUC of 0.792 (95% CI 0.63-0.93) in the low-risk stratum, and 0.746 (0.64-0.85) across the combined low- and intermediate-risk group, against 0.584 (0.44-0.74) among high-risk patients. A single risk threshold, fixed on training data and applied unchanged to held-out patients, flagged 26% of low- and intermediate-risk patients, whose observed 2-year progression rate was 44% versus 10% in those not flagged, against a 19% base rate (62% of progressors captured). BloodBytes matched 20/2/20 for ranking time to progression (C-index 0.650 versus 0.656), with the combined model higher than either alone (0.692).

Attention MIL on the same tile representations did not improve discrimination (AUC 0.702 alone, 0.767 with 20/2/20); in the high-risk stratum neither model discriminated well (0.562 for MIL vs 0.584).

Conclusions: The BloodBytes pathology foundation model provides independent and complementary prognostic information to the Mayo/IMWG 20/2/20 model by refining risk stratification among patients classified as low or intermediate risk, but it adds limited value in those already classified as high-risk. BloodBytes may be identifying additional biologic information beyond plasma cell percent including spatial organization and plasma cell morphology. These findings suggest that utilizing pathology foundation models and established clinical risk models together may improve risk stratification in SMM.

Jansen N Seheult, MBBCh, BAO, MSc, MS, MD: No relevant conflicts of interest to declare.