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

Paper Details

Abstracts, authors, affiliations, and disclosures.

SESSION

653. Multiple Myeloma: Clinical and Epidemiological: Poster III

Program: Poster Abstract Session
Monday, December 14, 2026
06:00 p.m. - 08:00 p.m.
MCCNO - Hall F
Access, Representation, Population Health, Technology and Procedures, Imaging, Treatment Considerations, Combination Therapy, Supportive Care, Therapy Sequence, Biological Therapies, Transplantation, Bispecific Antibody Therapy, (CAR)-T Cell Therapies, Immunotherapy, Monoclonal Antibody Therapy, Non-Biological Therapies, Pharmacology, Radiation Therapy, Women’s Health, Adverse Events, Artificial Intelligence (AI), Machine Learning, Biological Processes, Clinical Practice (Health Services and Quality), Clinical Guidelines, Patient Safety, Practice Models, Quality Improvement, Quality of Care, Value, Diseases, Bone Marrow Failure Syndromes, Hemoglobinopathies, Infectious Diseases, Viral, Lymphoid Malignancies, Plasma Cell Disorders, Myeloid Malignancies, Research, Translational Research, Clinical Research, Clinical Trials, Epidemiology, Patient-Reported Outcomes, Real-World Evidence, Registries, Survivorship, Study Population, Human, Adult, Elderly
PAPER

5732 Modular machine learning-based risk stratification in multiple myeloma: An ASH research collaborative analysis

Arwa Bohra, MD, Yelak Biru, MSc2, Susan Geyer, PhD3, Levi Pederson, MS4, Peyton Schumacher5, Travis Dockter, MS6, Emily Semmel, MS7, Shandra James, DNP8, Shambavi Richard, MD9, Manisha Bhutani, MD10, Jeffrey A. Zonder, MD11, Venkatesh Charya, MS12, Anita D'Souza, MD13, Jodi Lipof, MD14, Ajay K Nooka, MD, MPH15, Poy Theprungsirikul, MD16, Hamza Hashmi, M.D.17, Jiefu Wang, PhD18, Aaron S Rosenberg, MD19, Samuel Rubinstein, MD MSCI20, Kenneth Anderson, MD21, William Wood22, Saad Z. Usmani, MD23, Shaji Kumar, MD24
1 Mayo Clinic, Rochester, United States
2 ASH Research Collaborative, Washington, United States
3 Mayo Clinic, Rochester, United States
4 Mayo Clinic, Rochester, United States
5 Mayo Clinic, Rochester, United States
6 Mayo Clinic, Rochester, United States
7 ASH Research Collaborative, Washington, United States
8 ASH Research Collaborative, Washington, United States
9 Mount Sinai Cancer Center, New York, United States
10 Atrium Health Levine Cancer Institute, Wake Forest University School of Medicine, Forest City, United States
11 Barbara Ann Karmanos Cancer Institute / Wayne State University School of Medicine, Detroit, United States
12 University of North Carolina at Chapel Hill, Chapel Hill, United States
13 Medical College of Wisconsin, Wisconsin, United States
14 University of California, San Francisco, United States
15 Emory University School of Medicine, Atlanta, United States
16 Dana-Farber Cancer Institute, Boston, United States
17 Memorial Sloan Kettering Cancer Center, New York, United States
18 University of California, Davis, United States
19 University of California, Davis, United States
20 University of North Carolina at Chapel Hill, Chapel Hill, United States
21 Dana-Farber Cancer Institute, Boston, United States
22 University of North Carolina at Chapel Hill, Chapel Hill, United States
23 Memorial Sloan Kettering Cancer Center, New York, United States
24 Mayo Clinic, Rochester, United States

Background: Multiple myeloma (MM) risk stratification traditionally relies on staging systems built on linear hazard assumptions. Real-world clinical data are heterogeneous across sociodemographic, cytogenetic, and serologic domains, with effect modification arising from the coexistence of potential risk attributes not otherwise captured in existing risk systems (for example, Black or African American (AA) race or Hispanic ethnicity), warranting a need for more integrated modeling. We developed and cross-validated a modular machine learning framework that independently models each domain, selects the best-performing algorithm per domain by held-out discrimination, and combines domain-specific risk scores into a single multivariable model.  

Methods: To generate a robust risk stratification model, we developed a methodological approach to integrate and analyze data in a multi-phase, machine-learning framework that will be applied to newly diagnosed MM pts from 12 academic centers across the United States that participate in the ASH Research Collaborative’s Data Hub Program. To test and refine this complex framework, we initiated an interim model using data on pts from one site (Mayo Clinic). Overall survival (OS) was defined from plasma cell disorder diagnosis date to date of death or last follow-up. Three independent modules were built: (1) sociodemographic (age at diagnosis, self-reported race (missing=26, excluded from analysis), ethnicity (not documented=45, retained as covariate), birth sex; (2) bone marrow module, utilizing bone marrow plasma cell percentage and FISH abnormalities from biopsies obtained within 3 months of diagnosis or prior to first therapy, with the sample closest to diagnosis retained when duplicates existed; and (3) monoclonal protein isotype. Within each module, seven survival models were compared via stratified 5-fold cross-validation using C-index: standard Cox regression, Firth's penalized Cox regression, ridge, lasso, and elastic net-regularized Cox regression, Random Survival Forest, and gradient boosting. The best-performing model per module generated a patient-level risk score. The three risk scores were combined with serum beta-2-microglobulin (β2M) ≥5.5 mg/L in a final multivariable Cox regression model. An optimal risk dichotomization cutpoint was derived using the maximally selected rank statistic.  

Results: Overall, 2,187 pts were included (median follow-up 96.8 mo). Sociodemographic module (n=2,161; 562 events): standard Cox performed the best (C=0.63); significant predictors included age at diagnosis, HR 1.04/year (95% CI 1.03–1.05, p<0.01); AA race, HR 1.7 (1.–2.7, p=0.03); Hispanic ethnicity, HR 1.91 (1.18–3.10, p=0.008); and female birth sex, HR 0.8 (0.7–1, p=0.02). Bone marrow module (n=1,423; 345 events): Ridge regression was used for its clinically meaningful weights (C=0.61); top markers by HR were 1q gain/amp (1.3), del(17p) (1.3), del(1p) (1.3), and t(4;14) (1.28), with t(11;14) showing only a modest HR (1.05). Isotype module (n=2,175; 566 events) showed weaker discrimination (C=0.53, p<0.001). In the final combined model (n=1,116 with complete data across all domains; 274 events): sociodemographic risk score (3.07, 2.3–4.1, p<0.01), bone marrow risk score (2.7, 1.9–4, p<0.01), isotype risk score (1.5, 1.1–2.2, p=0.02), and β2M≥5.5 (1.6, 1.2–2.0, p<0.01) were each independently prognostic. (CI: 0.709). Cutpoint analyses showed the ability to generate risk-based classes. This risk score cutpoint separated 353 high-risk from 763 low-risk patients (HR 3.82); median OS was 119 mo in the high-risk group vs. not reached in the low-risk group. These approaches will be used to enable cross-site validation, and analyses will be extended to the full Data Hub cohort prior to presentation. Further, clinical validation of the proposed model is warranted before clinical adoption.  

Discussion: This modular framework extends the current staging paradigm by incorporating biology with sociodemographic disparities and age. It also allows for clinical representation of FISH and serological categories that are under-represented in purely linear modeling. While it utilizes readily available clinical data to significantly improve discriminatory power to 0.71 from 0.64-0.66 of previous clinical models, the architecture is extensible and allows for separate inclusion of modules associated with emerging biomarkers such as omics, minimal residual disease, and imaging features. 

Arwa Bohra, MD: No relevant conflicts of interest to declare.