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

Paper Details

Abstracts, authors, affiliations, and disclosures.

Sunday, December 13, 2026
06:00 p.m. - 08:00 p.m.
MCCNO - Hall F
Access, Representation, Population Health, Technology and Procedures, Assays, Molecular Testing, Treatment Considerations, Combination Therapy, Drug Development, Supportive Care, Therapy Sequence, Biological Therapies, Antibody Therapy, Bispecific Antibody Therapy, (CAR)-T Cell Therapies, Non-Biological Therapies, Pharmacology, Adverse Events, Artificial Intelligence (AI), Machine Learning, Natural Language Processing, Biological Processes, Pathogenesis, Genomics, Multi-Systemic Interactions, Clinical Practice (Health Services and Quality), Clinical Guidelines, Quality of Care, Diseases, Lymphoid Malignancies, Lymphoid Leukemias, Lymphomas, Measurable Residual Disease, CLL, Non-Hodgkin Lymphoma, B Cell Lymphoma, Indolent Lymphoma, Aggressive Lymphoma, Research, Translational Research, Clinical Research, Clinical Trials, Epidemiology, Health Outcomes Research, Health Disparities Research, Real-World Evidence, Registries, Study Population, Human, Adult, Young Adult
PAPER

3879 CLL-MPI: Development of a molecular prognostic index in chronic lymphocytic leukemia assisted by machine learning

Daniel Sungchul Choi, MD, Christian Brieghel, MD, PhD2, Yue Ren, MS3, Roberta Azevedo, MD4, Stacey M. Fernandes, BS5, Jennifer R. Brown, MD, PhD6, Matthew S Davids, MD, MMSc7, Eliezer Van Allen, MD8, Carsten Utoft Niemann, MD, PhD9, Svitlana Tyekucheva, PhD10, Inhye E Ahn, MD11
1 Dana-Farber Cancer Institute, Boston, Massachusetts, United States
2 Copenhagen University Hospital, Rigshospitalet, Copenhagen, Denmark
3 Dana-Farber Cancer Institute, Boston, Massachusetts, United States
4 The University of Texas MD Anderson Cancer Center, Houston, Texas, United States
5 Dana-Farber Cancer Institute, Boston, Massachusetts, United States
6 Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States
7 Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States
8 Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States
9 Copenhagen University Hospital, Rigshospitalet, Copenhagen, Denmark
10 Dana-Farber Cancer Institute, Boston, Massachusetts, United States
11 Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States

Introduction: Despite comprehensive annotation of key genetic drivers of CLL, existing prognostic models account for a limited list of genetic variables. We report a molecular prognostic index for CLL (CLL-MPI) that incorporates genetic variables identified from a hybrid analytic approach combining traditional Cox regression and machine learning (ML) methods.

Methods: The discovery cohort included 1,058 patients (pts) with CLL in the Dana-Farber CLL database, excluding pts with monoclonal B cell lymphocytosis. Validation analyses will be conducted in an independent cohort of CLL pts in the Danish Lymphoid Cancer Research (DALY-CARE) database. In the discovery cohort, we performed feature selection with 10-fold cross-validation using L1-penalized (lasso) Cox proportional hazards (CoxPH) analysis and ML modeling with Shapley value-guided recursive feature elimination to identify variables relevant to overall survival (OS) and time to first CLL-directed therapy (TTFT). Five ML algorithms (random survival forest, gradient boosting survival analysis, XGBoost, DeepSurv, CoxTime) were utilized for feature selection. We used the Kaplan-Meier method and the log-rank test to compare outcomes across risk groups, and Uno’s c-statistic to assess each model’s performance.

Results: 760 pts had all pre-treatment biomarkers of interest available for the analyses, including FISH, IGHV, 88-gene panel NGS, beta-2 microglobulin (B2M), and lactate dehydrogenase (LDH). Median follow-up was 7.2 years. Median age at CLL diagnosis was 62 years. 61% were male. 44% had unmutated IGHV. Most pts (80% or 343/427 pts who received ≥1 therapy) received targeted agents as first-line therapy (38% with a BTK inhibitor [BTKi], 24% with a BCL2 inhibitor [BCL2i], 19% with the BTKi/BCL2i combinations, all subgroups with or without anti-CD20 antibody). Mutations of CLL driver genes (15% NOTCH1, 14% SF3B1, 13% TP53, 7% ATM, 7% RAS/RAF, 4% MYD88, 3% XPO1) and those associated with clonal hematopoiesis (9% DNMT3A, 3% TET2) were detected.

Lasso-Cox regression selected variables relevant to OS (age, B2M, IGHV, TP53 aberration (ab), SF3B1MUT, NOTCH1MUT) and TTFT (all variables relevant to OS plus sex, LDH, ATM ab, RAS/RAFMUT, XPO1MUT). Parallel ML analyses identified highly concordant lists of variables. Genetic variables had different significance for OS and TTFT, with more driver gene mutations being associated with TTFT and each additional mutation exerting an additive effect. Unlike other mutations, SF3B1MUT had a mixed effect on outcome by showing favorable OS and inferior TTFT. TP53 ab had a hazard ratio (HR) comparable to that of IGHV for OS prediction (HR: 1.62 for unmutated IGHV [U-IGHV], 1.80 for TP53 ab) and lower than that of IGHV for TTFT prediction (HR: 2.70 for U-IGHV, 1.33 for TP53 ab).

We built multivariable CoxPH models based on selected variables. Using the coefficient of each variable as a score, we assigned pts to four risk groups, which demonstrated distinct 10-year OS (98%, 98%, 85%, and 53% for low, intermediate, high, and very high-risk groups, respectively; p<0.001) and 5-year TTFT (as rates of cumulative incidence of initial CLL therapy: 13%, 25%, 56%, 72%; p<0.001). When restricting the analyses to pts who received targeted agents as first-line therapy, the OS model continued to distinguish outcomes in the very high-risk group (5-year OS: 71% in very high-risk compared to 93-97% in others; p<0.001), indicating its relevance to the targeted therapy setting (c-index: 0.79).

We next compared CLL-MPI and CLL-IPI in the discovery cohort. CLL-MPI improved prognostic accuracy over CLL-IPI for both OS (c-index: 0.80 vs. 0.72) and TTFT (c-index: 0.71 vs. 0.68). Over half of the pts (OS: 317/532 pts, TTFT: 306/532 pts) were reclassified by CLL-MPI (36% upstaged and 24% downstaged for OS, 48% upstaged and 9% downstaged for TTFT). Pts reclassified to the very high-risk group in CLL-MPI from lower-risk groups in CLL-IPI (n=81) had poor outcomes (10-year OS: 50%, 95% CI: 32-76%). Data from validation analyses in an independent pt cohort will be presented at the meeting.

Conclusions: CLL-MPI is a new prognostic tool that predicts OS and TTFT in pts predominantly treated with targeted therapy. Incorporation of CLL driver gene mutations improved the predictive performance of the models relative to CLL-IPI.  Further evaluation of CLL-MPI in larger datasets, including in prospective clinical trials, is warranted.
Daniel Sungchul Choi, MD: No relevant conflicts of interest to declare.