Prognostic survival models for diffuse large B-cell lymphoma using statistical and machine learning approaches
Adugnaw Zeleke Alem1,2, Itismita Mohanty3, Nalini Pati3,4,5
1Department of Epidemiology and Biostatistics, College of Medicine and Health Sciences, Institute of Public Health, University of Gondar, Gondar, Ethiopia. Adugnaw.Alem@canberra.edu.au.
NPJ Precision Oncology
|July 16, 2026
Summary
New prognostic models for diffuse large B-cell lymphoma (DLBCL) show improved risk stratification for overall survival (OS) and progression-free survival (PFS) compared to existing tools. These models offer better patient outcome prediction in DLBCL management.
Area of Science:
- Hematology
- Oncology
- Biostatistics
Background:
- Diffuse large B-cell lymphoma (DLBCL) is a heterogeneous hematologic malignancy.
- Current prognostic indices like R-IPI and NCCN-IPI have limitations in capturing DLBCL heterogeneity and lack validation in competing-risks frameworks.
- Accurate prognostic tools are crucial for optimizing treatment strategies and patient management in DLBCL.
Purpose of the Study:
- To develop and validate novel prognostic models for overall survival (OS) and progression-free survival (PFS) in DLBCL patients.
- To compare the performance of machine-learning (ML) and standard regression models against existing prognostic indices (R-IPI, NCCN-IPI).
- To evaluate prognostic model performance within a competing-risks framework.
Main Methods:
- Development and validation of prognostic models using machine-learning and standard regression approaches in a cohort of 2769 DLBCL patients.
- Inclusion of clinical variables (age, stage, performance status, chemo-immunotherapy, creatinine, lactate dehydrogenase, anaemia) and specific biomarkers (extranodal involvement, BCL6 expression).
- Performance assessment using metrics like Area Under the Curve (AUC) and risk stratification analysis, including competing-risks modeling.
Main Results:
- Both ML and regression models demonstrated acceptable discrimination for OS and PFS, outperforming or matching existing indices.
- The Cox model achieved a 1-year OS AUC of 0.770, superior to R-IPI (0.722) and comparable to NCCN-IPI (0.746).
- Nomogram and random survival forest models showed superior 5-year OS risk separation, highlighting improved prognostic capability, especially in high-risk groups. Competing-risks analysis revealed underestimation of survival by conventional methods.
Conclusions:
- The developed prognostic models offer enhanced risk stratification for DLBCL patients compared to R-IPI and NCCN-IPI.
- Machine-learning and regression approaches provide valuable tools for predicting OS and PFS in DLBCL.
- External validation of these novel models is recommended to confirm their generalizability and clinical utility.
