Prognostic survival models for diffuse large B-cell lymphoma using statistical and machine learning approaches

Although diffuse large B-cell lymphoma (DLBCL) is a heterogeneous disease, existing prognostic tools, including the revised International Prognostic Index (R-IPI) and the National Comprehensive Cancer Network (NCCN-IPI), do not fully capture the heterogeneity and their performance in the competing-risks framework has not been validated. We developed and validated prognostic models for overall survival (OS) and progression-free survival (PFS) using both machine-learning (ML) and standard regression approaches in 2769 patients from the Lymphoma and Related Diseases Registry. Predictors included age, stage, performance status, chemo-immunotherapy, creatinine, lactate dehydrogenase, and anaemia, with extranodal involvement additionally used for OS and BCL6 expression for PFS. Both ML and regression-based models showed similar performance, with acceptable discrimination, particularly at 1- and 2-years. In validation, the Cox model achieved a 1-year OS AUC of 0.770, outperforming R-IPI (0.722) and comparable to NCCN-IPI (0.746). Both the nomogram and random survival forest models demonstrated greater 5-year OS risk separation, ranging from 26% in high-risk to 96% in low-risk patients and from 25 to 93%, respectively, compared with the R-IPI (50-91%) and NCCN-IPI (34-96%). Competing-risks analyses demonstrated that conventional methods underestimated survival, particularly in high-risk groups. While our models provided promising risk stratification, external validation is warranted.