Background: Accurate risk stratification is necessary for epidemiologic and outcomes research in diffuse large B-cell lymphoma (DLBCL). We evaluated performance characteristics of the clinically derived International Prognostic Index (IPI) and revised IPI (R-IPI) with a regression model-based score using the National Cancer Data Base.
Methods: We studied DLBCL patients diagnosed in 2004-2011, divided into derivation and validation cohorts. The model-based score was calculated from a Cox model incorporating variables routinely recorded by cancer registries. Calibration and discrimination of the indices with regard to overall survival were evaluated in the validation cohort.
Results: The IPI was recorded in 19,511 of 119,942 patients, of whom 79 % received chemotherapy. Both clinical indices provided good survival discrimination (5-year estimate range 33-74 % for the IPI, and 41-87 % for the R-IPI), but explained only 16 % of variation in survival. Survival predictions among chemotherapy-treated patients were similar to estimates from published clinical cohorts. The model-based score had significantly better discrimination characteristics (5-year survival estimate range 22-87 %) and explained 23 % of variation in survival.
Conclusions: We validated the IPI and R-IPI as recorded by cancer registries to provide robust risk stratification in the general population with DLBCL, but a prognostic model using raw registry data provides superior performance. Explicit recording of prognostic factors is preferable to abstracting coarsened clinical indices for the purpose of population-based epidemiologic research. Considering low variation of survival explained by the standard clinical variables, incorporating molecular markers into registry data is necessary to improve risk stratification.