Development of a Conditional Event-Free Survival Tool in Diffuse Large B-Cell Lymphoma

Abstract

Background: The International Prognostic Index (IPI) and related prognostic models for diffuse large B-cell lymphoma (DLBCL) use baseline clinical factors to predict outcomes at the time of initial therapy. These models are static and do not update over the disease course. Although most patients respond to initial therapy, patients and their treating physicians need better tools to understand recurrence risk, which decreases over time. For example, a young woman who completed treatment two years ago and is now contemplating pregnancy may wish to know about her recurrence risk in the next year. Similarly, physicians and patients alike may want to design a surveillance strategy that is commensurate with recurrence risk over time. Conditional event-free survival (cEFS) accounts for both patient risk factors plus the declining probability of event recurrence the longer a patient has been in remission. We report the development and validation of a dynamic risk prediction tool for DLBCL patients following completion of front-line immunochemotherapy (IC). Methods: Data from the LEO/MER cohort study and several prospective NCTN clinical trials (Alliance/CALGB 50303 [NCT00118209], S1001 [NCT01359592], E1412 [NCT01856192], and E4494 [NCT00003150]) were used to construct a dynamic prognostic model. Patients who had completed initial therapy and achieved a complete response (CR) (when response data were available; i.e., trial data) or did not experience an event within 12 weeks of end of therapy (when response data were not available; i.e., LEO/MER cohorts) were selected. Baseline variables included age, sex, stage, Eastern Cooperative Oncology Group (ECOG) performance status, pathology, and labs. Cox proportional-hazards regression modeling with backwards stepwise selection was utilized for model development. Model performance was assessed at fixed timepoints, using time-dependent areas under the receiver operating characteristic curve (tAUC of ROC) and Brier scores, and over the entire span of time in the data, using C-statistics. Model performance was compared to the performance of the IPI. Internal validation was performed using bootstrapping. External validation was separately performed using the Danish National Lymphoma Registry (LYFO). To develop the dynamic prediction tool, a landmark method was used, in which time since completion of therapy was also included as a key model input. An R Shiny application was developed to implement this dynamic prediction and facilitate point-of-care utilization, with the output being probability of remaining event-free during a user-specified time frame. Results: 2897 patients were included in model development, with the final model consisting of age (transformed to account for the non-linear association with outcome), sex, ECOG performance status, stage, and lactate dehydrogenase (LDH). Model parameters are shown in the Table. The final model outperformed a model containing only IPI in terms of C-statistic, tAUC, and Brier scores. The C-statistic was 0.66 (95% CI: 0.64, 0.68). Brier scores varied from 0.07 (0.06, 0.08) at 12 months to 0.19 (0.17, 0.20) at 84 months. The tAUCs varied from 0.65 (0.62, 0.69) at 12 months to 0.71 (0.68, 0.74) at 84 months. In contrast, in the IPI-only model, the C-statistic was 0.63, while the Brier scores ranged from 0.07 to 0.21 and tAUCs ranged from 0.64 to 0.65 between the same time intervals. After applying the model to 3250 patients from LYFO, the Brier scores and tAUCs were similar, ranging from 0.11 to 0.22 and 0.69 to 0.71, respectively, also between the same time intervals. In the Figure, we show that the 2-year cEFS of a 38-year-old female with stage III disease, a performance score of 1, and a LDH ratio of 4 having survived one year post-treatment is 0.908, compared to 0.921 for a female with the same model parameters but having survived 2 years. Conclusions: We present a dynamic risk prediction tool for patients with DLBCL having achieved a CR following front-line IC that allows for personalized prognostication of EFS both at the time of therapy completion and during follow-up. We are actively exploring prospective clinical trials of rational post-treatment surveillance based on individual cEFS, as a step towards evidence-based follow-up strategies. Support: U10CA180821 (CALGB 50303), https://acknowledgments.alliancefound.org, U10CA180820, UG1CA233339, and UG1CA232760 (ECOG-ACRIN), and U10CA180888 (SWOG). Figure 1View largeDownload PPTFigure 1View largeDownload PPT Close modal