A Highly Effective Model of Risk Assessment in Diffuse Large B Cell Lymphoma (DLBCL) Using Classification and Regression Tree (CART) Analysis.

Numerous biological factors affect the outcome of diffuse large B cell lymphoma (DLBCL), and it is important that this information is used effectively to enhance the quality of patient care. Classification and Regression Tree Analysis (CART) is a novel statistical approach that has the advantage of being able to handle complex interactions between variables, making it ideal for the generation of clinical decision rules. The aim of this study was to use CART to define and validate a tool for assessing the prognosis of patients with DLBCL based on routinely available clinical and biological parameters. 361 presenting DLBCL patients, where there was an intention to treat with CHOP chemotherapy, were classified according to BCL2 and FOXP1 expression, germinal centre phenotype, P53 status, and the presence of BCL6 rearrangement and the t(14;18). All of these factors that were used in the analysis have previously been shown to have prognostic significance. Median age was 60 (range 17–94) and median overall survival (OS) was 6.8 years with a median follow up of 3.5 years (range 0–20). The terminal nodes generated by CART analysis were classified into poor or good prognostic groups and survival estimates were calculated using Cox proportional hazards regression. Models predicting OS at 1 and 2 years and at last follow-up were generated using biological data alone and in combination with age or IPI. Each of these CART models was highly effective at discriminating poor and good risk. Using biological data alone in the 1-year model, 55% of patients were good risk, with an 81% chance of survival compared to a 66% chance in poor risk patients (Hazard ratio (HR) 1.86; 95% Confidence Interval (CI) 1.27–2.82, log rank test p=0.003). Addition of patients’age into the analysis significantly improved risk stratification, with 1-year OS of 82% (n=307) and 33% (n=54) respectively (HR 5.71; CI 3.74–8.73, p<0.0001), however the IPI did not enhance the model further (HR 4.00; CI 2.44–6.57, p<0.0001). In the 2-year model, 77% of patients were good risk, with a 76% chance of survival and 23% were poor risk (HR 4.37, CI 3.07–6.23, p<0.0001). In order to test the validity of the analysis and to determine whether addition of rituximab had an impact on the prognostic model, we applied the 1-year model, using age and biological factors, to a series of CHOP-R treated patients (n=94) with a median follow-up of 0.6 years (range 0–1.6). 53% of patients were identified with good risk (90% 1-year OS) and 47% of patients had a 59% chance of survival at 1-year (HR 4.77; CI 1.75–12.94, p=0.0007). Using CART analysis we have defined a highly effective prognostic model that has been used successfully to assign risk in DLBCL patients treated with CHOP chemotherapy, and that is particularly effective at identifying the majority of patients who have an excellent prognosis. Although follow up is short for the CHOP-R series, the data suggests that the same predictive model applies equally in this group as to patients receiving CHOP, however the probability of survival in both good and poor prognostic may be improved by the addition of rituximab. Using CART we have used routine clinical and biological data to generate a robust model that can be used to predict outcome in patients with DLBCL and this will improve clinical decision making in this heterogeneous group.