Iris Publication

Risk stratification is essential for the delivery of optimal treatment in childhood acute lymphoblastic leukemia. However, current risk stratification algorithms dichotomise variables and apply risk factors independently which may wrongly assume identical associations across biologically heterogeneous subsets and reduce statistical power. Accordingly, we developed and validated a prognostic index (PIUKALL) which integrates multiple risk factors and uses continuous data. We created discovery (n=2,405) and validation (n=2,313) cohorts using data from four recent trials (UKALL2003, COALL-03, DCOG-ALL10, NOPHO-ALL2008). Using the discovery cohort, multivariate Cox regression modelling defined a minimal model that included white cell count at diagnosis, pre-treatment cytogenetics and end of induction minimal residual disease. Using this model we defined PIUKALL - a continuous variable that assigns personalised risk scores. The PIUKALL correlated with risk of relapse and validated in an independent cohort. Using PIUKALL to risk stratify patients improved the C-index for all endpoints compared to the traditional algorithms. We used PIUKALL to define four clinically relevant risk groups which had differential relapse rates at 5 years and were similar between the two cohorts: discovery - low 3% (95% CI 2-4), standard 8%(6-10), intermediate 17%(14-21), high 48%(36-60) and validation low 4%(3-6), standard 9%(6-12), intermediate 17%(14-21), high 35%(24-48). An analysis of the area under the curve confirmed the PIUKALL groups were significantly better at predicting outcome than the algorithms employed in each trial. The PIUKALL developed in this study provides an accurate method for predicting outcome and a more flexible method for defining risk groups in future studies.