P243 Survival prediction in malignant pleural mesothelioma: fundamental limitations of routinely available clinic predictors

Introduction Accurate prognostication is difficult in Malignant Pleural Mesothelioma (MPM). Published prognostic models antedate optimal staging, a range of emerging predictors and use methods that cannot be up-scaled to incorporate these. Most existing models allocate patients to risk groups rather than precisely predicting survival. We developed robust computational models that can be up-scaled and provide quantitative statistics regarding the predictions offered. Here we report their performance using routinely available clinical data, on which previous models are based. Materials and Methods Baseline information regarding 20 candidate predictors was collected for 269 MPM patients diagnosed in the West of Scotland (January 2008 – April 2014). Patients were allocated to balanced training (n=169) and validation sets (n=100). Prognostic signatures (minimal length best-performing multivariate trained models) were generated by Least Absolute Shrinkage and Selection Operator (Lasso) regression for Overall Survival (OS), OS Results Median OS was 270 (IQR 140–450) days. The primary OS model assigned high weights to 4 predictors: age, performance status, white cell count and serum albumin, and after cross-validation performed significantly better than would be expected by chance (mean DXY 0.332 (+/-0.019) figure 1). However, validation set DXY was only 0.221 (0.0935–0.346), equating to a 22% improvement in survival prediction than would be expected by chance. 6- and 12 month OS signatures included the same 4 predictors, in addition to epithelioid histology plus platelets and epithelioid histology plus C-reactive protein (mean AUC 0.758 (+/-0.022) and 0.737 (+/-0.012), respectively). The Discussion The prognostic value of the basic clinical information contained in these, and previously published models, is fundamentally of limited value in accurately predicting MPM prognosis. The methods described are suitable for expansion using emerging predictors, including tumour genomics and volumetric staging.