THE JOURNAL OF UROLOGYâ
Vol. 193, No. 4S, Supplement, Monday, May 18, 2015
e727
Source of Funding: none
MP58-16 Source of Funding: None
MP58-15 PREDICTING SURVIVAL AFTER RADICAL CYSTECTOMY: VALIDATION OF THE SPARC SCORE Brian Hu*, Los Angeles, CA; Manuel Eisenberg, Stephen Boorjian, Igor Frank, Rochester, MN; Leo Dalag, Kamran Movassaghi, Los Angeles, CA; Prabin Thapa, Rochester, MN; Gus Miranda, Siamak Daneshmand, Los Angeles, CA INTRODUCTION AND OBJECTIVES: The Survival Prediction After Radical Cystectomy (SPARC) score incorporates clinical and pathologic features to predict cancer specific survival (CSS) for urothelial carcinoma of the bladder after radical cystectomy (RC). Validation of this model would improve its generalizability. METHODS: Using the IRB-approved bladder cancer database at the University of Southern California, we identified patients who underwent RC for urothelial carcinoma of the bladder for curative intent from 1971-2009. Clinical factors (Charlson comorbidity index, ECOG performance status, hydronephrosis, adjuvant chemotherapy, smoking status) and pathologic factors (pathologic T stage, nodal status, multifocality, and lymphovascular invasion) included in the SPARC score were obtained. Patients were excluded if there were missing variables or if they underwent neoadjuvant chemotherapy. Associations between clinicopathologic factors and CSS were evaluated using Cox proportional hazards. Calibration plots were generated comparing actuarial CSS with SPARC predicted CSS by deciles. A c-index was generated to determine accuracy of the prediction. Kaplan Meier curves estimated CSS stratified by SPARC score and were compared with the log rank test. RESULTS: A total of 2045 patients underwent RC and 1123 (55%) met inclusion criteria with a median follow-up of 4.7 years (IQR 2.0-8.9 years). Of the 1123 patients, 332 (30%) died of bladder cancer. All the clinical and pathologic variables used in the SPARC scoring model were associated with CSS except for smoking status and tumor multifocality. Calibration plots demonstrated concordance between the SPARC-predicted and actuarial CSS with a c-index of 0.75. Kaplan Meier curves demonstrated significant differences in CSS based upon SPARC score, p<0.001 (Figure). CONCLUSIONS: The SPARC score represents a valid instrument for predicting CSS after RC for urothelial carcinoma of the bladder. The model can be utilized to better tailor adjuvant therapy and surveillance.
PREDICTION OF CANCER-SPECIFIC SURVIVAL IN PATIENTS WITH RADICAL CYSTECTOMY FOR BLADDER CANCER USING ARTIFICIAL NEURAL NETWORKS Philipp Nuhn, Munich, Germany; Atiqullah Aziz, Hamburg, Germany; Matthias May, Straubing, Germany; Michael Staehler, Munich, Ger€rg Ellinger, many; Michael Gierth, Regensburg, Germany; Jo €ller, Bonn, Germany; Florian Wagenlehner, Stefan C. Mu Wolfgang Weidner, Giessen, Germany; Rudolf Moritz, € nster, Germany; Florian Hartmann, Edwin Herrmann, Mu Marc-Oliver Grimm, Jena, Germany; Chris Protzel, Oliver Hakenberg, €nter Janetschek, Salzburg, Rostock, Germany; Lukas Lusuardi, Gu €rdu €k, Jan Roigas, Berlin, Germany; Austria; Murat Go Maximilian Burger, Regensburg, Germany; Margit Fisch, Hamburg, Germany; Christian G. Stief, Munich, Germany; Patrick Bastian, €sseldorf, Germany; Tobias Grimm, Alexander Buchner*, Munich, Du Germany INTRODUCTION AND OBJECTIVES: The outcome of patients after radical cystectomy (RC) shows substantial variability, and there is still an ongoing search for prognostic parameters. Artificial neural networks (ANN) can be trained to learn and recognize complex patterns of input data. In recent years, ANN have been increasingly used in medical studies, e.g. for the identification of certain patient subgroups. In this study, we utilized ANN for risk stratification of patients after RC. METHODS: The Prospective Multicenter Radical Cystectomy Series 2011 (PROMETRICS 2011) database with 679 consecutive RC patients from 18 European centers was used for this study. The median follow-up time was 21 months, maximum 41 months. Age, body mass index, pack years of smoking, ASA score, TNMG classification, status of the surgical margin, lymphovascular invasion, presence of carcinoma in situ, focality and size of the tumor were used as input data for the ANN (StatSoft, Tulsa, OK, USA). At the time of analysis, 433 complete datasets with all of these data were available. Seventy percent of the cases were selected randomly and used for the training process, and the remaining cases served as two independent validation data sets. Target variable was cancer-specific survival after two years. ANN performance was judged by accuracy and ROC analysis, and the ANN results were compared with regression models. RESULTS: Cancer-specific death occurred in 25% (109/433) of the patients within two years after RC. After network training had been completed, the ANN correctly assigned the survival status to 82% of patients in the training data set and to 81% and 83% in the two validation data sets, respectively. In ROC analysis, the area under curve (AUC) was 0.825 for the whole study cohort. A logistic regression model was built with the same variables that were used for the neural network. The predictive accuracy of the regression model was 78%. CONCLUSIONS: The survival status two years after RC was accurately predicted by ANN, based on clinical and histopathological routine parameters; ANN outperformed the corresponding regression model. Neural networks are a promising approach for risk stratification after RC and may help to optimize the therapeutic strategy. Source of Funding: none