Presentation Authors: David Thurtle, Cambridge, United Kingdom, Ola Bratt, Gothenburg, Sweden, Pär Stattin, Uppsala, Sweden, Paul Pharoah, Vincent Gnanapragasam*, Cambridge, United Kingdom
Introduction: PREDICT Prostate is a novel prognostic model that provides individualised long-term prostate cancer-specific and overall survival estimates using baseline clinico-pathological and co-morbidity data. The model, derived from UK data and validated in an East Asian cohort, estimates potential treatment benefit. Here we aimed to:- Externally validate the model in a large independent dataset- Review performance by treatment groups
Methods: Data on age, PSA, clinical T stage, grade group, biopsy involvement, primary treatment and comorbidity were retrieved from the nation-wide population-based Prostate Cancer data Base Sweden (PCBase). Men with non- metastatic prostate cancer and PSA < 100 ng/ml diagnosed between 2000 and 2010 were included. _x000D_
15-year survival estimates were calculated using the PREDICT Prostate algorithm, with coefficients for each of the above parameters used alongside baseline hazards to calculate prostate cancer -specific mortality (PCSM) and all-cause mortality (ACM) in a competing risk model. Discrimination was assessed using Harrells concordance(c)-index for right-censored data. Calibration was evaluated using actual follow-up to censorship at 15 years. Data analyses were performed in Stata (TX, USA) and R.
Results: 69,206 men were included with 13 years of median follow-up. Overall discrimination of PREDICT was excellent with C-indices of 0.85 for PCSM and 0.79 for ACM. Calibration of the model was excellent with 25,925 deaths predicted and 25,849 deaths observed in PCBaSe. _x000D_
20,384 men underwent conservative management and 32,842men received radical treatment. Within these treatment groups c-index for 15-year PCSM was 0.81 and 0.78 respectively. Among men on well-defined active surveillance, c-indices were further improved at 0.88 for PCSM and 0.75 for ACM. Calibration also remained good within treatment groups with differences between observed and predicted numbers of overall deaths 1.4%, 2.4% and 3.1% among men who received active surveillance, external beam radiotherapy, and prostatectomy respectively.
Conclusions: This large external validation demonstrates that PREDICT Prostate is a robust and generalisable model. Unlike available models it incorporates competing risk for death and provides individualised prognostic estimates of mortality, with and without upfront radical treatment, for up to 15 years from diagnosis.
Source of Funding: The Urology Foundation (UK Charity)