Double-blind evaluation and benchmarking of survival models in a multi-centre study

Accurate modelling of time-to-event data is of particular importance for both exploratory and predictive analysis in cancer, and can have a direct impact on clinical care. This study presents a detailed double-blind evaluation of the accuracy in out-of-sample prediction of mortality from two generic non-linear models, using artificial neural networks benchmarked against a partial logistic spline, log-normal and COX regression models. A data set containing 2880 samples was shared over the Internet using a purpose-built secure environment called GEOCONDA (www.geoconda.com). The evaluation was carried out in three parts. The first was a comparison between the predicted survival estimates for each of the four survival groups defined by the TNM staging system, against the empirical estimates derived by the Kaplan-Meier method. The second approach focused on the accurate prediction of survival over time, quantified with the time dependent C index (C^td). Finally, calibration plots were obtained over the range of follow-up and tested using a generalization of the Hosmer-Lemeshow test. All models showed satisfactory performance, with values of C^td of about 0.7. None of the models showed a systematic tendency towards over/under estimation of the observed survival at τ = 3 and 5 years. At τ = 10 years, all models underestimated the observed survival, except for COX regression which returned an overestimate. The study presents a robust and unbiased benchmarking methodology using a bespoke web facility. It was concluded that powerful, recent flexible modelling algorithms show a comparative predictive performance to that of more established methods from the medical and biological literature, for the reference data set.