Evaluation of Different Survival Analysis Models for Nki Breast Cancer Data

Abstract

The goal of this study is to evaluate different Survival Analysis Models in terms of their predictive capabilities, accuracy in determining significant covariates within the data, as well as their respective results compared across standard indices. Highest Concordance Index and Lowest Akaike Information Criterion (AIC) are used as the basis of selecting the ideal Survival Analysis model as a template for the construction of the Survival Prediction model for NKI Breast Cancer Data. 6 Survival Analysis Models were used in this study. For the semi-parametric survival models, Classical Cox, Cox-Lasso, and Cox-Ridge Regressions. For the parametric models, 3 Accelerated Failure Time (AFT) models were implemented. These are: Weibull AFT, Log-logistic AFT, and Log-Normal AFT models. Right-censoring was performed in the data since it has been assumed that there are subjects which were not called back anymore for the entire, 18-year clinical trial where the data was taken from. A proportional hazards test was then performed to find out if the covariates in the data are fit to be modeled using Cox Regression and its derivatives. A test for the distribution on the time of event was also done to find whether it follows a specific distribution or not. This was done to verify the usability of the parametric survival analysis models on the data. It has been found out that in terms of Concordance Index and AIC, the Cox-Ridge Regression model outperforms its 2 other semi-parametric counterparts, having the least AIC of 752.6703 and Highest Concordance Index of 0.7709. As for the other 3 parametric models, Log-Normal AFT outperformed the Weibull AFT and Log-Logistic AFT models by a Concordance Index of 0.780 with a corresponding AIC of 608.822. This result also suggests that the time of event of the subjects is best fitted by Log-Normal Distribution. By comparing the 2, best-performing models, it has been reported that Log-Normal AFT outperforms Cox-Ridge Regressions, therefore suggesting to use this Parametric Survival Analysis Model as the basis for a Survival Prediction model suited for NKI Breast Cancer data.