MSc Stats Project Presentation "Exploring Prediction Error in High-Dimensional Data with Cox-Boosting and Random Survival Forests: A Simulation Study and Applications to DLBCL Data" by Ronne Tamayo

CANDIDATE:  Ronne Tamayo

ABSTRACT:

The Cox proportional hazards model is one of the most widely-used approaches for estimating the effect of covariates and predicting risk in survival analysis. However, prediction performance for the Cox model is highly affected by the presence of high-dimensional data. This project outlines three alternative approaches from the machine learning field that can be used to overcome limitations of the Cox proportional hazards model. Two different boosting adaptations of the Cox model, in addition to random survival forests, are evaluated using prediction error measures. Expected loss functions are computed for the Brier and C-index scores and evaluated for right-censored data simulated from R-CHOP for Diffuse Large B Cell Lymphoma and a Cox-Exponential distribution. Prediction performance is analyzed with varying censoring percentages in the presence of a considerably high number of predictors. Results indicate that the Cox model is the strategy with the lowest prediction performance, when the number of covariates and censoring percentages increase.