Skip to content

Latest commit

 

History

History
69 lines (49 loc) · 2.71 KB

README.md

File metadata and controls

69 lines (49 loc) · 2.71 KB
Raw

megalearner

megalearner is series of R scripts, found in the folder R, that implement the illustrative example described in the paper Gonzalez Ginestet, P. et al. (2021). “Survival stacking with multiple data types using pseudo-observation-based-AUC loss”.

megalearner builds a more accurate risk prediction model than standard methods such as Cox proportional hazard or single machine learning algorithms that have been already adapted to handle right-censored observations. megalearner optimally stacks predictions from any method based on different survival outcome types (either binary or continuous). In the paper, we show that our proposed method can improve on single survival based methods such as survival random forest or on other strategies that use a pre-processing step such as inverse probability of censoring weighted or pseudo-observations.

We illustrate our methodology in two datasets:

  1. We use the Rotterdam data set as training data set and we validate the predictions in an external data set. For that, we use the data from the German Breast Cancer Study Group (GBSG). Both data sets are available in the survival R package [Therneau, 2000].

  2. We use the METABRIC cohort for two endpoints: overall survival (OS) and recurrence-free survival (RFS) in months.

Organization of the folders

The scripts depend on each other in the following way:

03-validate.R -> depends on -> 02-opt-stack.R -> depends on -> 01-train-models.R -> depends on 00-model-specs.R

01-train-models.R: saves the fit of each algorithm (“fullfits.rds”), the matrix Z (“Zout.rds”) where the predictions of each algorithms are recorded and the folds used in the cross-validation (“split.rds”)

02-opt-stack.R: saves the optimal coefficient used to combine the algorithms (“opt-coeffs.rds”) and the AUC at 5 years cross-validated of the stacking (“cv-aucs-stack.rds”)

The folder METABRIC contains the R codes to run the procedure in the METABRIC dataset for the two predictive examples.

The folder figure contains the plots in the paper.

Installation

Depends on the following R packages: broom, patchwork, ggplot2, xtable, class, glmnet, e1071, riskRegression, CoxBoost, prodlim, Matrix, randomForestSRC, survival, eventglm, pseudoloss, splines, stats, graphics, grDevices, utils, datasets, methods, base, plotROC

The following packages CoxBoost, eventglm and pseudoloss can be installed as follows:

remotes::install_github("binderh/CoxBoost")
remotes::install_github("sachsmc/eventglm")
remotes::install_github("sachsmc/pseudoloss")