💡 Penalized Competing Risks Analysis using Case-base Sampling

📍 Repository with my code and my final report for my MSc. thesis

🎯 We develop an alternative penalized cause-specific hazards model that extends on the casebase package for competing risks survival analysis of high-dimensional biological data

📂 Repository organization

• bash_script_template - Template for bash scripts for Compute Canada. Creates individual bash scripts to be submitted to the cluster for one simulation run. Requires runscripts and logs folder to be setup

• doc - Thesis write up using ubcdiss template

• Final_results - Contains all final scripts,figures and results generated for the thesis for variable selection and prediction performance

• mtool - R package for fitting a penalized multinomial regression model with the K-1 logit parameterization

• mtool_fit_diagnostics - Documentation of known issues with mtool

• Papers - Papers for background information. To be updated.

• simulation_scripts - Template script for simulation from a two-cause model using the replicate function

• src - Helper fitting functions including the cross-validation function, simulation functions, and survival performance metrics functions

• survsim_mod - Contains modified funtions from the survsim R package to generate competing risks data with normally distributed covariates having 1) AR(1) correlation structure and 2) Block correlation structure

• updates - Weekly meeting updates

📃 Abstract

Genome-wide transcriptome profiling and advances in experimental technologies have greatly increased the generation of high-dimensional genomic data, particularly microarray data correlated with survival outcomes such as patient survival time or time to cancer relapse. The analysis of such genomic time-to-event data becomes more complicated when there are competing events, i.e., the failure of a patient can occur due to one of multiple distinct causes. We develop an alternate elastic-net penalized competing risks analysis method that is able to produce easily interpretable hazard ratios akin to the Cox regression model. This approach is also able accurately produce smooth-in-time predicted estimates patient risk, in a variety of settings, such as non-proportional hazards as well. We examine the performance of this method in a simulation study, looking at both variable selection, as well as patient risk estimation performance in both the $N > p$ as well as the $p > N$ scenarios.

🎺 The original model is based on the casebase paper by Sahir Bhatnagar, Maxime Turgeon, Jesse Islam and James Hanley, which is based on the sampling methodology proposed by James Hanley and Olli Miettinen.

💻 The optimization for the stochastic gradient descent in mtool was written by Dr. Yi Lian.

📌 The relaxed LASSO branch contains a relaxed LASSO implementation for the casebase penalized model (WIP) by Alex Romanus.