CE_07C: Applied Machine Learning Ideas for Time-to-Event Analyses (Added Fee)

Time-to-event problems are encountered in numerous areas of scientific research. For example, in biomedical studies it is common to evaluate the risk of adverse events, or to predict patient disease-free survival after a medical procedure. In insurance, it is important to characterize customer retention risk profiles and to predict their lifetime value. For e-commerce recommender systems, interest is in describing the distribution of the customer time to recommendation adoption, which may be fully observed (recommendation accepted) or incomplete (postponed decisions). Whether dealing with singular or repetitive events, it is important to recognize the cross-cutting thematic commonality: these problems are all addressable in the time-to-event framework. Risk factor identification or outcome prediction in such settings typically involves high-dimensional covariates, and possibly sparsity. Machine learning methods are now commonly used for variable selection and prediction problems in time-to-event problems. In this short course, we will discuss several fundamental machine learning concepts and techniques, including regularized and boosted Cox regression, survival trees/random forests, and neural networks. Practical implementation via relevant R packages will be discussed using illustrative simulated and real-world datasets. Attendees are expected to be familiar with basic time-to-event and linear regression concepts, as well as R programming at intermediate level. This course will be structured in four one-hour parts, as follows: Part 1 will be an overview of key relevant time-to-event concepts: survival function; hazard function; Kaplan-Meier estimator; Cox proportional hazards model, including partial likelihood development, covariate functional form assessment, time-dependent covariates, and model diagnostics. Illustrative examples will be used to assist in grasping these notions and R code will be provided as appropriate. R packages to be used for illustration purposes will include survival and timereg. Part 2 will consist of an overview of fundamental linear regression and machine learning notions: simple and multiple linear regression, cross-validation and bootstrap, model selection and regularization, regression trees and random forests. Numeric examples in R will be presented to demonstrate these key concepts. W will use the R packages ISLR and glmnet. In Part 3, we will connect time-to-event concepts and machine learning by presenting regularized Cox regression, including ridge, and (adaptive) LASSO/elastic net, and gradient boosted Cox models. For implementation purposes, we will refer to R packages glmnet, CoxBoost, SurvBoost, and ggRandomForests. Part 4 will be dedicated entirely to applications of these methods to real datasets, including examples from biomedical studies. Although attendees will be encouraged throughout the course to ask clarificatory questions, at least 20 minutes will be reserved to clarify and remaining conceptual or implementational aspects.