Machine Learning for Risk Prediction using Oblique Random Survival Forests
Byron Casey Jaeger, Assistant Professor at Wake Forest University School of Medicine
Risk prediction can reduce the burden of disease by guiding strategies for prevention and treatment. The random survival forest (RSF) is frequently used for risk prediction due to its minimal assumptions and its aversion to overfitting. Decision trees in the RSF traditionally use one variable at a time to grow new branches. Using linear combinations of variables instead of a single variable to grow new branches, a technique known as 'oblique' splitting, can substantially improve prediction accuracy for the RSF. However, few R packages offer oblique splitting because its computational complexity scales poorly. The 'aorsf' package fills this gap, offering computationally efficient oblique splitting for RSFs and an intuitive design for fitting, interpreting, and computing predictions with them. In this session, Dr. Byron Jaeger covers the basics of risk prediction and will show how to fit, tune, interpret, and predict risk with oblique RSFs using aorsf. Attendees also learned how to include oblique RSFs in machine learning workflows with both tidymodels and mlr3, and how to customize oblique RSFs with user-defined functions for finding oblique splits.
PDF: https://static.sched.com/hosted_files...
Main Sections
0:00 Introduction
1:49 Machine learning
2:28 Supervised learning
4:05 Decisions trees and random forests
8:29 Random forest
17:13 Random survival forest (RSFs)
18:05 Oblique RSFs
21:02 Demo with aorsf
39:50 Variable importance
43:45 But how accurate is aorsf’s predicted risk?
45:58 Benchmarking aorsf
51:40 Extending aorsf
57:40 Closing thoughts
More Resources
R/Medicine Event site: https://events.linuxfoundation.org/r-...
Main Site: https://www.r-consortium.org/
News: https://www.r-consortium.org/news
Blog: https://www.r-consortium.org/news/blog
Join: https://www.r-consortium.org/about/join
Twitter: / rconsortium
LinkedIn: / r-consortium
Mastodon: https://fosstodon.org/@RConsortium