Accelerated and Interpretable Oblique Random Survival Forests.

Area of Science:

• Machine Learning
• Biostatistics
• Computational Statistics

Background:

• Oblique random survival forests (RSF) offer high prediction accuracy for right-censored data.
• Standard RSF uses single predictors per tree, while oblique RSF uses linear combinations, increasing computational cost.
• Limited methods exist for estimating variable importance (VI) in oblique RSFs.

Purpose of the Study:

• To enhance the computational efficiency of oblique RSF.
• To introduce a reliable method for estimating variable importance (VI) in oblique RSFs.
• To provide an accessible R package (aorsf) for these methods.

Main Methods:

• Implemented a computationally efficient oblique RSF using Newton-Raphson scoring.
• Developed a 'negation VI' method by assessing the impact of predictor coefficients on out-of-bag accuracy.
• Benchmarked against existing oblique RSF software and compared VI methods via simulation.

Main Results:

• The new oblique RSF implementation is hundreds of times faster than existing software, maintaining prediction accuracy.
• 'Negation VI' demonstrated superior performance in discriminating relevant from irrelevant numeric predictors compared to permutation VI, Shapley VI, and ANOVA-based VI.
• The aorsf R package provides access to these advanced oblique RSF methods.

Conclusions:

• The developed methods significantly improve the speed and usability of oblique RSF for survival analysis.
• The 'negation VI' method offers a more accurate approach to variable importance assessment in oblique RSFs.
• These advancements facilitate broader application of sophisticated survival modeling techniques.