aorsf

accelerated oblique random survival forests

By Byron C Jaeger in Machine learning survival analysis R package

July 30, 2022

website code
hex1 hex1 hex1 hex1 hex1

Project Status: WIP – Initial development is in progress, but therehas not yet been a stable, usable release suitable for thepublic. Codecov testcoverage R-CMD-check pkgcheck Status at rOpenSci Software PeerReview

aorsf provides optimized software to fit, interpret, and make predictions with oblique random survival forests (ORSFs).

Why aorsf?

  • over 400 times faster than obliqueRSF.

  • accurate predictions for time-to-event outcomes.

  • negation importance, a novel technique to estimate variable importance for ORSFs.

  • intuitive API with formula based interface.

  • extensive input checks + informative error messages.

Installation

You can install the development version of aorsf from GitHub with:

# install.packages("remotes")
remotes::install_github("bcjaeger/aorsf")

Example

The orsf() function is used to fit ORSFs. Printing the output from orsf() will give some descriptive statistics about the ensemble.

library(aorsf)

fit <- orsf(data = pbc_orsf,
            formula = Surv(time, status) ~ . - id)

print(fit)
#> ---------- Oblique random survival forest
#> 
#>           N observations: 276
#>                 N events: 111
#>                  N trees: 500
#>       N predictors total: 17
#>    N predictors per node: 5
#>  Average leaves per tree: 24
#> Min observations in leaf: 5
#>       Min events in leaf: 1
#>           OOB stat value: 0.84
#>            OOB stat type: Harrell's C-statistic
#> 
#> -----------------------------------------

How about interpreting the fit?

  • use orsf_vi_negate() and orsf_vi_anova() for variable importance

    orsf_vi_negate(fit)
#>          bili           age       protime       ascites       albumin 
#>  0.0129714524  0.0125026047  0.0085955407  0.0063554907  0.0061992082 
#>       spiders           sex        copper         edema           ast 
#>  0.0057303605  0.0049489477  0.0044280058  0.0027783566  0.0018753907 
#>        hepato          trig         stage      alk.phos      platelet 
#>  0.0018232965  0.0016670140  0.0005209419 -0.0006772244 -0.0008856012 
#>          chol           trt 
#> -0.0016670140 -0.0025526151

  • use orsf_pd_ice() or orsf_pd_summary() for individual or aggregated partial dependence values.

    orsf_pd_summary(fit, pd_spec = list(bili = c(1:5)))
#>    bili      mean        lwr      medn       upr
#> 1:    1 0.2376652 0.01286480 0.1331328 0.8669733
#> 2:    2 0.2904995 0.04154464 0.1889294 0.8943053
#> 3:    3 0.3441193 0.06285072 0.2538335 0.9146453
#> 4:    4 0.3973732 0.10091026 0.3247072 0.9271855
#> 5:    5 0.4424865 0.14005297 0.3807676 0.9315099

  • use orsf_summarize_uni() to show the top predictor variables in an ORSF model and the expected predicted risk at specific values of those predictors. (The term ‘uni’ is short for univariate.)

    # take a look at the top 5 variables 
# for continuous predictors, see expected risk at 25/50/75th quantile
# for categorical predictors, see expected risk in specified category

orsf_summarize_uni(object = fit, n_variables = 5)
#> 
#> -- bili (VI Rank: 1) ---------------------------
#> 
#>        |---------------- risk ----------------|
#>  Value      Mean    Median     25th %    75th %
#>   0.80 0.2337275 0.1267164 0.04190426 0.3716894
#>   1.40 0.2526511 0.1452163 0.05476829 0.3998048
#>   3.52 0.3730196 0.2841226 0.16033444 0.5673070
#> 
#> -- age (VI Rank: 2) ----------------------------
#> 
#>        |---------------- risk ----------------|
#>  Value      Mean    Median     25th %    75th %
#>   41.5 0.2736650 0.1332843 0.04270990 0.4575449
#>   49.7 0.2993351 0.1656678 0.04910598 0.5203980
#>   56.6 0.3314766 0.2113831 0.06956163 0.5472662
#> 
#> -- protime (VI Rank: 3) ------------------------
#> 
#>        |---------------- risk ----------------|
#>  Value      Mean    Median     25th %    75th %
#>   10.0 0.2815397 0.1430703 0.04704506 0.4987604
#>   10.6 0.2947893 0.1526181 0.05252053 0.5393637
#>   11.2 0.3176024 0.1831842 0.06925762 0.5566000
#> 
#> -- ascites (VI Rank: 4) ------------------------
#> 
#>        |---------------- risk ----------------|
#>  Value      Mean    Median     25th %    75th %
#>      0 0.2954662 0.1462719 0.04778412 0.5409697
#>      1 0.4597940 0.3742798 0.25999819 0.6551916
#> 
#> -- albumin (VI Rank: 5) ------------------------
#> 
#>        |---------------- risk ----------------|
#>  Value      Mean    Median     25th %    75th %
#>   3.31 0.3176012 0.1789175 0.05525473 0.5785138
#>   3.54 0.2934144 0.1527704 0.04236004 0.5225606
#>   3.77 0.2791547 0.1412617 0.04199867 0.4937002
#> 
#>  Predicted risk at time t = 1788 for top 5 predictors

References

Byron C. Jaeger, D. Leann Long, Dustin M. Long, Mario Sims, Jeff M. Szychowski, Yuan-I Min, Leslie A. Mcclure, George Howard, Noah Simon (2019). Oblique Random Survival Forests. Ann. Appl. Stat. 13(3): 1847-1883. URL https://doi.org/10.1214/19-AOAS1261 DOI: 10.1214/19-AOAS1261

Funding

The developers of aorsf receive financial support from the Center for Biomedical Informatics, Wake Forest University School of Medicine. We also receive support from the National Center for Advancing Translational Sciences of the National Institutes of Health under Award Number UL1TR001420.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.