/machine-learning-for-survival-ecml2020

Code repository for "A General Machine Learning Framework for Survival Analysis" published at ECML 2020

Primary LanguageR

Overview

This is the code repository for "A General Machine Learning Framework for Survival Analysis" published at ECML 2020

The analyses in the publication were based on a prototype implementation of Piece-wise-exponential models (PEMs) using XGBoost in R. This prototype implementation is available as an R package pem.xgb and must be installed to run the benchmarks. The package is available from a separate repository. If you are only interested in how to make XGBoost estimate PEMs head over to that repository. Note that the package is a prototype and should not be used in production. In the near future we are planing to integrate the general framework, including GBT (PEM) into mlr3 via mlr3proba and mlr3pipelines.

If you experience any problems or need advice on how to run the models, don't hesitate to open an issue or contact the first author.

Benchmark experiments

To run the code for the benchmark experiments you first need to install two packages:

# install pem.xgb
devtools::install_github("adibender/pem.xgb") # PEM via XGBoost
devtools::install_github("adibender/pammtools", ref = "ecml")# Data trafo

Folder structure

  • Code for benchmarks based on real data sets is contained within folder benchmark

    • prep_data.R: preprocesses the different data sets used for benchmarking
    • problems.R: contains functions that generate an instances during each iteration of the benchmark
    • algorithms.R: contains wrapper functions for the different algorithms that
    • are compared to each other
    • deep_c_index.R: an R version of the C-index as implemented in the DeepHit package
    • calling_deephit.R: an R version of the DeepHit package wrapped in a random search function
    • benchmark-deephit.R: contains the code to perform the random search for DeepHit on all benchmark data sets
    • benchmark.R: contains the code that will performs the benchmark experiments and is based on package batchtools.
    • bench-time.R: benchmark w.r.t. different strategies w.r.t. cut-point selection (scaling experiment)
    • benchmark-metabric.R: benchmarks for metabric data set for comparison with DeepHit
    • benchmark-pbc-tvf: benchmark of the PBC data sets (with time-dependent covariates)
    • benchmark-mgus2.CR: benchmark of the MGUS data with competing risks for comparison with DeepHit
    • extract-benchmark-instances.R script that extracts results from benchmark as well as instances to facilitate comparison with DeepHit

  • Code for benchmarks based on synthetic data sets is contained within folder simulation

    • sim_funs.R contains functions that simulate data for right-censored and competing risks data
    • tv-effects.R: contains code to run benchmark w.r.t. to performance in presence of time-varying effects/non-proportional hazards (right-censored and competing risks)

  • Code for the aggregation and evaluation of benchmark studies are contained in folder paper

    • results.R: reads in the results from different experiments, combines results, creates raw tables (manually processed for publication)

Usage of DeepHit wrapper in R

In order to use DeepHit in R, clone the DeepHit repository into the benchmark folder. The calling_deephit.R wrapper that calls DeepHit from R needs a suitable Python environment (Python 2.7) and will import the functions class_DeepHit.py, import_data.py and utils_eval.py.