dcoxph

About

This repository hosts the R implementation of the distributed Cox Proportional Hazards algorithm as described by Lu et al. that can be used with the Personal Health Train distributed learning infrastructure.

The repository has been forked from the master branch. This repository has been tested with the new ppDLI infrastructure.

How to use this algorithm?

1. Getting the code

First, clone the repository and enter the directory:

git clone https://github.com/IKNL/dcoxph.git
cd dcoxph

2. Run the code in R-Studio

In order to run the algorithm, the following two approaches can be taken The following steps assume you have R (and RStudio) and git installed. If you run into trouble, please create an issue in the tracker.

2.1 Installing dependencies

Next, install the required packages in R. Either run the following in bash:

RScript install_packages.R

or run the following in R:

packages <- c(
  "abind",
  "dplyr",
  "httr",
  "rjson"
)

install.packages(packages)

2.2 Running the algorithm

This step assumes you have access to a central server and know your username, password and collaboration id.

A researcher then runs the analysis by:
i. Creating a client that communicates with the distributed learning infrastructure
ii.Calling the method dcoxph with the appropriate parameters

This is illustrated by the following R code: ```R source("Client.R") source("dl_coxph.R")

      # Create a client object to communicate with the server.
      #Example :  client <- Client("http://137.117.138.98:5000/api", "myname", "my_password", 2)

      client <- Client(host, username, password, collaboration_id)   
      client$authenticate()

      # Parameters used to interpret the hub's datastore
      expl_vars <- c("Age","Race2","Race3","Mar2","Mar3","Mar4","Mar5","Mar9",
                   "Hist8520","hist8522","hist8480","hist8501","hist8201",
                   "hist8211","grade","ts","nne","npn","er2","er4")
      time_col <- "Time"
      censor_col <- "Censor"

      results <- dcoxph(client, expl_vars, time_col, censor_col)
      ```

or, you can run the test_maastro.R script with correct username, password and collaboration id.

3. Run the code in Jupyter Notebook

2.1 Installing dependencies

Install [anaconda distribution] (https://www.anaconda.com/distribution/)
Install r package in anaconda run jupyter notebook in anaconda

4. Optional: encapsulating the distributed code in a Docker image

The code is split into a local and a distributed part. Both parts are implemented in the same R script dl_coxph.R. The Docker registry at https://docker-registry.distributedlearning.ai already hosts an image with the distributed code. If you are using your own installation of the infrastructure, you/the researcher should create a Docker image that holds the distributed code (see also build_docker.sh) and push the image to a (private) Docker registry. This requires Docker to be installed on the machine.

If you do not have access rights to https://docker-registry.distributedlearning.ai, a similar image can be extracted from here

For an overview of the working of the algorithm, see the figure below: