Neural Fine Gray

This repository allows to reproduce the results in Neural Fine Gray - Please use the release: CHIL for exact reproduction.
A neural network approach to the problem of competing risks, leveraging monotone neural networks to model the cumulative incidence functions.

Model

The model consists in two neural networks: one models the cumulative incidence function and the other the balance to ensure that they add up to one.

How to use ?

To use the model, one needs to execute:

from nfg import NeuralFineGray
model = NeuralFineGray()
model.fit(x, t, e)
model.predict_risk(x, risk = 1)

With x, the covariates, t, the event times and e, the cause of end of follow up (0 is censoring).

A full example with analysis is provided in examples/Neural Fine Gray on FRAMINGHAM Dataset.ipynb.

Reproduce paper's results

To reproduce the paper's results:

0. Clone the repository with dependencies: git clone git@github.com:Jeanselme/NeuralFineGray.git --recursive
1. Create a conda environment with all necessary libraries pycox, lifelines, pysurvival
2. Add path export PYTHONPATH="$PWD:$PWD/DeepSurvivalMachines:$PYTHONPATH"
3. Run examples/experiment_competing_risk.py FRAMINGHAM to run all models on the FRAMINGHAM dataset
4. Repeat with PBC, SYNTHETIC_COMPETING and SEER to run on each dataset
5. Analysis using examples/Analysis.ipynb to measure performance

Note that you will need to export the SEER dataset from https://seer.cancer.gov/data/. The previous scripts allow you to reproduce all the models presented in the paper except the Fine-Gray appraoch that requires: 0. Install R and the libraries: riskRegression, prodlim, survival, cmprsk and readr

1. Create a folder data/ in examples/ to save the generated files
2. Run examples/process_data.py FRAMINGHAM to create a csv files with the same data split used in the Python scripts
3. Run examples/FineGray.R to create the predictions of a Fine-Gray model (Note that you will need to change the content of this file for running on a subset of datasets)

Compare to a new method

Adding a new method consists in adding a child to Experiment in experiment.py with functions to compute the nll and fit the model. Then, add the method in examples/experiment_competing_risk.py and follow the previous point. TODOs have been added to make the addition of a new method easier.

Setup

Structure

We followed the same architecture than the DeepSurvivalMachines repository with the model in nfg/ - only the api should be used to test the model. Examples are provided in examples/.

Clone

git clone git@github.com:Jeanselme/NeuralFineGray.git --recursive

Requirements

The model relies on DeepSurvivalMachines, pytorch, numpy and tqdm.
To run the set of experiments pycox, lifelines, pysurvival are necessary.