Skip to content

jaredleekatzman/DeepSurv

Repository files navigation

DeepSurv

DeepSurv implements a deep learning generalization of the Cox proportional hazards model using Theano and Lasagne.

DeepSurv has an advantage over traditional Cox regression because it does not require an a priori selection of covariates, but learns them adaptively.

DeepSurv can be used in numerous survival analysis applications. One medical application is provided: recommend_treatment, which provides treatment recommendations for a set of patient observations.

For more details, see full paper DeepSurv: Personalized Treatment Recommender System Using A Cox Proportional Hazards Deep Neural Network.

For an updated implementation of the Cox loss function in PyTorch, please see Feature Selection using Stochastic Gates (STG) by Yamada et al..

Installation:

From source

Download a local copy of DeepSurv and install from the directory:

git clone https://github.com/jaredleekatzman/DeepSurv.git
cd DeepSurv
pip install .

Dependencies

Theano, Lasagne (bleeding edge version), lifelines, matplotlib (for visualization), tensorboard_logger, and all of their respective dependencies.

Running the tests

After installing, you can optionally run the test suite with

py.test

from the command line while in the repo's main directory.

Running Experiments

Experiments are run using Docker containers built off of the floydhub deep learning Docker images. DeepSurv can be run on either the CPU or the GPU with nvidia-docker.

All experiments are in the DeepSurv/experiments/ directory.

To run an experiment, define the experiment name as an environmental variable EXPRIMENTand run the docker-compose file. Further details are in the DeepSurv/experiments/ directory.

Training a Network

Training DeepSurv can be done in a few lines. First, all you need to do is prepare the datasets to have the following keys:

{ 
	'x': (n,d) observations (dtype = float32), 
 	't': (n) event times (dtype = float32),
 	'e': (n) event indicators (dtype = int32)
}

Then prepare a dictionary of hyper-parameters. And it takes only two lines to train a network:

network = deepsurv.DeepSurv(**hyperparams)
log = network.train(train_data, valid_data, n_epochs=500)

You can then evaluate its success on testing data:

network.get_concordance_index(**test_data)
>> 0.62269622730138632

If you have matplotlib installed, you can visualize the training and validation curves after training the network:

deepsurv.plot_log(log)