Survival Analysis using Deep Learning for Prediction
SurvivalDNN
SurvivalDNN is a library for survival analysis using deep learning, with a focus on prediction tasks and ease-of-use. Survival analysis models the time until events occur, for example the failure time of industrial components, patient survival times in medicine, customer churn in business, and credit risk in finance.
Getting Started
Install from PyPi:
$ pip install survivaldnn
Usage
- Specify the model architecture, loss function, and other hyperparameters
from survivaldnn import SurvivalDNNModel
model = SurvivalDNNModel()
# Default: Residual network with negative loglikelihood loss
model.compile(
numFeatures,
numSupport=100,
loss='loglik',
architecture='resnet',
layers=3
)
- Train the model
model.fit(X, Y,
numSupport=100,
epochs=100)
- Make predictions of interest
# Estimated time until failure
model.predict(X)
# Estimated time until failure, given that `elapsed` time has occurred
model.predict(X, elapsed)
# Estimated full survival function and support
survFunc, support = model.predict_survival_function(X)
For a more in-depth example, see this notebook.
Technical Note
This library builds on and draws from other proposed methodologies such as DeepSurv (Katzman et al., 2018), Nnet-Survival (Gensheimer and Narasimhan, 2018), and DNNSurv (Zhao and Feng, 2021), with a few modifications for practical usage, stability during training, and efficiency.
With discrete-time survival models, the outcome space is discretized into a finite number of support points ${t_1,\dots, t_k}$ with $t_{i-1} < t_{i}$ such that
$$\text{supp}(Y) = \bigcup_{i=1}^{k} (t_{i-1}, t_i]$$
Deep learning approaches to survival analysis often model the conditional survival probabilities $P(Y > t_j|X, Y > t_{j-1})$, or the probability of surviving each time interval given that an individual has survived up to the previous interval. In this library, the outputs of the neural network are instead a $k$-dimensional vector that models the survival function directly:
$$\Big( P(Y > t_1|X), P(Y > t_2|X), \dots, P(Y > t_k|X) \Big)$$
which is done by enforcing monotonicity restrictions in the final layer, such that $P(Y > t_{i-1}|X) \geq P(Y > t_i|X)$. The main reasoning for doing so is that for prediction tasks, interest is typically on statistics based on features of $P(Y > t_i|X)$. We can of course still obtain the marginal survival probabilities using the conventional approach with
$$P(Y > t_j|X) = \prod_{i}^{j} P(Y > t_i|X, Y > t_{i-1})$$
However this can be numerically unstable and can compound small errors especially when the number of support points is large. By modeling the survival function directly with the additional structure from enforcing monotonicity, this implementation can be more stable during training as well as being more computationally efficient for large-scale prediction tasks.
References
Gensheimer, M. F., & Narasimhan, B. (2019). A scalable discrete-time survival model for neural networks. PeerJ, 7, e6257. https://doi.org/10.7717/peerj.6257
Katzman, J. L., Shaham, U., Cloninger, A., Bates, J., Jiang, T., & Kluger, Y. (2018). DeepSurv: Personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Medical Research Methodology, 18(1), 24. https://doi.org/10.1186/s12874-018-0482-1
Zhao, L., & Feng, D. (2020). Deep neural networks for survival analysis using pseudo values. IEEE Journal of Biomedical and Health Informatics, 24(11), 3308–3314. https://doi.org/10.1109/JBHI.2020.2980204
Metadata
Release files for SurvivalDNN 1.0.1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| survivaldnn-1.0.1.tar.gz | 11.1 kB | Without provenance | Details for survivaldnn-1.0.1.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| survivaldnn-1.0.1-py3-none-any.whl | Python 3 | none | any | Details for survivaldnn-1.0.1-py3-none-any.whl |
Total release size: 22.3 kB
Release files / survivaldnn-1.0.1.tar.gz
| Download URL | survivaldnn-1.0.1.tar.gz |
|---|---|
| Size | 11.1 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
d2fda29acdcd89764b0b6692e870a534a9e9e49eadbf2b97b983920bc5dbccd4
|
|
BLAKE2b-256 checksum How to use checksums |
87e2ceeccaef7f88c2388dfa22774f99c7f15b760f03388fed21d8db905661b2
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.1.0 CPython/3.10.14
|
Release files / survivaldnn-1.0.1-py3-none-any.whl
| Download URL | survivaldnn-1.0.1-py3-none-any.whl |
|---|---|
| Size | 11.1 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
cd98c4bc1203de5c702662ded5dc54d11dcc0ad25da9e26b9f1fe7514b122f9f
|
|
BLAKE2b-256 checksum How to use checksums |
03a8286d15a17425c26525f249fad437899f0911af7688042a0c9d2a14e72e58
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.1.0 CPython/3.10.14
|
Security for SurvivalDNN 1.0.1
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.