Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/pr38/survivalpredict

Estimating survival predictions
https://github.com/pr38/survivalpredict

cox-proportional-hazard cox-regression python reliability-analysis scikit-learn survival-analysis

Last synced: 13 days ago
JSON representation

Estimating survival predictions

Awesome Lists containing this project

README 

          
# SurvivalPredict



A python package centered around Survival Analysis Statistical Learning, for predicting survival curves. The code in this repo is lovingly written without any stochastic generative processes.



See documentation



WIP. A pypi release should be released soon. In the meantime, the code in this repo can be installed via `pip install git+https://github.com/pr38/survivalpredict`. Ideally, before the first pypi release docstrings and example notebooks will be added. With the goal of finishing left-censoring support, sparse data support as well as  tree-based, ensemble, and exotic neural network  models further down the line. 



General walkthrough-demo



Demo for interfacing with scikit learn



## Estimators

The estimators implemented in the `survivalpredict.estimators` sub-module.



    

        Estimators

        Description

        Stratifiable

        Left-censorable

    

    

        CoxProportionalHazard

        

        Cox Proportional Hazards model is a linear semi-parametric relative risk model. A staple of survival analysis. Fast and efficient to train. Survivalpredict's implementation has many optimizations and is faster than other implementations available to Python. Both breslow and efron ties are supported. Currently, only the Breslow base hazard is available.

        

         Yes

        Yes

    

    

        ParametricDiscreteTimePH

         A fully parametric linear hazards model. Chen, weibull, log_normal, log_logistic, gompertz, gamma and additive_chen_weibull baseline hazards are available as hyperparameters. Maximum likelihood is estimated using a survival distinct time likelihood with censorship. Implemented with Pymc/Pytensor, with either a Jax or numba backend.

        Yes

        Yes

    

        

        CoxPHElasticNet

        Cox Proportional Hazards model model with Elastic-Net/Lasso penalty and feature shrinkage/selection. Uses a 'Newton Raphson-like' coordinate descent algorithm described in Simon, Noah et al. “Regularization Paths for Cox's Proportional Hazards Model via Coordinate Descent. Assumes breslow ties. The current literature is unclear on how to incorporate stratification support into said algorithm.

        

         No

        Yes

    

    

        KaplanMeierSurvivalEstimator

         Univariate non-parametric survival curve. Useful as a baseline/dummy estimator.

        Accepts strata, but builds a survival curve for each strata. 

        Yes

    

        

        KNeighborsSurvival

        K nearest neighbors for survival. An in-memory non-parametric model that builds a Kaplan-Meier survival curve based on neighbors.

        

        No

        Yes

    

        CoxNeuralNetPH

         A neural network model for estimating relative risk. Cox proportional hazards model's 'negative log likelihood for Breslow ties' is used as a loss function. Breslow's base hazard for relative risk is used to estimate survival across time. Implemented using Jax.  

        Yes

        Yes

    

        

        AalenAdditiveHazard

         Linear multivariate non-parametric estimation of hazard. Allows for each interval of time and feature to have an associated coefficient, allowing for the effects of features to change over time.

        

        No

        Yes



## Metrics



Survivalpredict focuses on metrics that directly measure prediction performance. Hence, the `survivalpredict.metrics` module intentionally excludes metrics based on ranking relative risk(i.e., ' c-index').



    

        Metrics

        Description

    

    

        brier_scores_administrative

        Squared error between the true survival and prediction for each time of interest. Censored intervals are ignored. Averaged by the number of rows not censored at a given interval of time. Ideal in cases of 'administrative' censorship, where 'survival time' is modeled after the time of an individual in the experiment, and not calendar time. This metric is ideal for cases of churn, conversion and operational failure. See here.

    

    

        integrated_brier_score_administrative

        Integral of administrative brier scores, to allow for a singular metric of performance. 

    

        integrated_brier_score_administrative_sklearn_metric

        scikit-learn metric wrapper around `integrated_brier_score_administrative` function, for accessing said metric in when using the SklearnSurvivalPipeline wrapper class when interfacing with scikit-learn.

    

    

        integrated_brier_score_administrative_sklearn_scorer

        scikit-learn scorer wrapper around `integrated_brier_score_administrative` function, for accessing said metric in when using the SklearnSurvivalPipeline wrapper class when interfacing with scikit-learn.

    

    

        brier_scores_ipcw

        Brier scores with inverse probability of censoring weights. The squared error between the true survival and prediction is weighted using a Kaplan-Meier curve with inverted events, depending on censoring and failure at different points in time. This is a common metric within the field of biostatistics and is used in clinical trials. See here.

    

    

        integrated_brier_score_ipcw

        Integral of brier scores with probability of censoring weights, to allow for a singular metric of performance.

    

    

        integrated_brier_score_ipcw_sklearn_metric

        scikit-learn metric wrapper around `integrated_brier_score_ipcw` function.

    

        integrated_brier_score_ipcw_sklearn_scorer

        scikit-learn scorer wrapper around `integrated_brier_score_ipcw` function.

    



## Strata Preprocessing



The `survivalpredict.strata_preprocessing` module allows for the creation of strata to be used various estimators.



    

        Class

        Description

    

        StrataBuilderDiscretizer

        Builds strata keys from numeric data. Allows various splitting strategies. 

    

        StrataBuilderEncoder

        Builds strata keys from categorical data.

    

        StrataColumnTransformer

        Allows various StrataBuilders to be stacked and simultaneously to be run on different columns to build the strata. Modeled after scikit-learn's ColumnTransformer. 

    

        make_strata_column_transformer

        Generates the StrataColumnTransformer class without having to name each transformation directly, like scikit-learn's make_column_transformer. 

    



## Pipeline



Due to various reasons, survivalpredict intentionally breaks with scikit-learn's api in several ways. The `survivalpredict.pipeline` module allows for creating wrappers around various survivalpredict classes, in order for survivalpredict to interpolate with the greater scikit-learn ecosystem (i.e., for feature selection or hyperparameter tuning); in addition of the various utility of a conventional scikit-learn's pipeline. 



    

        Class

        Description

    

    

        build_sklearn_pipeline_target

        Builds a singular target array from the times and events arrays. Used as the 'y'/observed for scikit-learn ecosystem.

    

    

        SklearnSurvivalPipeline

        Stacks various sklearn transformers and survivalpredict strata_builders and estimators into single class. It assumes the output of the `build_sklearn_pipeline_target` function as the 'y'/observed.

    

    

        make_sklearn_survival_pipeline

        Generates a SklearnSurvivalPipeline class without having to directly name all the steps.

    



## Validation



survivalpredict comes with some native model validation capability, within `survivalpredict.validation`.



    

        Class

        Description

    

        sur_cross_val_score

        survivalpredict's equivalent to scikit-learn's cross_val_score.

    

        sur_cross_validate

        survivalpredict's equivalent to scikit-learn's cross_validate.

    



## Model Selection



Scikit-learn's model_selection is also mimicked within `survivalpredict.model_selection`



    

        Class

        Description

    

        Sur_GridSearchCV

        survivalpredict's equivalent to scikit-learn's GridSearchCV

    

        Sur_RandomizedSearchCV

        survivalpredict's equivalent to scikit-learn's RandomizedSearchCV