survivalpredict documentation

Survivalpredict is a Python package for survival analysis with a statistical learning lens.

Survival analysis is a branch of statistics that frames its modeling on the assumption that all observations are being pushed towards failure as time moves forward. Each observation is assumed to have either failed at a given point in time or has yet to fail at the last known interval in time, and is hence censored. Survival curves are the estimated probability of failure at different intervals of time. Survival analysis was originally developed in the context of clinical trials but has applications in customer churn, lead conversion, financial defaults, operational failures, and other areas.

SurvivalPredict has a singular focus on survival curves and being able to tune models using statistical learning methodology. The ability to easily perform cross-validation, hyperparameter tuning, and feature selection on survival models is core to SurvivalPredict.

Installation

You can easily install Survivalpredict via the pip command.

pip install survivalpredict

Core features of Survivalpredict are:

  • Vectorised code base, with some necessary parts written in numba. Resulting in performat code.

  • Scikit-learn inspired api. All estimators return survival curves on predict. Survivalpredict also has tooling to interlope with Scikit-learn directly.

  • Ability to easily cross-validate estimators with Brier scores. Making tuning directly influenced by the performance of survival curves.

  • First-class support for left-censorship and model stratification.

Survivalpredict makes some assumptions:

  • The times array, the last known observed time of an individual before the event or censorship, is to be encoded as an integer. It is assumed that time starts with 1, and each interval of time is equally important. It is advised to engineer time to max out at a few thousand; large values in the time array can trigger expensive computation on several estimators.

  • The events array should be of boolean type, True if the individual experiences an event (e.g., charged, death, conversion, etc.), and False otherwise.

  • If the data is left-censored, in cases of time-varying effects or recurrent events, the smallest value of the times_start array should be 0, and the times_start array should be smaller than the times array.

  • When calling predict on an estimator, columns of the output will correspond to all times till max time, starting at time 1.

Walkthroughs:

API: