Multistate competing risk models in Python

Read the Docs
For details, read the JOSS paper

Hagai Rossman, Ayya Keshet, Malka Gorfine 2022

PyMSM is a Python package for fitting competing risks and multistate models, with a simple API which allows user-defined model, predictions at a single or population sample level, statistical summaries and figures.

Features include:

• Fit a Competing risks Multistate model based on survival analysis (time-to-event) models.
• Deals with right censoring, competing events, recurrent events, left truncation, and time-dependent covariates.
• Run Monte-carlo simulations for paths emitted by the trained model and extract various summary statistics and plots.
• Load or configure a pre-defined model and run path simulations.
• Modularity and compatibility for different time-to-event models such as Survival Forests and other custom models.

Installation

pip install pymsm

Requires Python >=3.8.

Alternatively if you want to work with the latest development version, you can also directly install it from GitHub. To do that, you will need to

1. Clone the repository to REPO_FOLDER (choose your own location)
2. Got to the location of the repository cd $REPO_FOLDER (Note! Not the pymsm folder, but the one above)
3. Run pip install -e pymsm. This imports the package for your Python interpreter

Quick example

# Load data (See Rotterdam example for full details)
from pymsm.datasets import prep_rotterdam
dataset, states_labels = prep_rotterdam()

# Define terminal states
terminal_states = [3]

#Init MultistateModel
from pymsm.multi_state_competing_risks_model import MultiStateModel
multi_state_model = MultiStateModel(dataset,terminal_states)

# Fit model to data
multi_state_model.fit()

# Run Monte-Carlo simulation and sample paths
mcs = multi_state_model.run_monte_carlo_simulation(
              sample_covariates = dataset[0].covariates.values,
              origin_state = 1,
              current_time = 0,
              max_transitions = 2,
              n_random_samples = 10,
              print_paths=True)

    stateDiagram-v2
    s1 : (1) Primary surgery
    s2 : (2) Disease recurrence
    s3 : (3) Death
    s1 --> s2: 1518 
    s1 --> s3: 195 
    s2 --> s3: 1077 

Loading

Full examples

1. Rotterdam Illness-death example
2. EBMT multistate example
3. COVID hospitalizations multistate example

Background and Motivation

Multi-state data are common, and could be used to describe trajectories in diverse health applications; such as describing a patient's progression through disease stages or a patient’s path through different hospitalization states. When faced with such data, a researcher or clinician might seek to characterize the possible transitions between states, their occurrence probabilities, or to predict the trajectory of future patients - all conditioned on various baseline and time-varying individual covariates. By fitting a multi-state model, we can learn the hazard for each specific transition, which would later be used to predict future paths. Predicting paths could be used at a single patient level, for example predict how long until a cancer patient will be relapse-free given his current health status, or at what probability will a patient end a trajectory at any of the possible states; and at the population level, for example predicting how many patients which arrive at the emergency-room will need to be admitted, given their covariates.

Capabilities

PyMSM is a Python package for fitting multi-state models, with a simple API which allows user-defined models, predictions at a single or population sample level, and statistical summaries and figures. Features of this software include:

• Fitting a Competing risks Multistate model based on various types of survival analysis (time-to-event) such as Cox proportional hazards models or machine learning models, while taking into account right censoring, competing events, recurrent events, left truncation, and time-dependent covariates.
• Running Monte-carlo simulations (in parallel computation) for paths emitted by the trained model and extracting various summary statistics and plots.
• Loading or configuring a pre-defined model and generating simulated data in terms of random paths using model parameters, which could be highly useful as a research tool.
• Modularity and compatibility for different time-to-event models such as Survival Forests and other custom ML models provided by the user.
  The package is designed to allow modular usage by both experienced researchers and non-expert users. In addition to fitting a multi-state model for a given data - PyMSM allows the user to simulate trajectories, thus creating a multi-state data-set, from a predefined model. This could be a valuable research tool - both for sharing sensitive simulated individual data and as a tool for any downstream task which needs individual trajectories.

Citation

If you found this library useful in academic research, please cite:

@article{Rossman2022, doi = {10.21105/joss.04566},
url = {https://doi.org/10.21105/joss.04566},
year = {2022},
author = {Hagai Rossman and Ayya Keshet and Malka Gorfine},
title = {PyMSM: Python package for Competing Risks and Multi-State models for Survival Data},
journal = {Journal of Open Source Software} }

Also consider starring the project on GitHub

This project is based on methods first introduced by the authors of Roimi et. al. 2021.
Original R code by Jonathan Somer, Asaf Ben Arie, Rom Gutman, Uri Shalit & Malka Gorfine available here. Also see Rossman & Meir et. al. 2021 for an application of this model on COVID-19 hospitalizations data.