care-survival

Upgrading survival models with CARE.

This repository contains a Python implementation of the methodology proposed by Underwood, Reeve, Feng, Lambert, Mukherjee and Samworth, 2025.

Building the Python package

This project uses uv to manage Python dependencies. To build the Python package, run uv build. To update the lockfile, use uv lock.

Running the simulation scripts

For running scripts, just and parallel should be installed. To execute all of the simulations and generate the plots, run just. The recipes can be found in the justfile.

The scripts to run the SCORE2 data analysis are also available in the bin directory, but require access to UK Biobank data.

Example usage

import numpy as np
import care_survival

# generate some sample data
n = 60
d = 2
X = np.random.random((n, d))
T = X.sum(axis=1) + 1 + np.random.random(n)
I = (np.random.random(n) > 0.8) * 1
f = (-X.sum(axis=1) / 2 + np.random.random(n)).reshape(-1, 1)

# split data into training and validation sets
n_train = int(np.ceil(n / 2))
X_train = X[0:n_train]
T_train = T[0:n_train]
I_train = I[0:n_train]
f_train = f[0:n_train]
X_valid = X[n_train:n]
T_valid = T[n_train:n]
I_valid = I[n_train:n]
f_valid = f[n_train:n]

# define a kernel and use the feature map optimisation method
a = 1
p = 2
kernel = care_survival.PolynomialKernel(a, p)
method = "feature_map"

# set up the kernel tuning parameters
n_gammas = 20
gamma_min = 1e-6
gamma_max = 1e1

# set up the simplex tuning parameters
simplex_resolution = 0.05

# compute concordance score on all data
with_concordance = ["train", "valid"]

# fit CARE
care = care_survival.care(
    X_train,
    T_train,
    I_train,
    f_train,
    X_valid,
    T_valid,
    I_valid,
    f_valid,
    kernel,
    method,
    n_gammas,
    gamma_min,
    gamma_max,
    simplex_resolution,
    with_concordance,
)

# view diagnostics
best = care.best["aggregated"]["ln"]["valid"]
print("best theta value:", best.theta)
print("best gamma value:", best.gamma)
print("concordance index:", best.score["concordance"]["valid"])

Publishing to PyPI

First build the package with

uv build

To publish to TestPyPI with an API token, run

uv publish --index testpypi --token <token>

To publish to PyPI with an API token, run

uv publish --token <token>