Gradient-based Explanations for Deep Learning Survival Models

This repository contains the code and material to reproduce the results of the manuscript "Gradient-based Explanations for Deep Learning Survival Models" accepted for publication in the proceedings of the Forty-Second International Conference on Machine Learning (ICML) 2025.

The reproduction material is based on the accompanying R package survinng available on github/bips-hb.

📁 Repository Structure

• setup.R: R environment setup script that installs required packages, the necessary conda environment Survinng_paper, and sets global options. It also installs the survinng package from the mentioned GitHub repository.
• Sim_time_dependent.Rmd: Simulation for time-dependent features. The results used in the paper are stored in the notebook Sim_time_dependent.html and figures are saved in the figures_paper/ directory.
• Sim_time_independent.Rmd: Simulation for time-independent features. The results used in the paper are stored in the notebook Sim_time_independent.html and figures are saved in the figures_paper/ directory.
• Sim_GradSHAP: Simulation for comparing GradSHAP(t) and SurvSHAP(t) on time-independent features regarding runtime, local accuarcy and feature ranking.
• real_data/: Scripts for reproducing the results on the real data example.
• figures_paper/: Directory for storing the figures used in the paper.

🚀 Reproducing the Results

• To reproduce the results, from Section 5.1.1 TIME-INDEPENDENT EFFECTS, run the RMarkdown file Sim_time_independent.Rmd and the results will be stored Sim_time_independent.html and the figures in the figures_paper/ directory.

• To reproduce the results, from Section 5.1.2 TIME-DEPENDENT EFFECTS, run the RMarkdown file Sim_time_dependent.Rmd and the results will be stored Sim_time_dependent.html and the figures in the figures_paper/ directory.

• To reproduce the results, from Section 5.2 GradSHAP(t) vs. SurvSHAP(t), run corresponding scripts in the directory Sim_GradSHAP/, i.e.,

  • sim_locacc.R: for the local accuracy comparison
  • sim_runtime.R: for the runtime comparison
  • sim_global_imp.R: for the global feature ranking comparison. The figures will be stored in the figures_paper/ directory. Note: This simulation is computationally expensive and conducts a simulation study using batchtools.

• To reproduce the results, from the Section 5.2 Practical Feasibility and Section 5.3 Example on Real Multi-modal Medical Data, we refer to the README file in the folder real_data/.

📚 Requirements

The script setup.R tries to install the necessary packages and the conda environment Survinng_paper (see file env_survinng_paper.yml). It installs the following R packages:

Survival packages

• simsurv
• survival
• survminer
• SurvMetrics
• survinng (from github/bips-hb)
• survex
• survivalmodels
• torch (necessary for the survinng package)

Plotting and other useful packages

• ggplot2
• cowplot
• viridis
• dplyr
• tidyr
• reticulate
• callr
• here
• data.table
• batchtools