SDPM: Survival Diffusion Probabilistic Model for Continuous-Time Survival Analysis

This repository contains the code and preprocessed data used for the experiments in the article:

SDPM: Survival Diffusion Probabilistic Model for Continuous-Time Survival Analysis

Preprint paper

SDPM is a diffusion-based model for continuous-time survival analysis. The repository includes the SDPM model implementation, baseline wrappers, experiment runners, preprocessed benchmark datasets, and notebooks for additional analyses and table generation.

Repository Contents

• sdpm/ - Python package with the SDPM implementation and experiment utilities.
• sdpm/sdpm.py - main SDPM model class.
• sdpm/diffusion/ - diffusion process, scheduler, and time-conditioning components.
• sdpm/baseline/ - neural-network baseline components used by SDPM.
• sdpm/experiments/ - comparison experiment code and baseline model wrappers.
• sdpm/experiments/results/final_results.csv - aggregated final experiment results.
• data/ - preprocessed survival datasets used in the experiments.
• comparison_script.sh - convenience script for running the main comparison experiments.
• ablation_K.ipynb, ablation_r.ipynb, ablation_sf.ipynb, log_delta_ablation.ipynb - notebooks with ablation and additional analysis experiments.
• final_tables.ipynb - notebook for aggregating results and preparing final tables.
• environment.yml - Conda environment specification.

Installation

Create and activate the Conda environment:

conda env create -f environment.yml
conda activate sdpm

Run commands from the repository root so that the local sdpm package and data/ directory are available.

Data

The repository includes preprocessed .npz versions of the ten datasets used in the experiments:

• flchain
• ovarian
• pbc
• retinopathy
• rotterdam
• seer
• support
• tcga_gbm
• vlbw
• whas500

Each dataset is loaded by name from data/<dataset>.npz.

Running Comparison Experiments

The main comparison experiments can be run with:

bash comparison_script.sh <dataset> [device]

For example:

bash comparison_script.sh vlbw cuda:0

If no device is provided, the script uses cuda:0 for SDPM. Classical and non-SDPM baselines are run on CPU.

The script runs the following models:

• sdpm_mlp - proposed SDPM model
• rsf - Random Survival Forest
• deepsurv - DeepSurv
• deephit - DeepHit
• gbm_wb - XGBSEStackedWeibull
• gbm_km - XGBSEKaplanNeighbors

Results are appended to:

sdpm/experiments/results/results.csv

The comparison runner can also be invoked directly:

python -m sdpm.experiments.comparison \
  -data vlbw \
  -model sdpm_mlp \
  -device cuda:0 \
  -trials 100 \
  -repeats 10 \
  -threads 16 \
  -table_filename sdpm/experiments/results/results.csv

Available dataset names are listed above. Available model names are sdpm_mlp, rsf, deepsurv, deephit, gbm_wb, and gbm_km.

Experimental Setup

The comparison experiments evaluate SDPM against five survival-analysis baselines on ten real-world datasets. The evaluation uses repeated 4-fold cross-validation with Optuna hyperparameter optimization inside each fold.

The reported metrics are:

• Harrell's C-index
• Integrated time-dependent AUC
• Integrated Brier score (IBS)

The main experiment results used for the article are stored in:

sdpm/experiments/results/final_results.csv

Ablation Experiments and Tables

Additional experimental analyses are provided as Jupyter notebooks:

• ablation_K.ipynb - influence of the number of generated samples K.
• ablation_r.ipynb - influence of the number of diffusion steps r.
• ablation_sf.ipynb - survival-function related ablation analysis.
• log_delta_ablation.ipynb - additional ablation for the event label and time representation.
• final_tables.ipynb - result aggregation and final table preparation.

Citation

@unpublished{kirpichenko2026sdpm,
      title={SDPM: Survival Diffusion Probabilistic Model for Continuous-Time Survival Analysis}, 
      author={Stanislav R. Kirpichenko and Andrei V. Konstantinov and Lev V. Utkin},
      year={2026},
      eprint={2605.22776},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.22776}, 
}