MICCAI 2025 CHIMERA Challenge Task 1 2nd Place Solution

Challenge Overview

This repository contains our 2nd place solution for the MICCAI 2025 CHIMERA Challenge Task 1, which focuses on predicting time to biochemical recurrence (BCR) or follow-up in prostate cancer patients after prostatectomy using the Concordance Index (C-index) as the evaluation metric.

Solution Overview

This directory contains a complete pipeline for training and evaluating survival models on clinical data. The pipeline consists of three main steps:

1. Fold Creation: Create stratified K-fold splits and save as CSV
2. Model Training: Train 6 models per fold and save top-3 performers
3. Ensemble Evaluation: Load all saved models and evaluate ensemble performance

File Structure

chimera/
├── create_fold.py          # Step 1: Create folds with stratification
├── unimodel_main.py        # Step 2: Train models and save top-3 per fold
├── ensemble_main.py        # Step 3: Evaluate ensemble of all saved models
├── models/                 # Model implementations
│   ├── mlp_model.py
│   ├── real_mlp_model.py
│   ├── tabm_model.py
│   └── traditional_model.py
└── README.md 

Required Python Packages

Core dependencies:

pip install numpy pandas scikit-learn scikit-survival torch xgbse tabm rtdl-num-embeddings

Usage

Step 1: Create Folds

Creates CSV files with fold assignments using uniform stratification.

cd chimera
python create_fold.py

Output:

• clinical_data_with_folds_age.csv - Stratified by age at prostatectomy
• clinical_data_with_folds_followup.csv - Stratified by follow-up time

Step 2: Train Models

Trains 6 models on each fold and saves the top-3 performers.

python unimodel_main.py

Models Trained:

1. MLP
2. RealMLP
3. TabM
4. CoxPH
5. RandomSurvivalForest
6. SurvivalXGBoost

Output:

• saved_models/followup_fold_XX/ - Top-3 models per fold
• results/unimodel_evaluation_results.json - Training results summary

Total Models Saved:

• 10 folds × 3 models = 30 models (for single split)
• With age split: 10 folds × 3 models × 2 splits = 60 models

Step 3: Evaluate Ensemble

Loads all saved models and evaluates ensemble performance.

python ensemble_main.py

Ensemble Method:

• Simple Average: Equal-weighted average of all model predictions

Output:

• results/ensemble_evaluation_results.json - Ensemble evaluation results

Results

Model	CV Test C-index	Public Leaderboard
RandomSurvivalForest	0.965 ± 0.042	-
TabM	0.939 ± 0.055	0.7273
RealMLP	0.879 ± 0.122	-
MLP	0.820 ± 0.123	0.7521 (2nd)
CoxPH	0.761 ± 0.279	-
XGBoost	0.616 ± 0.359	-
Ensemble (60 models)	0.963 ± 0.042	0.7438

Note: CV = Cross-Validation (10-fold on followup split)

Citations

If you use this repository in your research, please cite:

@software{chimera2025,
  author = {Your Name},
  title = {CHIMERA: Clinical Survival Model Pipeline},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/yws0322/chimera}
}

This project uses implementations from the following works:

@software{realmlp_standalone,
  author = {Holzmueller, David},
  title = {RealMLP-TD-S Standalone Implementation},
  year = {2024},
  publisher = {GitHub},
  url = {https://github.com/dholzmueller/realmlp-td-s_standalone}
}

@inproceedings{tabm2025,
  title={TabM: Advancing Tabular Deep Learning With Parameter-Efficient Ensembling},
  author={Gorishniy, Yury and Rubachev, Ivan and Kartashev, Nikolay and Shlenskii, Daniil and Kodryan, Akim and Babenko, Artem},
  booktitle={International Conference on Learning Representations},
  year={2025},
  url={https://arxiv.org/abs/2410.24210}
}

License

This project builds upon open-source implementations:

• RealMLP-TD-S: MIT License
• TabM: Apache-2.0 License

Please cite the original works if you use these models in your research.