⭐SurvTRACE: Transformers for Survival Analysis with Competing Events

This repo provides the implementation of SurvTRACE for survival analysis. It is easy to use with only the following codes:

from survtrace.dataset import load_data
from survtrace.model import SurvTraceSingle
from survtrace import Evaluator
from survtrace import Trainer
from survtrace import STConfig

# use METABRIC dataset
STConfig['data'] = 'metabric'
df, df_train, df_y_train, df_test, df_y_test, df_val, df_y_val = load_data(STConfig)

# initialize model
model = SurvTraceSingle(STConfig)

# execute training
trainer = Trainer(model)
trainer.fit((df_train, df_y_train), (df_val, df_y_val))

# evaluating
evaluator = Evaluator(df, df_train.index)
evaluator.eval(model, (df_test, df_y_test))

print("done!")

🔥See the demo

Please refer to experiment_metabric.ipynb and experiment_support.ipynb !

🔥How to config the environment

Use our pre-saved conda environment!

conda env create --name survtrace --file=survtrace.yml
conda activate survtrace

or try to install from the requirement.txt

pip3 install -r requirements.txt

🔥How to get SEER data

1. Go to https://seer.cancer.gov/data/ to ask for data request from SEER following the guide there.

2. After complete the step one, we should have the following seerstat software for data access. Open it and sign in with the username and password sent by seer.

3. Use seerstat to open the ./data/seer.sl file, we shall see the following.

Click on the 'excute' icon to request from the seer database. We will obtain a csv file.

4. move the csv file to ./data/seer_raw.csv, then run the python script process_seer.py, as

   python process_seer.py

   we will obtain the processed seer data named seer_processed.csv.

📝Functions

• single event survival analysis
• competing events survival analysis
• multi-task learning
• automatic hyperparameter grid-search

😄If you find this result interesting, please consider to cite this paper:

@article{wang2021survtrace,
      title={Surv{TRACE}: Transformers for Survival Analysis with Competing Events}, 
      author={Zifeng Wang and Jimeng Sun},
      year={2021},
      eprint={2110.00855},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}