Dynamic Survival Transformers for Causal Inference with Electronic Health Records
Accepted as a spotlight presentation to the NeurIPS 2022 Workshop on Learning from Time Series for Health
A deep learning model built in Pytorch Lightning. Dependencies are managed through Poetry, configurations through Hydra.
Our semi-synthetic data derives from the MIMIC-III Clinical Database (https://physionet.org/content/mimiciii-demo/1.4/). We use the MIMIC-Extract pipeline to preprocess the data (https://github.com/MLforHealth/MIMIC_Extract).
To generate the semi-synthetic dataset, you should have the MIMIC-Extract file all_hourly_data.h5 saved in a directory called data/. Then run
python run.py preprocess.do=True
To launch a multirun of DynST sweeping over several hyperparameters, you can enter the following command:
python run -m model.d_model=32,64 model.alpha=0,0.1,0.2