This is code for the paper, "Using Clinical Notes with Time Series Data for ICU Management" at EMNLP 2019 by Swaraj Khadanga, Karan Aggarwal, Shafiq R. Joty, Jaideep Srivastava.
Steps
Clone https://github.com/YerevaNN/mimic3-benchmarks and run all data generation steps to generate training data without text features.
Text Scripts
- Run extract_notes.py file under scripts folder.
- Run extract_T0.py file under scripts folder.
Configuration
- Update all paths and configuration in config.py file.
Models.
-
For IHM run ihm_model.py file under tf_trad.
Number of train_raw_names: 14681
Succeed Merging: 11579 - Model will train on this many episodes as it contains text.
Missing Merging: 3102 - These texts don't have any text for first 48 hours. -
For Decompensation, run decom_los_model.py file under tf_trad.
Text Not found for patients: 6897
Successful for patients: 22353 -
Lenght of Stay, run decom_los_model.py file under tf_trad.
Successful for episodes for training: 22353
Credits
The code is based on one of the popular MIMIC-3 benchmark repository https://github.com/YerevaNN/mimic3-benchmarks for the experiemntal setup and evaluation metrics.