This is a re-implementation of the GRU-D model with Python3 + Keras2 + Tensorflow.

Zhengping Che, Sanjay Purushotham, Kyunghyun Cho, David Sontag, and Yan Liu. "Recurrent Neural Networks for Multivariate Time Series with Missing Values", Scientific Reports (SREP), 8(1):6085, 2018.

An earlier version is available on arXiv (arXiv preprint arXiv:1606.01865).

• Python3.6
• For dependencies, see requirements.txt
• Python3 packages
  • h5py==2.10.0
  • Keras==2.2.0
  • numpy==1.14.0
  • scikit-learn==0.19.1
  • tensorboard==1.7.0
  • tensorflow==1.7.0

• We use [WD] to represent the working directory (i.e., working path, in which all related data/log/model/result/evaluation files are stored).
• We assume the data are saved in the folder [WD]/data/${dataset_name}. Please see data_handler.py for more information.
  • In [WD]/data/${dataset_name}/data.npz, there are input, masking, timestamp, label_${label_name}. Each of them is of the shape (n_samples, ...)
  • In [WD]/data/${dataset_name}/fold.npz, there are fold_${label_name}, mean_${label_name}, std_${label_name}. Each of them is of the shape (k_fold, 3, ...), for train/validation/test sets in k-fold cross validation.
• Our GRU models take (x, masking, timestamp) as the inputs. Please refer to models.py and nn_utils/grud_layers.py.
• Run.ipynb serves as an example script for model training and evaluation.
• To demonstrate the data file format, a (fake) sample data is provided at [WD]/data/sample. The sample data is generated by Generate-sample-data.ipynb.

The following steps will help you to conduct experiments for mortality predictions on the MIMIC-III dataset with the time series within the first 48 hours after the patient's admission. We rely on this (older version of) Benchmarking codebase to extract and preprocess the time series data from the MIMIC-III dataset and provide necessary scripts to convert the data for our GRU-D models.

1. Make sure you have the (older version of) benchmarking codebase and set up the database connection in the Requirements section.
2. Follow steps 1-6 in the Select admissions and all features section. I.e., execute
   • All 11 scripts named as [#]_***.ipynb for 0 <= [#] <= 9
   • Some scripts (e.g., 8_processing.ipynb) may take hours or a couple of days to complete.
3. Follow steps 1,2,4 in the Generate 17 processed features, 17 raw features and 140 raw features section. I.e., execute
   • run_necessary_sqls.ipynb
   • 10_get_17-features-processed.ipynb
   • 10_get_99plus-features-raw.ipynb
4. Follow step 3 in the Generate time series section with X=48(hours). I.e., execute
   • 11_get_time_series_sample_99plus-features-raw_48hrs.ipynb
5. Now you should have extracted necessary data files from the benchmarking codebase. Please set the directories in Prepare-MIMIC-III-data.ipynb and execute it to prepare the data for GRU-D.
6. Execute Run.ipynb and check the results!