Neural Record Captioning (NRC)

This repository contains code from the paper Natural Language Generation for Electronic Health Records.

what's included

1. Keras code for the NRC model.
2. Training and testing scripts for the model.
3. Example scripts for preprocessing EHR data to be used in the model.

getting started

1. Install the necessary Python modules (list below)
2. Use preprocessing/sparisfy.py to convert the discrete variables in your EHRs to sparse format
3. Use preprocessing/words_to_integers.py to convert your free text field to integers
4. Train the autoencoder on the sparse records with ae_training.py
5. Train the NRC model with caption_training.py
6. Generate text with caption_testing.py

required software

1. Python 3.x
2. Keras with the TensorFlow backend
3. Pandas, NumPy, h5py, and scikit-learn

hot tips

The default hyperparameters worked well for the data used in our paper, but they might not for yours, so feel free to experiment! Also, we recommend a GPU for training the captioning model. We used a single NVIDIA Titan X for our experiments, and training with ~2 million records took around 6 hours.