LAVA

This repository contains the source code for the paper: An, Sungtae, et al. "Longitudinal Adversarial Attack on Electronic Health Records Data." Proceedings of the 2019 World Wide Web Conference, WWW'19, ACM, 2019. Paper

These codes were tested with Python 3.6 and PyTorch 0.3.1 (built with CUDA 9.1). You may use the included conda environment file environment.yml.

0. Preprocess Data

0.0. Prepare MIMIC-III Dataset

While we used a different private EHR data for the experiments reported in the paper, here we use MIMIC-III as an example dataset since it is publicly available upon request. You should obtain access and download the CSV database first. Please refer to their web page.

Once the MIMIC-III CSV files are ready, you can use the following command to process the data into the required format.

python 0_0_process_mimic.py <ADMISSIONS FILE PATH> <DIAGNOSES FILE PATH> <PATIENTS FILE PATH> <OUTPUT PATH WITH A FILE NAME PREFIX>

For example,

python 0_0_process_mimic.py /mimic3/ADMISSIONS.csv /mimic3/DIAGNOSES_ICD.csv /mimic3/PATIENTS.csv ./Processed/outout

Among the multiple output files, output.seqs and output.morts are mostly used in this project.

output.seqs

It is a pickled Python object, which is a List of List of List of Int. First of all, all clinical (event) features such as diagnoses, procedures, and medications are encoded as integer values from 0 to N-1 where N is the number of clinical features. Then, each visit (or multiple events on the same date) made by a patient is grouped and represented as a List of Int. Similarly, a single patient's clinical history (prior to a certain index date determined) is represented as a List of visits (which is a List of Int above) ordered by time/date. Finally, the entire cohort (all patients) data are collected into a List. Therefore, the final form is a List of List of List of Int.

output.mort

It is a pickled Python List that contains 0 or 1 values as a label (mortality) for each patient. It has the same order of patients as output.seqs.

You may use any other datasets with this LAVA code once they are prepared in the same format described above.

0.1. Split the dataset

We split the prepared dataset into three folds: training, validation, and test sets. You may use your own script or the following code:

python 0_1_split_dataset.py <PICKLED SEQUENCE FILE> <PICKLED LABEL FILE> <OUTPUT DIR>

For example,

python 0_1_split_dataset.py Processed/output.seqs Processed/output.morts Processed/

It will generate <FOLD NAME>.seqs and <FOLD NAME>.labels for this LAVA project, and additional <FOLD NAME>_<NORMALIZED>.data_csr files which can be used as non-sequence input data, with MLP for example.

1. Train a source (RETAIN) model

Now that we have a pre-processed dataset, we train a source model to craft adversarial exmaples of longitudinal EHR by using LAVA. In the paper and so here, we use RETAIN as our source model to utilize dual-level attention mechanism. Please refer to LAVA paper and the RETAIN paper for more details.

You can train a source RETAIN model using the following code:

python 1_train_RETAIN.py <DATASET DIR>

For example,

python 1_train_RETAIN.py Processed/

Please note that you must set the number of available features in advance in the code or by using the argument --num_features. It applies same for all the codes below.

Please refer to the code or python 1_train_RETAIN --help to check available options for the model architecture or the training procedure.

2. Craft LAVA samples

Once we have a trained source model, we can craft LAVA adversarial examples:

python 2_craft_lava.py <CLEAN DATASET DIR> <TRAINED MODEL FILE>

For example,

python 2_craft_lava.py Processed/ Save/best_model.pth

Please refer to the code or python 2_craft_lava.py --help to check available options for the crafting procedure.

3. Test a target model

We can evaluate a trained model with either the clean test set or the crafted adversarial test set. While we reported the results for a variety of different target models in the paper, here we use the same source RETAIN model as the target model (i.e., white-box attack) for simplicity.

You can test/evaluate a trained RETAIN model using the following code:

python 3_test_RETAIN.py <PICKLED TEST SEQUENCE FILE> <PICKLED TEST LABEL FILE> <TRAINED MODEL FILE>

For example,

Clean test set

python 3_test_RETAIN.py Processed/test.seqs Processed/test.labels Save/best_model.pth

Crafted adversarial test set (note that we use the same label file)

python 3_test_RETAIN.py Crafted/adv_20.seqs Processed/test.labels Save/best_model.pth

Please contact Sungtae An [email protected] if you have any questions about either the paper or the code.