Bidirectional LSTM for Labeling Adverse Drug Reactions in Twitter Posts

This repository contains code and data from the paper:

Anne Cocos, Alexander G. Fiks, and Aaron J. Masino. Deep Learning for Pharmacovigilance: Recurrent Neural Network Architectures for Labeling Adverse Drug Reactions in Twitter Posts.

Please cite the following if you use these resources in your own work:

@article{doi:10.1093/jamia/ocw180,
author = {Cocos, Anne and Fiks, Alexander G and Masino, Aaron J},
title = {Deep learning for pharmacovigilance: recurrent neural network architectures for labeling adverse drug reactions in Twitter posts},
journal = {Journal of the American Medical Informatics Association},
volume = {24},
number = {4},
pages = {813},
year = {2017},
doi = {10.1093/jamia/ocw180},
URL = { + http://dx.doi.org/10.1093/jamia/ocw180},
eprint = {/oup/backfile/content_public/journal/jamia/24/4/10.1093_jamia_ocw180/1/ocw180.pdf}
}

Note that the code in this release has been updated from the original version to correspond to new versions of Keras and Theano. As a result, the model is missing a dropout layer which will impact the results. For the original release, see v1.0.

Quick Start

1. Ensure your Python 2.7 environment has all the requirements listed in ./requirements.txt, in particular Theano (v1.0.2) and [Keras]() (v2.2.2). We used an Enthought Canopy Python distribution to cover the rest. Note this code is not yet updated to work with Python 3.

2. Download the original Twitter ADR Dataset v1.0 and the word2vec_twitter_model from the respective dataset authors by running this script:

   ./download_orig_datasets.sh

3. It is against Twitter's Terms of Service to publish the text of tweets, so we (and the original Twitter ADR Dataset v1.0 authors) have provided tweet ID's instead, and a script to download the text of each annotated tweet directly from Twitter. Run it with the following command:

   python ./download_tweets.py

   Note that you will see numerous errors of the type Error downloading tweet 327323676317974528 (over 780 of these, as of 07-30-2018). This is because you cannot retrieve tweets by users with deleted or private accounts. It is a violation of the Twitter terms of use to distribute full-text tweets for deleted or private accounts.

4. Prepare the dataset (by running prep.py), and train, validate, and test the model (with adr_label.py). We've provided an example script that will allow you to create a 90/10 train/validation split from within the combined ASU and CHOP training sets, train the BLSTM, and test on the combined ASU and CHOP test sets:

   ./prep_train_test.sh

TL;DR

Contents

2. Files & Directories
3. Data
4. Code

Files & Directories

Data

With this model we are releasing the CHOP ADHD ADR Twitter Supplement, which we used to supplement the ASU Twitter ADR Dataset v1.0 for training data. See the included ReadMe.txt for details. As included, the data is divided into Twitter ID and annotation files, and you must run download_tweets.py to get the original Tweet text. Running that script will produce train and test data as tab-separated-value files in the directory ./data/seq_labeling/raw with the headings (id,start,end,semantic_type,span,reldrug,tgtdrug,text). The fields start and end denote the indices of the first and last word in the tweet that make up the annotated span (where words are separated by spaces). The semantic_type denotes whether the span is an ADR, Indication, Beneficial reaction, or none of these (Negative). The span gives the span text, the reldrug is the name of the drug related to the span, the tgtdrug is the drug which was searched for to obtain the tweet, and text gives the tweet text.

In general, you can input any sequence labeling to this model that follows the same format. If you need to update the semantic_type headings, you can do so in settings.py.

Before running the model, you should run prep.py in order to build the feature indices and pickle your data for input to the model.