Social Media Pharmacovigilance Research Project

By Samuel Kovaly and Matthew Demoy

Table of Contents

• About the Project
• What is included
• Results of data collection
• Results of Task

Description

All relevent files of a year long research project conducted with the mentorship of Professor Hong Yu. The goal of the research project was to train an AI using cutting edge natural language processing techniques to detect Adverse drug effects in tweets.

What is included in the repository

• Annoted drug mentioning Tweets
• Code BERT Model
  • Built using PyTorch
  • Trained on Google Cloud
• Analysis of gathered data using BERT model
• Papers and Powerpoint exploring findings

Motivation

Social Media platforms such as Twitter are popular mediums for people to share personal information. Some users share their current health conditions online such as drugs they are taking and the resulting effects of said drugs Autonomous collection of this kind of information for the purpose of pharmacovigilance is a subject of active research because of the value it could bring to monitoring how these drugs affect the general population

Link to Model Code

https://nbviewer.jupyter.org/urls/gist.githubusercontent.com/jyh1/2f162afedc2c6b47be63417c9acaebdf/raw/cd20035c67275b8fa9e4e7f3344d9dd2a6c26e63/bert_sentence_calssification.ipynb

Results of data collection

898 new tweets annotated(418 ADES, 480 Non-Ades)

Results of Task

F1 score of .58 with .65 precision and .53 recall.