Twitter-Based Alert System to Combat Large-Scale Vaccine Rollout Challenges

Archita Khaire

Overview

The Food and Drug Administration (FDA) estimates that adverse drug reactions (ADRs) are the 4th leading cause of death resulting in upwards of 106,000 deaths annually. The current infrastructure to combat ADRs is the Vaccine Adverse Event Reporting System (VAERS), which collects and analyzes public reports about vaccine reactions to detect any safety issues. However, the system faces significant limitations in that it is time-consuming, costly, and a manual process. Therefore, the goal of this project is to extract the sentiments and vaccine reactions and plot them on a US map by enhancing the geolocation using AI algorithms in real-time. The project was divided into the creation of a geolocation prediction model to enhance the plotting of tweets on a map and an ADR detection model to determine the ADR sensitivity of a tweet. A combination of Linear SVCs and Random Forest models was used on user-provided locations and tweets to enhance geotags. The ADR detection model was created using Gensim, which utilized a "Similarity Index" function to determine similarities between tweets and ADR vocab. After implementing both models and performing sentiment analysis on tweets, the results were plotted on a U.S. map and successfully achieved the engineering goal with 91% accuracy. The developed models can be used in conjunction with the VAERS system to gain early warnings during the vaccine rollout process. Future improvements include the generalization of the model in order to provide similar analytic information regarding different vaccines and their possible reactions.

My Story

Hello! My name is Archita and I am a junior at North Carolina School of Science and Mathematics. In the summer of my freshman year, I had the opportunity to join an outstanding team of teen researchers at a summer research program: Stanford AI4ALL. I thought I was just applying for a hands-on educational experience, however, I left with a new dream. AI4ALL was the first to develop my interest in the intersection of healthcare and artificial intelligence.

During the onset of the pandemic, I developed an interest in vaccine reactions and the notion of vaccine hesitancy. Inspired by the concept of trending topics on Twitter, I wanted to find trends in tweets in order to identify the location of vaccine reaction hotspots. Using machine learning algorithms, I could predict the location of a Twitter user who was experiencing a vaccine reaction. However, it was necessary to utilize natural language processing to filter through the mass of tweets to only detect those discussing negative side effects.

The human language is extraordinarily nuanced, and although I was able to read the text, I struggled to make my script comprehend tweets. TF-IDF scores and stopwords allowed me to simplify long tweets into shorter ones that my algorithm could understand. I was then able successfully plotted tweets on a map to display the locations of major vaccine reactions.