Introduction and aims

Infoveillance approaches (i.e. surveillance methods using online content) that leverage big data can provide new insights about infectious disease outbreaks and substance use disorder topics. We assessed social media messages about HIV, opioid use and injection drug use in order to understand how unstructured data can prepare public health practitioners for response to future outbreaks.

Design and methods

We conducted an retrospective analysis of Twitter messages during the 2015 HIV Indiana outbreak using machine learning, statistical and geospatial analysis to examine the transition between opioid prescription drug abuse to heroin injection use and finally HIV transmission risk, and to test possible associations with disease burden and demographic variables in Indiana and Marion County. Tweets from October 2014 to June 2015 were compared to disease burden at the county level for Indiana, and classification of census blocks by presence of relevant messages was done at the census block level for Marion County. Marion County was used as it exhibited the highest total count of Tweets.

Results

257 messages about substance abuse and HIV were significantly related to HIV rates (P < 0.001) and opioid-related hospitalisations (P = 0.037). Using 157 characteristics from the American Community Survey, a linear classifier was computed with an appreciable correlation (r = 0.49) to risk-related social media messages from Marion County.

Discussion and conclusions

Communities appear to communicate online in response to disease burden. Classification produced an accurate equation to model census block risk based on census data, allowing for high-dimensional estimation of risk for blocks with sparse populations.

Introduction

From late 2014 through 2015, Scott County, Indiana faced an HIV outbreak triggered by opioid abuse and transition to injection drug use. Investigating the origins, risk factors, and responses related to this outbreak is critical to inform future surveillance, interventions, and policymaking. In response, this retrospective infoveillance study identifies and characterizes user-generated messages related to opioid abuse, heroin injection drug use, and HIV status using natural language processing (NLP) among Twitter users in Indiana during the period of this HIV outbreak.

Materials and methods

Our study consisted of two phases: data collection and processing, and data analysis. We collected Indiana geolocated tweets from the public Twitter API using Amazon Web Services EC2 instances filtered for geocoded messages in the immediate pre and post period of the outbreak. In the data analysis phase we applied an unsupervised machine learning approach using NLP called the Biterm Topic Model (BTM) to identify tweets related to opioid, heroin/injection, and HIV behavior and then examined these messages for HIV risk-related topics that could be associated with the outbreak.

Results

More than 10 million geocoded tweets occurring in Indiana during the immediate pre and post period of the outbreak were collected for analysis. Using BTM, we identified 1350 tweets thought to be relevant to the outbreak and then confirmed 358 tweets using human annotation. The most prevalent themes identified were tweets related to self-reported abuse of illicit and prescription drugs, opioid use disorder, self-reported HIV status, and public sentiment regarding the outbreak. Geospatial analysis found that these messages clustered in population dense areas outside of the outbreak, including Indianapolis and neighboring Clark County.

Discussion

This infoveillance study characterized the social media conversations of communities in Indiana in the pre and post period of the 2015 HIV outbreak. Behavioral themes detected reflect discussion about risk factors related to HIV transmission stemming from opioid and heroin abuse for priority populations, and also help identify community attitudes that could have motivated or detracted the use of HIV prevention methods, along with helping identify factors that can impede access to prevention services.

Conclusions

Infoveillance approaches, such as the analysis conducted in this study, represent a possibly strategy to detect "signal" of the emergence of risk factors associated with an outbreak though may be limited in their scope and generalizability. Our results, in conjunction with other forms of public health surveillance, can leverage the growing ubiquity of social media platforms to better detect opioid-related HIV risk knowledge, attitudes and behavior, as well as inform future prevention efforts.