2014.ISDS.Abstracts.Final.pdf


ISDS Annual Conference Proceedings 2014. This is an Open Access article distributed under the terms of the Creative Commons Attribution-
Noncommercial 3.0 Unported License (http://creativecommons.org/licenses/by-nc/3.0/), permitting all non-commercial use, distribution, 
and reproduction in any medium, provided the original work is properly cited.

ISDS 2014 Conference Abstracts

Supplementing Obesity-Related Surveillance with 
Persistent Health Assessment Tools

Meredith Keybl*, John Henderson, Guido Zarrella, John Gibson and Maeve Kluchnik

MITRE Corporation, Bedford, MA, USA

Objective
We developed Persistent Health Assessment Tools, PHAT, to 

equip public health policy makers with more precise tools and 
timely information for measuring the success of obesity prevention 
programs. PHAT monitors social media to supplement traditional 
surveillance by making real-time estimates based on observations of 
obesity-relevant behaviors.

Introduction
In response to the rise in obesity rates and obesity-related healthcare 

costs over the past several decades, numerous organizations have 
implemented obesity prevention programs. The current method for 
evaluating the success of these programs relies largely on annual 
surveys such as the Centers for Disease Control and Prevention’s 
Behavioral Risk Factor Surveillance System (BRFSS) which provides 
state-by-state obesity rates. As a result, public health policy makers 
lack the fine-grained evaluation data needed to make timely decisions 
about the success of their obesity prevention programs and to allocate 
resources more efficiently.

Methods
We developed a practical interface that enables policy makers 

to leverage social media. Specifically, we developed models for 
predicting obesity rates from sets of tweets and developed a dashboard 
to provide interactive navigation and time slicing. We isolate tweets of 
interest and import them into the dashboard where policy analysts can 
query and browse to observe instances where a program’s outreach 
was successful and identify ways it can be improved. These tools 
enable policy makers to study relevant demographics, adjust content 
and messaging faster, and closely measure their program’s impact.

Model: Traditionally, survey data is imputed onto different 
populations using multilevel regression with post-stratification1. 
Our approach built models predicting a survey result from non-
surveillance indicator variables - namely text in social media posts, 
location, and time - by fitting margin-matching models. Those models 
were then used to predict the missing results on other demographic 
slices from that same data. We used 2011 BRFSS data2 state-aligned 
with a sample of 7.5m tweets (500,000 users) from U.S. users to 
produce logistic regression models of obesity rate as a function of 
tweet texts.

Dashboard: To develop the dashboard, we combined data from 
social media and the BRFSS and imported them into an interactive 
display that allowed the information to be viewed in multiple formats 
such as timelines or maps. Timelines allow policy makers to track the 
response after a health program is rolled out. Color-coded maps allow 
them to see how obesity rate and program impact vary with location.

Results
We utilized a separately-produced user demographics prediction 

model to build a database aligning states with twitter users. The 
model excluded users with irrelevant or low confidence geotags, 
fake accounts, and spam campaigns. A round-robin cross-validation 
of the obesity-rate prediction model was performed: 50 models were 
built, each holding out one of the states, and the result predicted for 

the testing state was compared with the BRFSS reference. Pearson 
correlation of the resulting 50 estimates was 0.82. A large portion of 
the obesity rate was captured in those models. This initial result was 
promising and encourages us to pursue more complex models based 
on the tweets. This initial success provided us with the necessary data 
to build the PHAT dashboard.

The dashboard allows policy makers to view traditional health 
survey results (such as the BRFSS) with greater granularity, to 
query tweets while filtering through demographic variables or health 
program goals, and to drill down on individual messages.

Conclusions
We demonstrated that signals in social media can supplement 

existing survey data to provide policy makers with better tools 
to evaluate the success of their obesity prevention programs. The 
development of PHAT, especially the interactive dashboard, provides 
a more timely understanding of a program’s impact.

Future works include tailoring PHAT to evaluate a specific 
program. We also plan to expand our raw data-tagging to other 
demographic areas (e.g. veteran status) and apply PHAT to other 
health behaviors (e.g. smoking).

Keywords
Social Media; Obesity; Surveillance

References

1. Park DK, Gelman A, Bafumi J. Bayesian multilevel estimation with
poststratification: state-level estimates from national polls. Political 
Analysis. 2004;12(4):375-85.

2. Centers for Disease Control and Prevention (CDC). Behavioral
Risk Factor Surveillance System Survey Data. Atlanta, GA: U.S. 
Department of Health and Human Services, Centers for Disease 
Control and Prevention, 2011.

*Meredith Keybl

E-mail: meredith.keybl@gmail.com    

Online Journal of Public Health Informatics * ISSN 1947-2579 * http://ojphi.org * (1):e86, 201