This presentation is a follow-up on the previous version: Spreading Research and Engaging Disease Communities – One Automated Tweet at a Time. Here we share new data and argue that Algorithmic content creation can serve as a potent model for ongoing value generation to foster patient loyalty and research participant recruitment.
Engaging Patients in Research: Does algorithmically created content have a role to play in patient engagement?
1. Engaging Patients in Research
Katja Reuter1, PhD, and Anirvan Chatterjee2
Bradley Voytek3, PhD, John Daigre1
1 Southern California Clinical and Translational Science Institute (SC CTSI), University of Southern California (USC)
2 Clinical and Translational Science Institute (CTSI at UCSF), University of California, San Francisco (UCSF)
3 University of California, San Francisco (UCSF), Department of Neurology
Does algorithmically created content have a role
to play in patient engagement?
Presented at AMIA, CRI 14, Apr 10, San Francisco
3. A Shifting Landscape of Opportunity
Source: Pew Research Center surveys, 1995-2014.
60% of U.S. adults
search for health
information online.
(PEW Research, 2009)
4. It’s Time to Rethink Scientific Outreach
“Scientists are failing at
communicating science to
the public.”
(The Welcome Trust, 2001; Wilcox, 2012)
5. Learning from the Publishing Industry
“
How Algorithmically Created Content will Transform Publishing: http://www.forbes.com/sites/danwoods/2012/08/13/how-
algorithmically-created-content-will-transform-publishing/
Algorithms can provide acceleration
for steps in content creation that are
better performed by machines.
Fred Zimmerman, CEO of Nimble Books
6. Our Key Question
Does algorithmically created
content have a role to play in
patient engagement?
7. We Developed an Information System that …
Content
Discovery
Conversion
Notifications
Editing
Automated
Publishing
Automatically scans data sources
for disease-specific
content, e.g., PubMed, Clinicaltrials.gov, U
niversity News, UCSF researchers/groups on
Twitter.
Automatically creates tweets using
disease-specific #hashtags and shortened
URLs.
Automatically schedules the tweet
for posting using social media content
management system.
10. Why Twitter?
Symplur: The Rise of Patient Communities on Twitter by Auden Utengen.
Sep 2010 June 2012
Growth of Patient Communities on Twitter (green bubbles)
11. Measuring Popularity of Hashtags
For example, within a 24-hour period there were…
1,500 tweets posted
using #diabetes;
reaching 1.6 million Twitter users
Source: Hashtracking.com, Oct 8th, 2012
14. List of accounts twitter.com/UCSFRemix/ucsf-disease-
research/members
Tweet stream:
twitter.com/UCSFRemix/lists/ucsf-disease-research
15.
16. Automatically Post Relevant Tweets
New scientific publications: PubMed
New clinical trials: ClinicalTrials.goc
Links to a researcher profile
University news articles
Retweets of relevant content from University groups
Retweets of relevant content, copyedited by a communicator
@UCSFDiabetes
17. Results after 6 Weeks
Key Metrics
Total number of…
After 6 weeks
Followers 867
Tweets generated and sent 1,042
Clicks by Twitter users 1,149
Active Outreach
18. Results after 1 year and 4 Months
Key Metrics
Total number of…
After 6 weeks After 1 year
and 5 Months
Followers 867 3,094
Tweets generated and sent 1,042 3,442
Clicks by Twitter users 1,149 2,365
Active Outreach No Active Outreach
19. What Content is Most Popular?
Clinical Trials
Researcher
Profile
Publication
University News
Story
Retweet
Copyedited
Retweet
-
0.05
0.10
0.15
0.20
0.25
0.30
- 0.50 1.00 1.50 2.00 2.50 3.00
Retweetspertweet
Clicks per tweet
Interactions per tweet, by tweet type
20. Retweets See the Most Engagement
AverageClicks
T-Test with Bonferroni Correction for Multiple Comparisons
22. We Thank …
This project was funded through an IT Innovation Contest
Award from the University of California, San Francisco
(UCSF) and supported by the Clinical and Translational
Science Institute (CTSI) at UCSF.
23. Conclusions
Algorithmic content creation can accelerate and enhance
the traditional process of content creation at little cost.
Retweets of research-related content see the highest
engagement.
Disease communities value such an effort.
University groups in charge of communications save time
while increasing their information output.
24. Recommendations
Algorithmic content creation can support new types of
hybrid content that is collaboratively created by humans
and machines.
Potent model for ongoing value generation to foster
patient loyalty and research participant recruitment.
More research is necessary to assess the effectiveness of
different types of content.
Consider involvement of influencers.
The green bubbles have grown in numbers and significance quite dramatically in past months.
We started off by looking at 186 different biomedical hashtags.For each one, we looked at three factors:1. How frequently people post using that hashtag — are people talking about this?2. How many publications on that topic were published at UCSF the previous year — do we have anything to say about this?3. The number of Twitter accounts that include that hashtag in the description — do people strongly self-identify as being involved in this community?From that list, we selected 8 hashtags that met our criteria all 3 ways — people cared about it, people were taking about it, and we had something to say about it
So we created eight new Twitter accounts, one for each of the topics. They were accounts like @UCSFDiabetes, @UCSFWeightLoss, and @UCSFDepressionAnd on each account, we’d automatically post tweets of several different types
So we created eight new Twitter accounts, one for each of the topics. They were accounts like @UCSFDiabetes, @UCSFWeightLoss, and @UCSFDepressionAnd on each account, we’d automatically post tweets of several different types
We started off doing six weeks of active outreach using social media best practices, trying to get users to follow and interact with us by doing 4 things:Follow other Twitter usersMentioning other influential Twitter users in our tweetsPosting #FollowFriday tweets mentioning other influential Twitter usersActively soliciting feedback from readers
And then we stopped. And let the system run on auto-pilot. After 17 months, our 8 accounts had a combined total of over 3,000 followers, and over 2300 clicks on the tweets we posted
We started off looking at which types of content got retweeted and clicked on the most. Unsuprisingly, having a tweet composed of the full name of a publication was pretty much the least popular option — probably because that text’s not written for lay people to read
T-tests with bonferroni correction for multiple comparisonsWhile there wasn’t enough data for us to validate all the tweet types, it’s clear that retweets saw the most engagementOf these, both types of RTs performed significantly better than either profiles or pubmed links (p < 0.01 corrected)
…and we got really positive feedback from patients