1. Using
Twi)er
Data
to
Predict
Flu
Outbreak
Son
Doan
Division
of
Biomedical
Informa2cs
University
of
California
San
Diego
BigData@UCSD
workshop
Nov
25,
2013
2. Seasonal
influenza
and
influenza-‐like
illness
• Seasonal
influenza
is
a
major
public
health
concern:
• 3-‐5
million
cases
of
severe
illness
• 250,000
to
500,000
deaths
worldwide
each
year
• Seasonal
influenza
has
main
syndrome
called
Influenza-‐
Like
Illness
(ILI)
• During
the
peak
of
a
major
outbreak
of
influenza,
more
cases
of
ILI
are
observed
à
Monitoring
ILI
can
help
in
predict
flu
outbreak
3. Tradi?onal
system
to
monitor
ILI:
ILINet
• ILINet:
CDC’s
U.S.
Outpa2ent
ILI
Surveillance
Network
– consists
of
>3,000
outpa2ent
healthcare
providers
– all
50
US
states
and
area
– reports
more
than
30
million
pa2ent
visits
each
year
• ILINet
monitors
influenza
through
ILI
rate
– ILI
rate
is
percentage
of
pa2ents
with
ILI
among
all
pa2ents
– Average
na2onal
baseline
ILI
rate
for
2013
is
2.0%
5. Let’s
revisit
the
process
Pa2ent
1
Pa2ent
2
visits
Healthcare
provider
Check
if
ILI
visits
Healthcare
provider
Check
if
ILI
Healthcare
provider
Check
if
ILI
…
Pa2ent
n
visits
ILINet
gather
data
and
then
calculate
ILI
rate
6. ILINet
issue
ILINet
needs
1-‐2
weeks
to
gather
and
process
data
Can
we
leverage
other
data
sources
to
predict
ILI
rate
faster?
7. Nowadays,
users
tend
to
find
informa?on
in
Internet
User
1
User
2
searches
searches
Internet
…
User
n
searches
8. …
or
tweet
their
personal
health
condi?ons
User
1
User
2
tweets
tweets
Internet
…
User
n
tweets
9. Es?mate
ILI
rate
using
user-‐generated
data
• Models
– Linear
model
[1]:
ILI
rate
=
(ILI-‐related
data)Ÿα
+
error
– Logis2c
regression
[2]:
logit(ILI
rate)
=
logit(ILI-‐related
data)Ÿα
+
error
• Key
point:
How
to
iden2fy
ILI-‐related
data?
• Hint:
ILI
is
defined
as
fever
(temperature
of
100°F
[37.8°C]
or
greater)
and
cough
and/or
sore
throat
[1]
Polgreen
et
al.
“Using
internet
searches
for
influenza
surveillance”,
Clinical
Infec2ous
Disease,
2008,
47(11):1443-‐8.
[2]
Ginsberg
et
al.
“Detec?ng
influenza
epidemics
using
search
engine
query
data.”,
Nature.
2009
Feb
19;457(7232):1012-‐4
10. GFT
es?mates
based
on
flu-‐related
queries
are
highly
correlated
to
ILI
rate
Repor2ng
lag
of
about
1
day
Source:
hVp://www.google.org/flutrends/about/how.html
11. GFT
is
good,
however…
• Researchers
cannot
access
original
data
• GFT
does
not
disclose
search
queries
Source:
Ginsberg
et
al,
Nature
457,
1012-‐1014
(19
February
2009)
13. Twi)er
corpus
Timeline:
36
weeks
for
the
US
2009
influenza
season
(Aug
30,
2009
to
May
8,
2010)
Name
Total
25 mil
Tweets
587,290,394
Unique
23,571,765
users
URL
136,034,309
Hash
Tags
20 mil
15 mil
10 mil
96,399,587
5 mil
Thanks
to
Brendan
O’Connor
(CMU)
and
TwiVer
Inc.
14. Related
work
Twi)er
corpus
ILI-‐related
tweets
Culo)a4
Signorini3
Chew3
flu
swine
h1n1
cough
flu
swine
flu
headache
influenza
swineflu
sore
throat
[3]
A.
CuloVa,
“Detec2ng
influenza
epidemics
by
analyzing
twiVer
messages,”
arXiv:1007.4748v1
[4]
A.
Signorini,
A.
M.
Segre,
and
P.
M.
Polgreen,
“The
Use
of
TwiVer
to
Track
Levels
of
Disease
Ac2vity
and
Public
Concern
in
the
U.S.
during
the
Influenza
A
H1N1
Pandemic,”
PLoS
ONE,
vol.
6,
no.
5,
p.
e19467,
05
2011.
[5]
C.
Chew
and
G.
Eysenbach,
“Pandemics
in
the
Age
of
TwiVer:
Content
Analysis
of
Tweets
during
the
2009
H1N1
Outbreak,”
PLoS
ONE,
vol.
5,
no.
11,
p.
e14118,
11
2010.
16. Correla?on
to
ILI
rate
(CDC
data)
Method
Google
Flu
Trends
Pearson
corr
with
ILI
rate
0.9912
Related
work
CuloVa4
0.9485
Filter
1
Respiratory
syndrome
+
“flu”
-‐
URL
0.9752
Filter
1+2
Nega2on
+
Emo2con
+
HashTags
+
Humor
+
Geo
0.9846
17. %
Correla?on
to
ILI
rate
(CDC
data)
S.
Doan,
L.Ohno-‐Machado,
N.
Collier,
"Enhancing
TwiVer
Data
Analysis
with
Simple
Seman2c
Filtering:
Example
in
Tracking
Influenza-‐
Like
Illnesses",
Proc.
of
the
2nd
IEEE
HISB
2012,
pp.62-‐71,
2012.
18. Big
Data
challenge
Twi)er:
140
millions
ac?ve
users
340
millions
tweets/day
Twitter API sampling rate is
small (1-5% data)
Filtered tweets: 0.2% of samples
Is
sampling
data
enough?
19. DIZIE:
system
for
syndromic
surveillance
using
Twi)er
Syndromic
surveillance
for
gastrointes?nal,
respiratory,
neurological,
dermatological,
haemorrhagic,
musculoskeletal
from
Tweets
in
40
world
ci2es.
20. Use
cases
• DIZIE
was
integrated
to
BioCaster,
our
news
media
biosurveillance
system
• DIZIE
was
used
by
European
Centre
for
Disease
Preven2on
and
Control
(ECDC)
to
track
syndromes
in
the
London
2012
Summer
Olympics
21. Poten?al
applica?ons
using
Twi)er
in
public
health
• Mental
Heath
Analysis
• Tobacco
surveillance
• Medica2on
use
in
social
media
22. Acknowledgements
• Nigel
Collier,
European
Bioinforma2cs
Ins2tute
• Mike
Conway,
UCSD
• Lucila
Ohno-‐Machado,
UCSD
24. Data
source
for
influenza
surveillance
•
•
•
•
•
Data
provided
by
physicians
and
laboratory
Over-‐the-‐counter-‐drug
sales
School
absentee
records
Health-‐related
phone
calls
Internet-‐based
data:
– News
media
– Mailing
list
– Social
media
25. Extract
respiratory
syndrome
keywords
achy
chest
cold
symptom
respiratory
failure
apnea
cough
runny
nose
asthma
dyspnea
short
of
breath
asthma?c
dyspnoea
shortness
of
breath
blocked
nose
gasping
for
air
sinusi?s
breathing
difficul?es
lung
sounds
sore
throat
breathing
trouble
pneumonia
stop
breathing
bronchi?s
rales
stuffy
nose
…
…
…
We
have
a
total
of
37
keywords
26. Knowledge-‐based
approach
Name
Example
Respiratory
syndrome
only
tweets
containing
syndrome
keywords
Barber just coughed
on me in the chair.
Respiratory
syndrome
+
“flu”
tweets
containing
syndrome
keywords
and
“flu”
I got flu n coughed a
lot.
Respiratory
syndrome
tweets
containing
+
“flu”
-‐
URL
syndrome
keywords
and
“flu”,
remove
links
7-year-old boy dies of
flu,pneumonia < URL>
27. Seman2c
level
filtering
Name
Examples
Nega?on
Remove
nega?on
in
tweets
I don’t have flu
Emo?con
Remove
tweets
containing
smiley
emo?cons,
e.g.,
:-‐),,:D
Glad to hear that you’re beating
the flu. :-) Hope you don’t get the
nasty cough that everyone’s
getting this year
HashTags
Keeps
tweets
containing
keyword
“flu”
Still coughing smh #swineflu
#h1n1
Humor
Remove
humor
features
in
tweets,
e.g.,
“haha”,”hihi”,
“***cough
…
cough***”
Hm Im kinda wanting to go to
NYC really soon ***cough …
cough*** @Ctmomofsix =)
Geo
Tweets
from
graphical
loca?ons
(e.g.,
US)
28. Seman2c-‐level
filtered
tweets
Types
Tweet
samples
Influenza
confirma?on
I got flu n coughed a lot. Now my voice is like
monster’s voice. Rrr
Influenza
symptoms
My day: flu-like symptoms (headache, body aches,
cough, chills, 100.9 fever). Swine flu not ruled out.
#H1N1
Flu
shots
I’m still getting flu shots, nothing is worth flu turning
into bronchitis into pneumonia
Self
protec?on
Cover your mouth if coughing, use a tissue, wash
your hands often & get a flu shot - protect and
defend your community from #H1N1
Medica?on
Wondering why I didn’t take the flu shot, laying in
bed with cough drops, medicine, and the remote