Post on 29-Nov-2018
From prediction to practice: Integrating forecasting
models into public health education and response
KACEY C. ERNST, ASSOCIATE PROFESSOR OF EPIDEMIOLOGY AND BIOSTATISTICS, UNIVERSITY OF ARIZONA, TUCSON, AZ
KERNST@EMAIL.ARIZONA.EDU 520-626-7374
Transmission risk in a given
geographic area
Transmission potential
Detection and Response infrastructure
Human Environment: Mobility patterns, Behavior and behavior
change, Social and political infrastructure
Human – Natural Environment interactions: Land use change, human-vector-zoonotic
host interactions
Baseline Natural Environmental Suitability – vectors, pathogens, zoonotic hosts
Some key concepts identified from: Arthur RF, Gurley ES, Salje H, Bloomfield LS, Jones JH. Philos Trans R Soc Lond B Biol Sci. 2017 May 5;372(1719).
Long term predictions to early
warning and early detection
Time
Ca
ses
50yrs 20yrs 10yrs 5yrs 1yr 6mo 3mo 1mo onset peak end
Early
detection
systems
Early warning
systems
Long-term
prediction
Uncertainty
Driven by:
-climate
change
scenarios
-Population
projections
Driven by:
-seasonal
forecasts
-current
census
information
Sources:
-syndromic
- data
mining
- HC-based
- CBP
Traditional
surveillance
Use in public health response and planning
Projecting long-term epidemic potential
Understand the system What is the process that
climate/weather influences the infectious disease
potential?
Determine patterns of seasonality
Conduct correlations across differing geographies
Laboratory experimentation
Mitigating factors
If all are exposed, who are the vulnerable?
Examine risk factors for the transmission
Determine the current and projected distribution of
these risk factors
Compile into projections
If we drive process models with projected climate
data coupled with data on the human dimension, what
happens?
Determine complex interactions among all the
determinants of the process and the potentially mitigating
risk factors
Map shows the range of the Aedes aegypti mosquito for present-day (1950-2000) and future (2061-2080; RCP8.5) conditions.
Larger cities have higher potential for travel-related virus introduction and local virus transmission. Adapted from:
Andrew J. Monaghan, K. M. Sampson, D. F. Steinhoff, K. C. Ernst K. L. Eb B. Jones, M. H. Hayden, Climatic Change (2016)
Present-DaySuitabilityforAedesaegyp (Types1-3)
FutureSuitabilityforAedesaegyp (Types1-3)
Coun eswithrecentlocaldengueorchikungunyatransmission
Philadelphia
NewYork
Washington
Miami
TampaOrlando
Houston
Dallas
LosAngelesSanDiego
Denver
Sacramento
Atlanta Charleston
SanAntonio
Phoenix
St.Louis
MetroPopula on
5,000,000
20,000,000
Ae. aegypti virus transmission suitability
Predicting risk must account for
changing human factors
0
20
40
60
80
100
120
140
160
180
200
0
100000
200000
300000
400000
500000
600000
700000
800000
2000 2005 2010 2015
Num
be
r o
f bed n
ets
dis
trib
ute
d
(in m
illio
ns)
Ma
laria d
eath
s u
nder
age 5
years
Malaria deaths and bed net distribution
Malaria Deaths Under Age 5 years Bednets distributed
Linear (Bednets distributed) Data from WHO
Malaria deaths
under age 5 years
Bednets
distributed
Risk influenced by socially-determined
behaviors
0
0.5
1
1.5
2
2.5
3
3.5
4
Primary School Secondary School More than secondarySchool
Odds tha
t a w
om
an w
ill o
wn a
bed
ne
t com
pare
d to
tho
se
w
ith
no
ed
ucation
Women with an education are more likely to own a bed net in the highlands of Kenya
Developing Early Warning systems
for infectious disease transmission Frameworks for early warning systems have been developed that include stages
for:
Watch: Developing prior assessment of risk of emergence event
Warning: Human disease has been detected
Emergency: Epidemic or outbreak is underway
Key challenges
Integration of data streams
Changing landscapes of risk
Human: behaviour and available control strategies, response capacity
Biological: phylodynamics, environmental conditions
Communication and adoption of early warning systems by key stakeholders
Investment in ongoing monitoring systems
Lack of integration of natural system risk with impact of human and social factors on transmission
Some concepts adapted from BA Han and JM Drake EMBO Reports 2016
Monaghan AJ, Morin CW,….Ernst
K. PLOS Currents (March 2016)
Zika Risk in CONUS
• Weather-driven
mosquito models
with
• travel,
• socioeconomic
conditions
• virus history
• Required rapid
analysis
• Designed for
widespread
dissemination to
stakeholders and the
public.
• One time assessment
• Used climate not
current weather
Early Warning
Example
Examples from other
fields
Sustainable forecasting has been actualized in other disciplines
Drought monitoring
Food security
Typically more directly related to environmental measurements readily available
Temperature
Rainfall
Do not include projects of health, economic or other downstream impacts in which social science has a stronger role
http://www.fews.net/
https://www.drought.gov/drought/
Early detection
Goal of early detection
Reduce the reproductive
number to minimize
transmission
Methods:
Reduce contacts
Reduce duration of
infectiousness
Target interventions
Maximize uptake
Components of Basic Ro:
Β = Probability of transmission given contact
c = Number of contacts per given time period
D = Duration of infectiousness
Can modify by: x – proportion susceptible - Targeted vaccination (measles OB) - Distribution of control measures (bed net)
Non-traditional Early Detection sources
Social listening: Using social media data to identify trends and respond (often used in marketing but can be used for infectious disease trends)
Sources of data
Twitter (predomínate source)
Internet searches
Key strategy
Develop algorithims for searching for key words/ phrases – machine learning (Kagashe I, Yan Z, Suheryani I J Med Internet Res 2017)
Monitor trends
Key Challenges
Biases in data (age, geography)
Noisy data – best suited for trends in high transmission diseases (influenza (n=50), dengue (some regions) etc.)
Other factors may influence discussion of a topic
Fig 2. Tweets are a useful tool for estimating Dengue activity at country level.
Marques-Toledo CdA, Degener CM, Vinhal L, Coelho G, Meira W, et al. (2017) Dengue prediction by the web: Tweets are
a useful tool for estimating and forecasting Dengue at country and city level. PLOS Neglected Tropical Diseases 11(7):
e0005729. https://doi.org/10.1371/journal.pntd.0005729
http://journals.plos.org/plosntds/article?id=10.1371/journal.pntd.0005729
Dengue in Brazil
• Tweets strongly
correlated
with case reports
• County level
• City level
• Nowcasting
• Forecasting up to 8
weeks
Non-traditional early detection
sources cont.
Community-based participatory surveillance
Recruited group of users report symptoms each week
Examples
FluNearYou
successfully implemented (Pearson corr. >90% adjusted)
Biased population
More females
Higher human development index
Smolinski, M. S., Crawley, A. W., Baltrusaitis, K., Chunara, R., Olsen, J. M., Wójcik, O., . . .
Brownstein, J. S. (2015). 105(10), 2124-2130.
Baltrusaitis K, Santillana M, Crawley AW, Chunara R, Smolinski M, Brownstein JS.
JMIR Public Health Surveill. 2017 Apr 7;3(2):
Kidenga – CBS for vector-borne
System:
Similar to Flu Near You
Monitoring of syndromes and mosquito activity in US-Mexico border región
Key lessons
Recruitment and retention of participants is challenging
Need to provide tailored messages to motivate action
Interest falls off after transmission wanes
Too high numbers required for identification of rare events
Current utility
Platform for educational messaging and dissemination of EWS messages
Intention – get it on everyone’s phone – periodic activitation
Kidenga 2.0: Iterate with alerts and cues to action
“relating it
to the weather”
“‘some kind
of notification
that there was activity”
‘a good
prompt for me in my
environment’
‘”check- ins”
Summary Identify the complex processes behind disease transmission – integrate human and natural processes
Parameterize models for predicting future risk – both long and short term
Invest in sustainable infrastructure to support forecasting efforts
Identify best practices to convey information to motivate action in stakeholders and public
Operationalize messaging at an appropriate and actionable spatial and temporal scale
Acknowledgements 19
University of Arizona
Dr. Mike Riehle
Dr. Kathleen Walker
Teresa Joy
Dr. Pablo Castro-Reyes
Dr. Yves Carriere
Dr. Andrew Comrie
Dr. Cory Morin
Steve Haenchen
Eileen Jeffrey
Daniel Williamson
NCAR
Dr. Mary Hayden
Dr. Andy Monaghan
Dr. Daniel Steinhoff
NIH/NIAID: Grant R56AI091843 and 1R01AI091843
and Skoll Global Threats Grant
Research was approved by
University of Arizona,
Directorate of Education, Quality and Research
Ethics Committee of the Department and
Medicine and Health Sciences of the
Universidad de Sonora
Universidad de Sonora
Dr. Gerardo Alvarez
Dra. Maria del Carmen Candia Plata
El Colegio de Sonora
Dra. Lucia Castro
Dr. Rolando Diaz Caravantes
Carmen Arellano Gálvez
Office of Border Health
Robert Guerrero
Orion McCotter
Sonoran Ministry of Health Department of
Epidemiology and Department of Health
Services and Directorate of Education
Dr. José Jesús Bernardo Campillo
García.
Dr. Francisco Javier Navarro Gálvez.
Dr. Sergio Olvera Alba.
Dr. Ariel Vázquez Gálvez.
Hospital General, Nogales
Mercedes Gameros
Centers for Disease Control
• Dr. Steve Waterman
• Alba Phippard