The LLNL FMD Decision Support System:

1
The LLNL FMD Decision Support System Concise
Description of Features and Output
Tanya Kostova T. Bates, C. Melius, S. Smith, A.
Robertson, S. Hazlett, P. Hullinger, Lawrence
Livermore National Laboratory
DIMACS Workshop March 2006 Data Mining and
Epidemiological Modeling
2
LLNL is developing a decision support system for
evaluation of the economic impact of FMD
epidemics

• Effort funded by the Department of Homeland
  Security
• DHS has numerous ST investments in research
  projects for agriculture security countermeasures
  and requires tools to help evaluate future
  investments

3
LLNL is developing a decision support system for
evaluation of the economic impact of FMD
epidemics

• Effort funded by the Department of Homeland
  Security
• DHS has numerous ST investments in research
  projects for agriculture security countermeasures
  and requires tools to help evaluate future
  investments
• Numerous FMD epidemiological models exist but
• They are not national in scale
• Current models target natural or accidental
  introduction not an intentional act
• Epidemiological and economic models are not
  coupled

4
GENERAL FEATURES OF THE EPIDEMIC MODEL
Agent-based spatially-explicit discrete-time
computational model
Time progresses in increments of 1 unit (1 day)
5
GENERAL FEATURES
Agent-based spatially-explicit discrete-time
computational model
Time progresses in increments of 1 unit (1 day)
In a time stepping agent based model, at each
time increment some of the agents change some of
their attributes depending on their previous
state and on the previous states of some of the
other agents.
6
GENERAL FEATURES
Agent-based spatially-explicit discrete-time
computational model
Time progresses in increments of 1 unit (1 day)
In a time stepping agent based model, at each
time increment some of the agents change some of
their attributes depending on their previous
state and on the previous states of some of the
other agents.
The FMD model agents are the animal facilities.
7
GENERAL FEATURES
Agent-based spatially-explicit discrete-time
computational model
Time progresses in increments of 1 unit (1 day)
In a time stepping agent based model, at each
time increment some of the agents change some of
their attributes depending on their previous
state and on the previous states of some of the
other agents.
The FMD model agents are the animal
facilities. Facilities are groups of animals
managed in a specific manner.
Farms, Markets, Feedlots, Slaughter houses
8
THE ATTRIBUTES OF THE FACILITY AGENT
Type (incl. species, size and
operation)

Spatial coordinates
Static
Average Number of Contacts (to and from),
Method of disease spread specific network of
contacts
Dynamic
Disease states
Change due to interaction
Availability Seasonal factors
Change externally and independently of interaction
9
THE ATTRIBUTES OF THE FACILITY AGENT Type
The current model version deals with 34 types of
animal facilities
Beef(B), Dairy(S), Dairy(M), Dairy(L), Dairy(B),
Grazing(S), Grazing(L), Feedlot(S), Feedlot(L),
Stocker(S), Stocker(L) Swine(B), SwineFWean(S),
SwineFWean(L), SwineFinish(S), SwineFinish(L),
SwineNursery(S), SwineNursery(L),
SwineFFeeder(S), SwineFFeeder(L), SwineFarFin(S),
SwineFarFin(L), Sheep(S), Sheep(L), Sheep(B),
Goats, Goats(B), Market, Market(Cattle),
Market(Swine), Market(Other), Market(L),
Market(C-L), DCalfHeifer(L)
10
THE ATTRIBUTES OF THE FACILITY AGENT
The spatial coordinates of each facility are
exact up to the county level
The NASS data supplies the numbers of different
facility types in each county
There are 1.2M facilities (according to NASS
data) with 160M animals. These do not include
markets which come from another database. Thus,
we model 1.2M facilities and their contacts.
11
THE ATTRIBUTES OF THE FACILITY AGENT
The spatial coordinates of the facilities are
generated using a random algorithm based on the
county-based data.
Hogs and pigs Cattle and cows Sheep
12
THE ATTRIBUTES OF THE FACILITY AGENT
Type (incl. species, size and
operation)

Spatial coordinates
Static
Average Number of Contacts (to and from),
Method of disease spread specific network of
contacts
Dynamic
Disease states
Change due to interaction
Availability Seasonal factors
Change externally and independently of interaction
13
THE ATTRIBUTES OF THE FACILITY AGENT

Depends on the size and type of facility and
determined for each specific facility as random
number drawn from a given probability
distribution obtained from survey data
Average Number of Contacts (to and from),
Method of disease spread specific network of
contacts
14
THE ATTRIBUTES OF THE FACILITY AGENT

Depends on the size and type of facility and
determined for each specific facility as random
number drawn from a given probability
distribution obtained from survey data
Average Number of Contacts (to and from) Method
of disease spread specific network of contacts
Direct (regional and inter-state) Indirect (high
risk and low risk)
15
THE ATTRIBUTES OF THE FACILITY AGENT
Type (incl. species, size and
operation)

Spatial coordinates
Static
Average Number of Contacts (to and from),
Method of disease spread specific network of
contacts
Dynamic
Disease states
Change due to interaction
Availability Seasonal factors
Change externally and independently of interaction
16
THE ATTRIBUTES OF THE FACILITY AGENT
Disease states
Suspected
Confirmed
Culled
S - Susceptible (healthy) L- Latent U-
Subclinically infectious I- Clinically
infectious W Vaccinated and susceptible V-
Vaccinated M- Immune P- Suspected F- Confirmed X
- Culled
Immune
Waning of immunity
L3
L2
L
L1
Latent (infected)
Subclinically infectious
Susceptible
Clinically infectious
?
Infection
Vaccinated
17
THE ATTRIBUTES OF THE FACILITY AGENT
The disease state attributes of each facility are
calculated by an intra-facility model (IFM)
The intra-facility model is a time-since
infection Reed-Frost type model Represents a
discrete-time system of difference equations
representing the number of animals on the
facility that are in each state S, L, I, U , V,
W, M
18
THE ATTRIBUTES OF THE FACILITY AGENT
The disease state attributes of each facility are
calculated by an intra-facility model (IFM)
The intra-facility model is a time-since
infection Reed-Frost type model Represents a
discrete-time system of difference equations
representing the number of animals on the
facility that are in each state S, L, I, U , V,
W, M The output of the IFM is used to calculate
the probability that an infected facility will
infect other facilities This is done by using a
spread model
19
THE ATTRIBUTES OF THE FACILITY AGENT
Average Number of Contacts (to and from) Method
of disease spread specific network of contacts
Availability Seasonal factors
These attributes are used by the Spread Model to
calculate the newly infected facilities
20
The Spread Model calculates the newly infected
facilities
Infected agents can spread the epidemic via
various methods along method-specific networks
Examples of methods - direct (movement of
animals) - indirect personnel movements -
inter-state direct movements
Truck routes network
For each method, the infection can be spread
within a predefined set of facilities specific to
the method. Thus, an infected facility will
spread the infection to the facilities within
the networks to which it belongs.
Vet routes network
not infected
infected
21
The epidemic spread is modeled by a random
process
Uses information about the Average Number of
adequate Contacts ANC of the infected facility
by each of the methods The daily number
of adequate contacts RANCmi is obtained from a
Poisson process with mean ANC
A contact originating from a facility that can
cause infection is an adequate contact. An
adequate contact that actually infects a target
facility is an effective contact.
STEP 1
For each method of infection m
For each infected facility i - A
probability density function Pmi(j)
defined on each of the nodes j of the
network Smi of m and i is calculated
- For each node j of Smi the probability Cmj
is calculated
STEP 2
Pmi(j) is the probability that facility j will
get a contact with facility i by method m.
Distance dependent
Cmj is the probability that an adequate contact
to facility j will cause infection.
Pmi(j) is used in a roulette algorithm to
determine which facilities receive an adequate
contact Cmj is used to determine which of the
contacted facilities become infected RANCmi,
Pmi(j) and Cmj are used to trace back the cause
of infection of j
STEP 3
22
The Spread Model involves factors sampled from
PDFs
Pmi(j) depends on - the average number of
m-type contacts received by j - size of the
facility j - seasonal factors -
control measure factors - distance between i
and j - frequency of contacts between i and j
Cmj depends on - the fraction of vaccinated
animals on the facility - control measure
factors - probability that a contact of type m
would cause infection
Many of these factors are uncertain or involve
variability and are sampled from probability
density functions.
23
The Control Measures Component
Control measures include
Vaccination
Culling Contact
restrictions
Isolation Increased
detection
Control measures are applied regionally

24
Events during one increment of time
25
AGGREGATION ALGORITHMS
Our model is of US - national scale however to
keep calculations to a minimum - We do not
calculate all facilities at all times. - Only
facilities in infected and their neighboring
counties are initialized - Intra-facility
model calculated only for infected facilities
- Counties and states that have not been yet
infected are considered as aggregated entities
if a contact happens to in such a county, it gets
disaggregated.
26
OUTPUTS
A simulation is made of N MC runs N O(102) -
O(103)
27
OUTPUTS
A simulation is made of N MC runs N O(102) -
O(103)
A run implements M time steps M O(102),
usually 200-330 days or until a
certain criterion is met (epidemic comes to end)
28
OUTPUTS
A simulation is made of N MC runs N O(102) -
O(103)
A run implements M time steps M O(102),
usually 200-330 days or until a
certain criterion is met (epidemic comes to end)
At each time step we keep track of the number P
of facilities that are currently involved in the
epidemic (i.e. the ones that are infected or in
the neighborhoods of infected facilities. P
O(102) - O(105) ???
29
OUTPUTS
A simulation is made of N MC runs N O(102) -
O(103)
A run implements M time steps M O(102),
usually 200-330 days or until a
certain criterion is met (epidemic comes to end)
At each time step we keep track of the number P
of facilities that are currently involved in the
epidemic (i.e. the ones that are infected or in
the neighborhoods of infected facilities. P
O(102) - O(105) ???
For each facility the important data (current
states, costs, trace-back facilities) is O(101)
30
OUTPUTS
O(1010)
Thus, the total output of a simulation could be
in the range of or more.
Naturally, we do not keep all this output
although what we do not keep may be important for
the analysis
What do we keep currently?
31
OUTPUTS
Daily Numbers of facilities of the 34 types that
are in the 9 disease states
L- Latent U- Subclinically infectious I-
Clinically infectious W Vaccinated and
susceptible V- Vaccinated M- Immune P-
Suspected F- Confirmed X - Culled
Numbers of facilities that have just acquired a
new state Numbers of facilities that have ever
been in some disease state Total numbers of
infected, vaccinated, culled facilities Daily
and total numbers of infected, vaccinated, culled
animals of different species
32
OUTPUTS
Durations Lengths of time for which the 34
types of facilities were in some disease
state Duration of total epidemic
Costs associated with epidemic and control
measures
33
OUTPUTS
Currently, output is in Excel spreadsheet format
and is used for visualization
As well as to calculate statistics (means,
quantiles, skewness, kurtosis, etc.) of MC output.
34
Epidemic model outputs and data mining
Question How can modern data mining tools help
in the analysis of output data generated by a
large-scale epidemic model?
35
Epidemic model outputs and data mining
Question How can modern data mining tools help
in the analysis of output data generated by a
large-scale epidemic model? Specifically, can
data mining help uncover important relations
between - scope of epidemic and spatial
distributions of facilities? - how
control measures are applied and the cost of the
epidemic?
36
Epidemic model outputs and data mining
Further, can data-mining tools help
Identify sources (infected facilities), likely
transmission mechanisms? Classify of
outbreaks into "natural" vs. "intentional" to
help policy makers develop correct response
strategies?
37
Epidemic model outputs and data mining
Further, can data-mining tools help
Identify sources (infected facilities), likely
transmission mechanisms? Classify of
outbreaks into "natural" vs. "intentional" to
help policy makers develop correct response
strategies? Identify key facilities/locations
for surveillance? Identify which control
mechanisms are having the largest impact?
38
Epidemic model outputs and data mining
Further, can data-mining tools help
Identify sources (infected facilities), likely
transmission mechanisms? Classify of
outbreaks into "natural" vs. "intentional" to
help policy makers develop correct response
strategies? Identify key facilities/locations
for surveillance? Identify which control
mechanisms are having the largest impact?
Evaluate new technologies? Evaluate
vulnerability of different industries and regions
of the country?
39
If the answer is yes to at least some of our
questions, which are the recommended data mining
tools? Are they available?