Bayesian Models for General Water Quality and Public Health Information Fusion Model

kettlecatelbowcornerAI and Robotics

Nov 7, 2013 (4 years and 6 days ago)

69 views

Bayesian Model
s

for
General Water Quality and Public Health
Information Fusion Model

Z
.

Mnatsakanyan, PhD
,
H
. Burkom, PhD, L.
Ramac
-
Thomas, PhD

Johns Hopkins University Applied Physics Laboratory


I
NTRODUCTION
:

Waterborne pathogens can be a serious public

health threat causing sever gastrointestinal
-
related (GI) illness. As a result of this concern,
the GI syndrome was created for use in
syndromic surveillance systems.
However
e
xisting

water distribution system
monitoring
and public health surveillance sys
t
ems
do not
perform

integrated

surveillance.
.
This project

create
s

information fusion algorithms that utilize
inputs from general water quality monitoring
sensors and information about medical facilities
visits. The objective is to detect anomalous
water
contamination

patterns and correlate them
with the possible health events in the
community.

M
ETHOD
:

GI related c
hief complaint data for different age
groups, were queried and processed with
statistical anomaly detection algorithms
[1]

using
Early Notificati
on of Community
-
based
Epidemics (ESSENCE) system
[
2
]
. Water quality
data
was
collected by the local

water

utilities
.
Our fusion process is a network of three
Bayesian Networks(BN)
[
3
](figure 1)
: GI BN
estimates detection algorithm outputs and
classify epidem
iologically significant alerts from
the ones that are just mathematical anomalies in
the GI data; Water BN estimates probability of
the water contaminations based on
data from
multiple sensor
s
; and Health/Water Fusion
BN
(Figure 2)

that estimates likelihood

that GI
events detected by the GI BN are waterborne.



Figure 1
. Bayesian Networks based system
architecture


Figure 2 Water and Health Fusion BN.


R
ESULTS
:

Eleven GI outbreaks were simulated. Each
o
utbreak involved 14

23 cases within 2

4 days.
Each case was represented by a randomly
selected GI relevant chief complaint. Outbreaks
were randomly injected in the background data.
The background data was created using one year
of archived real data that
was free of outbreaks.
The detection threshold was set up to 25% or
above probability of the outbreak. The model
detected 10 of the injected outbreaks, with 2
false positive detections during a 2
-
year period.
Water events was injected in the water quality
data. Health/Water BN
was alerting

when
high
probability of GI event was

following high
probability water contamination within 14
days.


C
ONCLUSION
:

Bayesian model for GI monitoring showed good
sensitivity to simulated outbreaks. Preliminary
results for He
alth/Water BN showed that water
contamination and public health events can be
successfully detected by fusing public health
information and water quality sensors data.


R
EFERENCES
:

1.

Burkom H. Development, adaptation, and
assessment of alerting algorithms fo
r
biosurveillance.
Johns Hopkins APL Technical
Digest
2003;
24
: 335
-
342.

2.

Lombardo JS, Buckeridge DL, editors.
Disease Surveillance: A Public Health
Informatics Approach. Hoboken (NJ): John
Wiley & Sons, Inc.;2007.

3.

Pearl J. Fusion, propagation, and structur
ing
in belief networks.
Artificial Intelligence

1986;
29
(3): 241
-
288.

water_cont_severe
true
false
50.0
50.0
water_anomaly_sites_contin
none
one
more
33.3
33.3
33.3
anomaly_in_both
true
false
50.0
50.0
water_anomaly_sites_e_coli
none
one
more
33.3
33.3
33.3
zones_all
none
one
more
33.3
33.3
33.3
Adults_18_44
true
false
50.0
50.0
Children
true
false
50.0
50.0
Elderly
true
false
50.0
50.0
type_c_alert
true
false
50.0
50.0
water_anomaly_sites_e_coliform
none
one
more
33.3
33.3
33.3
Adults_45_64
true
false
50.0
50.0
Infants
true
false
50.0
50.0
any_age
true
false
50.0
50.0
health
true
false
50.0
50.0
diagnostic_cases
true
false
50.0
50.0
GI
true
false
50.0
50.0
type_a_alert
true
false
50.0
50.0
any_sensor_water
true
false
50.0
50.0
type_b_alert
true
false
50.0
50.0
conduct_all_sensors
none
one
two
more 10 percent
more more 25 percent
more 50 percent
16.7
16.7
16.7
16.7
16.7
16.7
turbidity_all_sensors
none
one
two
more 10 percent
more more 25 percent
more 50 percent
16.7
16.7
16.7
16.7
16.7
16.7
clorine_all_sensors
none
one
two
more 10 percent
more more 25 percent
more 50 percent
16.7
16.7
16.7
16.7
16.7
16.7
toc_all_sensors
none
one
two
more 10 percent
more more 25 percent
more 50 percent
16.7
16.7
16.7
16.7
16.7
16.7

Health Net

Estimates the Probability of

the Infectious GI Outbreaks


Water Net

Estimates Probability of

the Water Contamination


Water/Health Net

Estimates Probability of

the Wate
rborne Outbreak

W
W
a
a
t
t
e
e
r
r


Q
Q
u
u
a
a
l
l
i
i
t
t
y
y


S
S
e
e
n
n
s
s
o
o
r
r
s
s


ESSENCE