Download “NOISE” STILL PRESENT STRENGTHS HUMAN CASE

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
REMOTELY SENSED INDICATORS OF MICROCLIMATE IN PREDICTING NEW AREAS OF
HUMAN RISK OF LYME DISEASE USING SPATIAL
STATISTICS AND ARTIFICIAL NEURAL
NETWORKS
A PRESENTATION TO THE SUMMER COLLOQUIUM ON CLIMATE AND HEALTH
JULY 26, 2004, NCAR, BOULDER COLORADO
RUSSELL BARBOUR PH.D.
VECTOR ECOLOGY LABORATORY
YALE SCHOOL OF MEDICINE
NEW HAVEN CT.
PROBLEM STATEMENT
• HUMAN CASE DATA HAS BEEN PROVEN
AN UNRELIABLE INDICATOR OF LYME
DISEASE RISK
• UNDER REPORTING, MIS-DIAGNOSES,
AND OVER REPORTING DISTORT HUMAN
CASE DISTRIBUTION
• COLLECTION AND TESTING OF
INFECTED NYMPHS COSTLY
PROBLEMS CONTINUED
• Ixodes scapularis TICKS HAVE NOT EXPANDED
INTO ALL AREAS OF SUITABLE HABITAT
• INVADING TICKS ARE NOT NECESSARILY
INFECTED WITH Borrelia burgdorferei ( BACTERIAL
AGENT OF LYME DISEASE)
• ONLY INFECTED NYMPHAL TICKS POSE A
THREAT TO HUMANS
NEW APPROACH TO RISK ESTIMATION
AND PREDICTION
• INTEGRATE HUMAN CASE DATA WITH
LANDSCAPE INDICATORS OF THE NIDALITY
(FOCI) OF INFECTION OF Borrelia burgdorferi
• BUILD DATA LAYERS FROM REMOTELY SENSED
MICRO-CLIMATE INDICATORS, PUBLISHED
CANINE SEROPREVALENCE AND PREVIOUS
HUMAN CASE DATA
• DERIVE PROBABILITY OF INCREASING RISK
THROUGH MARKOV-BAYES MONTE CARLO
SIMULATIONS
KRIGING VERSUS
MARKOV-BAYES MONTE CARLO CHAIN (MMCC)
SIMULATION
•
KRIGING GIVES THE MOST LIKELY EVENT AT ALL LOCATIONS..
THE TOP OF A PROBABILITY DENSITY CURVE
•
KRIGING IS BASED ON JUST ONE “ ITERATION OF POSSIBLE
REALITY”
•
KRIGING DISHONORS THE ORIGINAL DATA …“ EVENTS MORE
PROBABLE THAN REALITY”
•
MMCC GIVES OTHER PROBABILITIES AT EACH LOCATION
•
MMCC HONORS THE ORIGINAL DATA
•
MMCC IS BASED ON METROPOLIS-HASTINGS RANDOM WALK
(ALGORITHM USED TO DEVELOP H-BOMB). THE NEXT STATE IS
ONLY DERIVED FROM THE CURRENT STATE
•
RANDOM WALK CREATES A NUMBER OF ITERATIONS
ALTHOUGH EVENTUALLY THEY WILL CONVERGE TO KRIGED
VALUES
EVI AS A FACTOR IN ESTIMATING
LYME DISEASE RISK
• MORE SENSITIVE TO PERIODS OF LIGHT
VEGETATION , SPRING AND FALL WHEN
NYMPHAL AND ADULT TICKS ARE ACTIVE
• DISTINGUISHES BETWEEN WOODED SUBURBS
AND TRUE FORESTS DURING THIS TIME PERIOD
• IDENTIFIES DISCONTINUITY IN LANDSCAPES
BETTER THAN NDVI
MODIS Products
MODIS
Ocean
Atmosphere
Land
Products:
MOD36 Ocean Color Products:
MOD09 Reflectance
MOD28 SST
…
MOD12 Snow Cover
MOD13 Vegetation
MOD14 Thermal
Products:
MOD04 Aerosols
MOD05 Water
Vapor
MOD06 Cloud
MOD35 Cloud Mas
…
Anomaly
…
SOURCE: http://modis.gsfc.nasa.gov/
MODIS Data products come in different
Spatial Resolution
250 m
500 m
1000 m
4 km
5 km
5 min
0.05 deg
0.25 deg
Etc…
Temporal Resolution
daily
8-day composites
16-day composites
96-day composites
Etc…
Versions
Version3 v003
Version4 v004
~But most products do NOT come with all these
resolutions and versions~
RELATIONSHIP BY DATE BETWEEN HUMAN
CASES AND EVI BY MOVING WINDOW ANALYSIS
HUMAN
CASES
MODIS EVI DATES
CORRELATION
1992
MAY25 2001
JULY 28 2001
.12
.01
1993
MAY 25 2001
JULY 28 2001
.11
.02
1994
MAY 25 2001
JULY 28 2001
.18
.06
1995
MAY 25 2001
JULY 28 2001
.14
.06
1996
MAY 25 2001
JULY 28 2001
.16
.06
1997
MAY 25 2001
JULY 28 2001
.14
.06
1999
MAY 25 2001
JULY 28 2001
.17
.06
2000
MAY 25 2001
JULY 28 2001
.20
.06
MATHEMATICAL DATA INTEGRATION
Most Abundant Data
NEW
REMOTELY SENSED
VEGETATION INDEX
(EVI)
PREVIOUS HABITAT
SUITABILITY MODEL
CANINE
SEROPREVALANCE
DATA POINTS
COMBINED
BY ANN
Sparse Data
1992- 2000
HUMAN CASE
DATA BY COUNTY
SPATIAL
STATISTICS &
MMCC SIMULATIONS
ESTIMATED HUMAN
CASES BY LOCATION
PREDICTIVE VALUE OF 1995 MID-WESTERN CASE DATA
WHEN INTEGRATED WITH PREVIOUS YEARS AND
LANDSCAPE INDICATORS OF INFECTION BY MULTILAYER
ARTIFICIAL NEURAL NETWORKS
YEAR
1995
1996
1997
1998
1999
2000
PREDICTIVE
VALUE
0.9986
0.9007
0.8420
0.8800
0.8779
0.8323
PREDICTIVE VALUE OF 1998 MID-WESTERN CASE DATA
WHEN INTEGRATED WITH PREVIOUS YEARS AND
LANDSCAPE INDICATORS OF INFECTION BY MULTILAYER
ARTIFICIAL NEURAL NETWORKS
YEAR
PREDICATIVE
VALUE
1998
.99
1999
.99
2000
.91
PROBABILITY OF HUMAN PREVALENCE
HIGHER THAN 25/100,000 FROM 1992 HUMAN
CASE DATA AND
LANDSCAPE INFECTION INDICATORS
URBAN AREAS
PROBABILITY
PROBABILITY OF HUMAN PREVALENCE
HIGHER THAN 25/100,000 FROM 2000 HUMAN
CASE DATA AND LANDSCAPE INFECTION
INDICATORS
URBAN AREAS
PROBABILITY
URBAN AREAS
1992 PROBABILITY OF
HIGH PREVALENCE
URBAN AREAS
2003 PROBABILITY OF
HIGH PREVALENCE
PROBABILITY MAP
MODEL AGREEMENT WITH CASE
DATA
• PREDICTED SPATIAL HUMAN LD PREVALENCE
BY FROM LANDSCAPE AND PREVIOUS HUMAN
CASE DATA AGREED WITH ACTUAL CASES BY
81%
WEAKNESS
• VEGETATION DATE TOO SPECIFIC
• LARGE AREAS OF UNCERTAINTY
• NO QUALITY CRITERIA FOR
ORIGINAL CASE DATA
• “NOISE” STILL PRESENT
STRENGTHS
• HUMAN CASE DATA LINKED TO NIDALITY OF
INFECTION
• REASONABLE PREDICTIONS OF HUMAN RISK
POSSIBLE
• THREE YEAR ADVANCE OF INFECTION WALL
APPEARS VISIBLE
WILDLIFE URBAN INTERFACE DATA
LOW DENSITY INTERFACE :
AREAS WITH HOUSING DENSITY BETWEEN 6.2 AND 49.4
HOUSING UNITS PER KM 2 AND 50% VEGETATION COVER
WITHIN ALL 2 KM AREAS WITH 75 % COVER
Source:
SILVIS Lab Spatial Analysis For Conservation And Sustainability
Forest Ecology & Management University Of Wisconsin - Madison
ASSOCIATED WITH
HUMAN LD CASES In WI
SPATIAL STRUCTURE OF 2000
HUMAN CASE DATA IN WISCONSIN
CROSS VARIOGRAM
2000 HUMAN CASES AND
LOW DENSITY WUI
CO-KRIGE OF 2000 HUMAN CASES
AND LOW WUI % LAND COVER
SPREAD OF INFECTED
NYMPHAL
Ixodes scapularis TICKS AS
ESTIMATED
FROM HUMAN CASES
FUTURE RESEARCH
• CUBIC SPLINE REGRESSION OF ALL
HUMAN CASE DATA TO REMOVE NOISE
• ADDITION OF MODIS ATMOSPHERIC
DATA TO CAPTURE HUMIDITY
• FILTER OF UNSUITABLE LANDSCAPES
FARMLAND
• CALCULATION OF THE RATE OF
INFECTION SPREAD, CURRENTLY ABOUT
6 KILOMETERS A YEAR, BASED ON
MMCC PROBABILITY MODELS, NOT
CASES