Transcript Document
Modeling West Nile virus Distribution from Surveillance Data Josh Bader 16 February 2009 University of California-Santa Barbara Department of Geography Outline • Biogeography Background • WNV Background • My Research – – – Conceptual model Predicting WNV distribution Identify optimal surveillance location Source: birds.cornell.edu Biogeography • Intersection of life sciences and geography – Also ecology, geology, molecular biology • Two divisions – Historical • Evolutionary perspective • Pleistocene Ice Age – Ecological • Modern persective • Why are species where they are (were)? Biology • Process behind spatial distributions • Without tolerance limits, a species will occupy all available areas--maximum dispersion • Limits determined by: – – – – Biotic factors Abiotic factors Genetics Population dynamics • Intraspecies and interspecies • Often related to fundamental niche of species Potential range map Deductive Approach Biological Information Geographic Information •Habitat Req. •Tolerance limits •Presence/Absence •Land cover, etc. Inductive Approach Biogeography Links • California Wildlife Habitat Relationships – http://www.dfg.ca.gov/bdb/html/cwhr.html • GAP Analysis – California: http://www.biogeog.ucsb.edu/projects/gap/gap_ proj.html – National: http://gapanalysis.nbii.gov/ http://www.biogeog.ucsb.edu/projects/gap/gap_proj.html Outline • Biogeography Background • WNV Background • My Research – – – Conceptual model Predicting WNV distribution Identify optimal surveillance location West Nile virus • First isolated in Uganda—1937 • First detected in US—1999 • Since spread to entire contiguous 48 states • Infection can cause range of symptoms – Mild: West Nile fever – Severe: Encephalitis and Meningitis • • • • • • 2003: 9862 cases & 264 deaths 2004: 2539 cases & 100 deaths 2005: 3000 cases & 119 deaths 2006: 4219 cases & 161 deaths 2007: 906 cases & 26 deaths 2008: 1370 cases & 37 deaths http://www.cdc.gov/ncidod/dvbid/westnile/index.htm Source: http://www.cdc.gov/ncidod/dvbid/westnile/Mapsactivity/surv&control06Maps.htm http://www.cdc.gov/ncidod/dvbid/westnile/Mapsactivity/surv&control07Maps.htm http://www.cdc.gov/ncidod/dvbid/westnile/Mapsactivity/surv&control08Maps.htm http://www.cdc.gov/ncidod/dvbid/westnile/index.htm Surveillance • Human – Mandatory reporting to CDC – Blood donations – Point of infection difficult to ascertain • Mosquito – Set trap locations – Trap placement important Surveillance • Sentinel chickens – Similar to mosquito – Show seroconversion – Effort > warning • Veterinary – Similar to human surveillance – Mainly equines – Vaccine available http://www.hhs.state.ne.us/wnv/ Surveillance • Dead bird – Good early indicators – Rely on public participation • Find a dead birdCall hotline –1-877-WNV-BIRD – Species and condition important – Volunteered geographic information Know Your WNV Hosts A. B. C. D. E. F. Surveillance Links • California – http://westnile.ca.gov/latest_activity.php • National (CDC) – http://www.cdc.gov/ncidod/dvbid/westnile/inde x.htm • National (USGS) – http://diseasemaps.usgs.gov/ http://westnile.ca.gov/ http://westnile.ca.gov/ http://diseasemaps.usgs.gov/ Why map WNV? • Ultimate goal: limit human infection • Map migration across the country • Identify areas of high risk for mosquito control and health alerts • Determine outbreak patterns – Perennial: Japanese encephalitis – Sporadic: St. Louis encephalitis http://www.cdc.gov/ncidod/dvbid/westnile/index.htm Previous Work • Disease Mapping • Largely descriptive • Little predictive value • No process behind pattern • Geographic Correlation • Sin Nombre—mice • Lyme—ticks • DYCAST • Predict hotspots from dead bird reports • Urban areas http://westnile.ca.gov/2005_maps.htm Outline • Biogeography Background • WNV Background • My Research – – – Conceptual model Predicting WNV distribution Identify optimal surveillance location Research Objectives I. Define a conceptual model for WNV distribution II. Predict WNV distribution from surveillance data and ancillary environmental variables III. Identify optimal areas for additional surveillance sampling I. Conceptual Model • Ecological/biogeographical approach • Mapping WNV as function of pertinent life cycle components • Virus propagation areas – Amplification and transmission – “Reproductive range” • WNV only needs reservoir host (birds) and vector (mosquito) • Human, sentinel, and veterinary instances can be considered sterile I. Habitat suitability models • Based on Hutchinson’s (1957) concept of niche – Hypervolume where species is found within suitable ranges for all variables – Biology reflected in habitat selection – Fundamental niche – Realized niche does not often match fundamental niche • Includes biotic interactions and competitive exclusion • Species is not at equilibrium – Number of multivariate techniques • Probabilistic techniques are similar – High suitability implies high presence probability I. Hosts + Vectors • Suitability/probabilistic techniques classify area for one species • Multiple reservoir and vector species – Need at least one host and one vector • P(H) = P(H1 U H2) = P(H1) + P(H2) – P(H1 ∩ H2) • P(V) = P(V1 U V2) = P(V1) + P(V2) – P(V1 ∩ V2) • P(WNV) = P(H ∩ V) Study Area • W Kern County – W of Sierra Nevada mountains – 10,000 sq. km • • • • Kern Co. MVCD Jepson’s ecoregions Rural & urban 20+ data points for for 2 intermediate hosts and 2 vectors – 2004 season Intermediate Hosts • Corvids most important – Susceptible, conspicuous, recognizable – 8 species within study area • American Crow – Corvus brachyrhynchus – cosmopolitan – Woodlands, grasslands, croplands, and urban areas • Western Scrub-jay – – – – Aphelocoma californicus More selective Woodlands & shrublands—Oak Residential urban areas Vectors • Genus Culex – Permanent water breeders – Bloodfeeding usually close to breeding sites • Culex tarsalis http://www.usask.ca – Western encephalitis mosquito – Irrigation ditches, riparian • Culex pipiens quinquefasciatus http://www.fehd.gov.hk – Southern House mosquito – Urban environments (e.g. sewer catch basins) Variable Selection • 7-10 Eco-geographic variables (EGV) per species – 1 km --- 10,000 pixels • General EGVs – Elevation, Percent Urban, Distance to water • Species Specific – Mosquitoes—hydrographic; Birds—land cover • Neighborhood layers will account for species range size – Mosquitoes—0.03-0.04 sq. km – Jays--0.03 sq km; Crow--0.1-0.5 sq km II. Presence/absence Methods • Presence/absence – Ex. regression – Potentially more predictive power – Reliable absences difficult to obtain for animals • Species is present, but not detected – Imperfect detectability of target species • Species is absent, even though habitat is suitable • Presence-only – Ex. ENFA – Trade off: predictive power vs. unreliable absences II. Bayesian Model • Conditional probabilities of Bayes Theorem • Probability of WNV positive given a series of EGV values • Advantages – Presence-only when EGVs known everywhere – Easy to integrate new presences • Presence and EGV data—rasters – Matlab • For each species, two sets of histograms – Global—EGV values over entire study area – Presence subset—EGV values at presence locations • Multivariate probability density functions II. Bayes Theorem P(H1|EGV) = P(H1) * P(EGV| H1) _______________________________________ P(H1)*P(EGV| H1) P(EGV) + P(absence)*P(EGV|absence) P(H1) P(EGV| H1) Denominator = probability of WNV positive intermediate host of species 1 over the entire study area = probability of EGV value within WNV positive H1 subset = probability of EGV value within global set II. Simulation • Problem: presence subset not exhaustive – P(H1) not known – P(EGV| H1) not fully characterized – Needs to be augmented • Total presence probability P(H1) estimated from focal species range within study area – Maximum WNV dispersal • Additional presences simulated until threshold met – Areas near presences are preferentially weighted – P(EGV| H1) updated • Presence probability map for simulated P(EGV| H1) • Many simulations (n=1000) – Distribution of presence probabilities for each pixel II. Flowchart • Composite probability map for each species Prob. Crow (H1) Prob. Host Prob. Jay (H2) Prob. WNV Prob. Tars. (V1) Prob. Vector Prob. Quin. (V2) • Combined using definition of P(WNV) II. Presence/Absence Maps • Convert WNV probability to binary presence/absence • Receiver Operating Characteristic plots – Determines threshold that most accurately separates 2 classes • Requires validation dataset – Sentinel data • For each threshold, sensitivity-specificity pair is calculated – Sensitivity: true positive fraction • a / (a + c) – 1 – Specificity: false positive fraction • d / (b + d) ROC plots • Tangent line defines optimal sensitivityspecificity pair – Corresponding threshold considered best separation value – Slope can be function of false positive and false negative costs • AUC can be used as index of overall model accuracy • Use threshold to change probability to binary map Fielding and Bell 1997 III. Optimizing Surveillance • Surveillance is expensive – Improve efficiency • Identify areas for additional sampling that provide the most information on virus activity – Improve separation between presence/absence classes • Optimal site – Ambiguous P(WNV)—near threshold • Presence simulated and change quantified – Change in AUC of ROC plot III. Optimizing Surveillance • Optimal sampling strategy – Number and locations of surveillance points • Loss function--monetary • Surveillance costs – Traps, testing, travel • Surveillance benefits – Improved efficiency of mosquito control – Less human cases • Suggestions to Kern Co. MVCD Conclusion • WNV endemic to US – Public health significance • Spatial aspect of WNV is important – Direct surveillance – Direct mitigation • Research Objectives – Provide a model for mapping zoonotic disease – Method of relating presence-only data to EGV – Assessing the value of additional surveillance • Address academic and management issues – GIS works for both Acknowledgements • • • • Dr. Michael Goodchild Dr. Phaedon Kyriakidis Dr. Keith Clarke Dr. Wayne Kramer Questions ??