A Valley Fever (Coccidioidomycosis)

[sunny]

National Center For Environmental Research

Final Report: A Valley Fever (Coccidioidomycosis) Public Health Decision Support System Based on Climate and Environmental Changes

EPA Grant Number: R832754
Title: A Valley Fever (Coccidioidomycosis) Public Health Decision Support System Based on Climate and Environmental Changes
Investigators: Comrie, Andrew C. , Yool, Stephen R.
Institution: University of Arizona
EPA Project Officer: Bloomer, Bryan
Project Period: December 14, 2005 through December 13, 2007 (Extended to December 13, 2009)
Project Amount: $265,004
RFA: Decision Support Systems Involving Climate Change and Public Health (2005)
Research Category: Global Climate Change , Health Effects

Description:
Objective:

Valley fever (coccidioidomycosis) is a disease endemic to arid regions in the Western Hemisphere, and is caused by the soil-dwelling fungi Coccidioides immitis and Coccidioides posadasii. Arizona is currently experiencing an epidemic with almost 4000 cases in 2004, greatly exceeding other climate-related diseases such hantavirus or West Nile Virus. Previous research has indicated relationships linking temperature and precipitation to outbreaks of coccidioidomycosis. Our preliminary statistical models can explain up to 80 percent of the seasonal variability in coccidioidomycosis based on antecedent precipitation and atmospheric dust concentrations. In this new research we sought to further our understanding of climaterelated factors and their role in coccidioidomycosis outbreaks in Arizona. Through collaboration with the Arizona Department of Health Services we proposed to put in place a set of climate-based seasonal models and associated databases that would enable anticipation of coccidioidomycosis outbreaks and improved public health actions to mitigate them. Specific components include temporal and spatial information on historic and forecast coccidioidomycosis incidence in tandem with the key climatic and other environmental variables that correlate to disease outbreaks.

In this study, we have used Arizona coccidioidomycosis case data for 1995-2006 to investigate temporal trends and variability in case rates, and to identify relationships between climate variables using correlation and regression techniques. Our latest research findings have confirmed and refined our understanding of linkages between climate and environmental conditions and seasonal coccidioidomycosis changes. Information generated through our research efforts has become an integral part of the public health decision-support knowledge base for coccidioidomycosis in Arizona.

[sunny]

Summary/Accomplishments (Outputs/Outcomes):

Project Activities
Year 1
We focused our efforts on developing and extending our existing data for our modeling and environmental studies. We also renewed our collaboration with ADHS to secure human case data for model validation. Our objective was to develop initial climate-satellite-derived data correspondences, and perform initial multi-layer, multi-scale remote sensing based spatial analysis of the surface landscape. The database developed would serve as a practical reference for health officials and as a model for health information support services in general.

Climate Modeling:
We analyzed the sensitivity of the seasonal modeling approach of Comrie (2005) using different quality control and adjustment procedures, evaluating model performance on coccidioidomycosis incidence data updated through 2005 (Comrie and Glueck, 2007). Issues addressed: (i) the role of quality-controlled case report data versus “raw” data; (ii) model performance using simple aggregate monthly incidence instead of exposure adjusted incidence data; and (iii) climate variability-detrended incidence relationships. Close relationships identified between climate conditions and coccidioidomycosis incidence further support the broader “grow and blow” hypothesis of regional climate influence on coccidioidomycosis

We found that models using case-level data adjustment do not suffer significantly if individual case report data are used “as is.” Eliminating the data quality control step (removal of the small % duplicates and other inconsistencies) speeds up processing with little impact on model performance. Further use of unadjusted “raw” monthly reports did not reduce important explanatory power of the model. Several predictor variables matched those of Comrie (2005). Use of the identical modeling approach (seasonal models with aggregate monthly reports) with detrended incidence data (long-term trend removed) resulted in models that explained a large portion of the remaining seasonal and annual variability. Antecedent fall precipitation appears to have some long-term influence on detrended incidence indicating that climate explains a large proportion of variability in coccidioidomycosis incidence. However, inclusion of updated data (to 2005) revealed that our models were unable to capture the recent overall upward trend in coccidioidomycosis incidence, despite being able to capture much of the seasonal and annual variability. Whether the longer trend was due to a non-climate factor (e.g., urban expansion, improved reporting, etc.) or some decadal-scale climate response could not be immediately resolved using these short periods of data.

Remote Sensing. Surface, and Exposure Modeling

The spatial generalization of the climate model to the three-county area of central Arizona (Pima, Maricopa, Pinal) required satellite data that tracked climate-related conditions over this large region. The MODIS satellite data used in this project were assessed for scalability as part of the model-building process. We examined NDVI (Normalized Difference Vegetation Index) data at two different resolutions in eastern Pima County, Arizona. Assessment of NDVI condition and change was thought to be an important part of landscape-scale monitoring of conditions that contribute to Valley Fever occurrence, for example in mapping inferred soil moisture variations over large areas (http://vf.grd.arizona.edu/findings).

We compared the same ground area imaged at 250 m resolution and at 1 km resolution. Results show that the average of the 250 m contributing pixels was essentially identical to the 1 km NDVI value (R2 = 0.96), confirming that the NDVI signal could be rescaled linearly, an important consideration when monitoring and modeling environmentally-mediated diseases using multiple scales.

The linkage between climate and surface models required understanding of how climate processes affect the satellite data used in the surface model. Therefore, we examined the relationship between precipitation and MODIS NDVI data measuring vegetation spectral response in eastern Pima County, Arizona, a “hot spot” for Valley Fever. Data spanned the period between March 2000 and December 2005. Monthly precipitation totals were adjusted to match the 16-day period covered by each NDVI image. Total precipitation and average NDVI were computed for seasons matching the climate-cocci model, and correlation coefficients were calculated. Results showed that seasonal mean NDVI could be predicted from antecedent precipitation with R2 values ranging between 0.99 (winter NDVI correlated with winter precipitation from two years previously) and 0.57 (monsoon NDVI correlated with precipitation from arid foresummer two years previously). Correlations for fall and arid foresummer fell between these two values. Seasonal variations of NDVI response were determined to be most likely driven by different vegetation types that reach their peak growth at different seasons with different response lags to precipitation.

[sunny]

Year 2
We modeled exposure dates, continued development of remote sensing databases related to surface moisture, fungal ecology and dust exposure. Remote sensing protocols for moisture modeling, ecological modeling and exposure modeling were also extended.

Climate Modeling
In Year 2, we carried out extensive bivariate and multivariate analyses to complete a thorough evaluation of potential predictors for detrended and variance-corrected incidence data. Interim results appeared to confirm that climate and associated fluctuations in seasonal and year-to-year environmental factors account for as much as 80 percent of the coccidioidomycosis incidence variability about the trend. Sensitivity analyses of the seasonal modeling approach of Comrie (2005) continued using different quality control and adjustment procedures, evaluating model performance on coccidioidomycosis incidence data updated through 2005 (Comrie and Glueck, 2007). We also used zip code level data from ADHS and developed a strategy for modeling exposure dates. Relationships identified between climate conditions and coccidioidomycosis incidence continued to support the broader “grow and blow” hypothesis of regional climate influence on Valley Fever.

Remote Sensing: Surface, Ecological, and Exposure Modeling
We used the AVHRR NDVI database assembled in Year 1 to compare NDVI time series with Valley Fever incidence--testing hypotheses linking soil moisture to fungal growth. The recent strong upward trend in incidence was not reflected in modeled soil moistures, consistent with other findings that the cause of the trend is non-climatic. Seasonal patterns of the disease and NDVI in all three counties display an apparent inverse bimodal relationship consistent with fungal spore dispersion in the dry seasons, but are not significant from year to year. Results of lagged regression modeling showed that moist soil in the early spring, resulting from winter precipitation, is significantly associated with increased incidence up to a year later in all three counties (Stacy et al. in review). We selected the Normalized Difference Vegetation Index (NDVI) as a proxy for surface moisture based on results from prior work.

We compiled a soils database to consider the use of MODIS data as a proxy for soil habitat modeling of Coccidioides. Because Coccidioides favors organic soils and (like hantavirus) may therefore associated with the presence of vegetation and rodents, we deployed a MODIS Organic Matter Index (OMI). Preliminary comparison of MODIS OMI data with a rodent field census (a separate project supported through the NASA Space Grant Internship Program) suggested that MODIS OMI may be sensitive to soil structural characteristics (e.g., hydraulic conductivity) thought to be favorable to rodents and Coccidioides.

Development of an exposure model was undertaken by a Ph.D. candidate external to this EPA grant. The objective was to identify areas of land surface disturbance as potential fungal spore dispersion regions. The Lower Sonoran Life Zone (LSLZ), a zone considered by mycologists to be prime habitat for Coccidioides, was masked. Change detection techniques were applied to six non-thermal bands of Landsat Thematic Mapper satellite images over the period of 2000 to 2005 in eastern Pima County, identifying locations and extent of human-related disturbance. Dust inspection permit data from Pima County were used to validate the change detection. Good cross-correlation was evident between the Thematic Mapper change detection protocols and dust inspection data indicating that Landsat TM could be used to extend the series back in time (mid- 1980’s).

[sunny]

Year 3 and Beyond
In year 3 and our no-cost extension period, we continued to refine our modeling efforts through refinement of computational methods, extension/addition of data and sensitivity analyses.

Climate Modeling
We conducted numerous univariate and bivariate lag analyses of case rates in Maricopa and Pima counties in 2008 and 2009. Precipitation from several stations in Pima and Maricopa were used in preference to the AZMET precipitation data because a greater number of stations were available. The case-data received from ADHS has a strong trend that we now conclude is an artifact of enhanced surveillance. To perform analyses on the data, we removed the trend and equalized the variance across time. Several epidemics were apparent in the corrected time-series of case rates. Both counties showed strong peaks in late-Fall, and Pima has an additional peak in late-Spring. Statistical analyses demonstrated strong-positive correlations between case-rates in late-Winter through early-Spring. The case rates during this period were also modulated by precipitation and humidity, supporting the “blow and grow” hypothesis. No relationship was identified between the case rates during the Spring months and early Summer. It is unknown what factors are related to this characteristic, but it occurs near the beginning of the monsoon season suggesting that precipitation may reduce the likelihood of infection during this time of year, also supporting the “grow and blow” hypothesis.

Remote Sensing. Surface, Ecological, and Exposure Modeling
We extended remote sensing protocols for moisture modeling, ecological modeling and exposure modeling. We used the Landsat Thematic Mapper data to characterize temporal trends in surface moisture at comparatively fine scales. We believe the operational scale (i.e., the scale of action) of Coccidioides is likely quite fine, and that these fine-scale spectral data will enable us to investigate the ‘grow’ scale of the fungus closer to its likely scale of action. The nominal 30m spatial resolution of Landsat provides higher quality areal database of surface moisture variability than possible from the 1km AVHRR. Landsat band 5 (i.e., the ‘water’ band: 1.55-1.75 μm) was used to track temporal trends in surface moisture and more readily determine correlations between surface moisture and disease incidence.

We acquired 90 free LANDSAT TM images of Maricopa (Phoenix), Pima (Tucson), and Pinal (Casa Grande) counties from U.S. Geological Survey to establish time series required to characterize relationships between fugitive dust production and incidence (Ph.D. dissertation work of F. Scott Pianalto) for the three scenes (p36 r37, p36 r38, and p37 r37) from 1984 to 2008 of the late spring (late May to early June). Subsets of the areas common to all images from the three scenes were radiometrically and atmospherically corrected. A band differencing technique (Landsat TM Band 5; 1.55 through 1.75 micrometers, mid-infrared spectral region) was investigated and implemented due to its simplicity and superior results achieved over the Principal Component Analysis technique. The band differencing change detection technique was tested with the Pima County satellite data set. Spatial change signal generated from the satellite data was compared with Pima County dust inspection permit data, consisting of spatial data for grading, trenching, road construction and subdivision development activities occurring in the county, to validate the change detection method. Results indicated good cross-correlation between the Thematic Mapper change detection protocol and dust inspection data. Images and results were added to our decision support database, a resource that will eventually include surface disturbance, soil moisture, soil organic content, soil temperature, and other ecological parameters.

[sunny]

Summary of Project Findings

We have been developing under the present agreement, and in collaboration with the ADHS, a public health decision-support database that builds on these findings. The database will serve as a practical reference for health officials and as a model for health information support systems in general. In Year 1, we worked with ADHS to secure human case data for model validation, developed initial climate-satellite-derived data relationships, and performed initial multi-layer, multi-scale remote sensing based spatial analysis of the surface landscape. In Year 2, we modeled exposure dates, continued development of remote sensing databases related to surface moisture, fungal ecology and dust exposure. In Year 3 and beyond, we continued to develop and validate models begun in Years 1 and 2, working towards a multi-disciplinary suite of models to be included in our decision support database.

Climate Modeling
We have developed extensive data quality controls and pre-processing techniques for effective evaluation of a set of climate variables against coccidioidomycosis case rate data for Pima and Maricopa counties, 1995-2006. Climate variables included temperature (min, max, mean), humidity (min, max, mean), precipitation, wind speed (mean, max) & direction, vapor pressure deficit, solar radiation, and dust (PM-10). We performed data reduction via principal components analysis of the climate data, and obtained two key underlying components reflecting (1) moisture/precipitation and (2) temperature. Leading variables in these two components were employed in a lag correlation analysis of case rates adjusted for exposure date and for trend and variance corrections resulting from data recording issues. Analyses (Figure 1) confirmed the utility of seasonal precipitation as the key driving variable for coccidioidomycosis cases in Arizona. In particular, a strong, simple and consistent climate effect emerges that is consistent with the “grow & blow” hypothesis. Results indicate a seasonal autocorrelation structure in case rates for Arizona that may be related to the ecology of the fungus, or an artifact of differential diagnosis and reporting lags. In the growth phase, October through December precipitation is significantly linked to coccidioidomycosis 6-18 months later. Regression analysis indicates that October-December precipitation is positively associated with case-rates the following fall and winter in both Maricopa County (R2 = 0.52, p = 0.013) and Pima County (R2 = 0.48, p = 0.019). High or low precipitation in this period leads to corresponding high or low case rates across most seasons in a lagged time window centered on a year later (Figure 2). Fall and winter case rates are negatively associated with concurrent precipitation in Maricopa (R2 = 0.69, p = 0.002) and Pima (R2 = 0.46, p = 0.02), possibly due to dust inhibition. In the dispersion phase, concurrent monthly precipitation is negatively correlated with coccidioidomycosis. Thus, a wet or dry month leads to corresponding low or high case rates at the same time, presumably because of soil wetting reducing dust and dispersion. A strong negative correlation between precipitation in successive falls and winters during the study period still makes it somewhat difficult to assess the independence of these associations.



Figure 1: (top) Lag correlation plot for Maricopa indicating the correlations between all month pairs. The colors indicate the Pearson’s correlation coefficient (R). The black-contours indicates p=0.01. The plot indicates strong positive correlations between case-rates from September-July. August is not strongly correlated with any of the previous 12-months. (bottom) Pima County. Positive correlations are experienced between September-March. However, the correlations are not as strong as they are in Maricopa. (Tamerius et al., in prep).

[sunny]

Figure 2: (top) Scatter plot displaying the association between September-May precipitation and case-rates of the sameperiod. It indicates that greater amounts of precipitation are negatively associated with case-rates in both Maricopa and Pima. (bottom) Scatter plot displaying the association between October-December precipitation and average exposurerates for the following September-March. (Tamerius et al., in prep.)

Climate Model Details
The climate-based model of coccidioidomycosis incidence refined herein was based on the multivariate regression model described in Comrie (2005). Primary predictor variables included monthly precipitation and PM10 dust data for Pima, Pinal, and Maricopa County. For our study here, case data were obtained from the Arizona Department of Health Services (ADHS) for the years 1995-2006. Incidence data was aggregated by county, but inadequate numbers of cases in most counties restricted this study to the populous Maricopa and Pima counties. Case data that did not indicate a diagnosis date were removed. A total of 18,954 cases in Maricopa County, Arizona and 4,645 cases in Pima County, Arizona were present during the study period. Exposure date was estimated for each case using information from an enhanced surveillance study conducted by ADHS which indicated that the median time from the onset of symptoms and the diagnosis date was 54 days. Thus, 54 days were subtracted from the diagnosis date of each case to estimate the time of symptom onset. In addition, 14 days were subtracted to account for the incubation period, producing a time-series of estimated exposure. Exposure dates were aggregated into monthly totals. Case rates per 100,000 population were determined by dividing through by linearly interpolated monthly population estimates provided by the US Census Bureau. The data were detrended by subtracting the best-fit line through the monthly exposure ofeach county.

Daily climate data were retrieved from the Arizona Meteorological Network (AZMET) for three stations in Maricopa County and a single station in Pima County. Variables included temperature (min, max, mean), humidity (min, max, mean), mean wind speed, max wind speed, mean wind vector, soil temperature (2 and 4-inch depths), vapour pressure deficit, precipitation, solar. Full details of the climate-based model of incidence can be found in listings under Publications and Presentations.

Remote Sensing. Surface, Ecological, and Exposure Modeling
We compiled from the free Landsat Thematic Mapper imagery the timeseries required to characterize relationships between fugitive dust production and incidence (Ph.D. dissertation work of F. Scott Pianalto). We have gained access to zip-code level incidence data that, when upscaled to county level, provide the reference against which we will validate exposure and ecological databases. We have been developing ecological databases using the Landsat data archive. These ecological databases include space-time variations in moisture and greenness. We have added soils data to our decision support database that will characterize differences in soil organic content related to Coccidioides habitat requirements.

Conclusions:

Our present modeling and data analyses results have further refined and confirmed that seasonal climate fluctuations, especially Fall-Winter precipitation, explain much of the coccidioidomycosis case rate variability about the trend from 1995-2006. Future efforts will focus on understanding spatial disturbance on finer temporal and spatial scales. This will require the incorporation of additional environmental data as well as higher quality incidence data. Formal field work will be required to better understand inherent disease ecology-incidence connections for the Arizona region. As appropriate new data (remote-sensing ground surface proxy and field collection) are acquired and analyzed, our attention will shift to development of a dynamic climate model of incidence. This type of model will allow us to better investigate lags in a spatial domain, thereby, making it possible to attempt coccidioidomycosis potential mapping.

Collaborations Years 1-3
USA: We are continuing collaborations with stakeholders at the ADHS to put in place a set of seasonal models and associated databases that will enable anticipation of coccidioidomycosis outbreaks and improved public health actions to mitigate them. We have maintained contact with Arizona State Epidemiologist Ken Komatsu, securing release of human incidence data for model validation and pre-processing released data to suit our project needs. We have also collaborated with Rebecca Sunenshine of CDC/ADHS (now at Maricopa County Health Dept.) with regards to the accuracy of the case data. “Grand Rounds” type research meetings were held once a year to bring researchers, UA colleagues, and ADHS personnel together for collaborative reviews of related project findings. Contacts at Pima, Maricopa, and Pinal County made PM10 available to our project at no cost. We have also engaged in discussions with numerous regional U.S. stakeholders through outreach and professional presentations.

International: We have engaged in discussions with international colleagues at University of Adelaide through an Invited Colloquium presentation (Yool).

Data Quality Assurance
We are confident we conform to 40 C.F.R. 30.54. Climate and satellite data derive from government sources. Satellite data are calibrated to achieve maximum accuracy in space and time, thus are optimal for environmental measurement and data generation. Human disease case data have been processed into GIS format following formal release by ADHS. We gained release of these human case data in conformance with federal human subjects guidelines through formal project review approval of University of Arizona Institutional Review Board and a similar body at Arizona Department of Health Services.