Early Warning System for West Nile Virus Risk Areas, California, USA

TOC Summary: This system effectively identified high-risk human population areas.


W est Nile virus (WNV; family Flaviviridae, genus
Flavivirus) is a mosquito-borne pathogen that has led to ≈30,000 reported (>325,000 estimated) human cases and 1,172 reported deaths in the United States since it was fi rst detected in New York, New York, in 1999 (1). The virus was fi rst detected in California in a pool of Culex tarsalis mosquitoes in July 2003 (2), and in 2004 and 2005 the state had the highest number of reported human cases (779 and 880, respectively) and deaths (29 and 19, respectively) in the United States (3). Humans are incidental, dead-end hosts of WNV and generally become infected after intense viral amplifi cation and spillover from local avian populations (4). Birds are the natural reservoir and amplifi cation hosts of WNV and infections can cause death rates up to 100% among avian species (5,6). Beginning in 2000, bird carcasses in California were submitted by local agencies to the WNV Dead Bird Surveillance Program (DBSP) at the California Department of Public Health (CDPH; previously known as the California Department of Health Services) as part of the California Mosquito-Borne Virus Surveillance and Response Plan (7,8). A toll-free telephone hotline and website for recording public reports of dead birds was established in 2002.
Previous efforts for the early detection and monitoring of WNV activity have used dead bird density or spatial scan statistic as a proxy for transmission risk for humans (9)(10)(11)(12)(13). However, aggregation of reports over nonuniform spatial units (i.e., counties and census tracts) may fail to detect WNV amplifi cation clusters that span regional boundaries or that are contained within large areas. In addition, temporal aspects of the WNV transmission cycle should be considered to avoid false-positive identifi cations in circumstances in which sustained but slow transmission leads to an accumulation of dead bird reports above the designated risk threshold but does not result in spillover to the human population.

Early Warning System for West Nile Virus Risk Areas, California, USA
Another approach is the DYCAST system (14,15), implemented in New York, New York, in 2001 and Chicago, Illinois, in 2002. This system detects statistically signifi cant spatiotemporal clustering of dead bird reports by modeling the WNV amplifi cation cycle using biological parameters; it also includes a statistical method for evaluating effectiveness of human case predictions in space and time. Results indicated that clusters of dead bird reports and human cases of WNV were signifi cantly associated in space and time (15). This association suggests that this procedure may be useful for predicting areas at high risk for WNV transmission to humans. Because there is no drug prophylaxis, human vaccine, or treatment available for WNV, integrated pest management and personal mosquito protection remain the only options for reducing human illness and death, and early warning of high-risk areas allows for these measures to be implemented in a timely and effective manner. The objective of the present study was to evaluate implementation of DYCAST as an early warning system in California to target public education campaigns, surveillance, and mosquito control efforts during an anticipated statewide outbreak of WNV.

Data
Public reports of dead birds were obtained from the DBSP. Through press releases and various types of media campaigns at state and local levels, citizens were encouraged to use the hotline (1-877-WNV-BIRD) and website (www.westnile.ca.gov) to report dead birds (7). Information regarding location, date found, and species was collected for each dead bird reported to the hotline; multiple dead birds included in a single report were treated as multiple reports. Hotline staff screened and entered these data into an Access database (Microsoft Corporation, Redlands, WA, USA); data were subsequently geocoded by using ArcMap version 9.1 and associated 2005 StreetMap USA Plus AltNames street dataset (Environmental Systems Research Institute, Inc., Redlands, CA, USA). WNV became a reportable disease in California in 2005, and human data were collected by local health departments by standardized case history forms. Data were subsequently stripped of personal identifi ers, and addresses were geocoded by using a CDPH batch geocoding service that used multiple reference databases (www.ehib.org). Use of human data was approved by the institutional review board at the California Health and Human Services Agency (project no. 05-06-51).

Procedure
The DYCAST procedure was implemented by using GIS software, Smallworld 3.2.1, and Magik programming language (General Electric Company, Fairfi eld, CT, USA). Regions comprising 32,517 km 2 among 16 participating agencies in 17 counties were superimposed by grids consisting of ≈0.44 km 2 (≈0.17 mi 2 ) cells ( Figure 1). Clustering of dead bird reports was quantifi ed by using a Knox test (16,17) implemented from the center of individual cells; spatial and temporal parameters were defi ned by using biologically relevant values ( Figure  2). The 2.4-km (1.5-mi) radius of the spatial domain represents 2× the daily feeding distance (14) of Culex spp. mosquitoes in California (18). The effective fl ight range of these mosquitoes is also 2.4 km (19), which corresponds to the maximum distance from breeding sites over which a suffi cient number of vectors are able to disperse a mosquito-borne disease (20). The temporal domain of 21 days was based on a 7-day extrinsic incubation period of WNV, which ranges from 5 to 8 days for Culex spp. mosquitoes at 28ºC in California (21), plus 2 avian infection cycles of 7 days each (approximate maximum time from infection to death ; 5,14,22). Candidate values for defi ning proximity of dead birds in space (0.40, 0.56, and 0.64 km) and time (3,4, and 5 days) were based on the limited mobility (caused by lethargy, ataxia, and reluctance to fl y) and lifespan of infected amplifi cation hosts (5,14,23). During model calibration, locations of human cases were compared with DYCAST risk maps generated by using various combinations of candidate values; 0.40 km (0.25 mi) and 3 days were selected as the optimal combination for the fi nal model. For this calibration, the daily DYCAST procedure was run retrospectively (once during May 2005) by using dead bird and human case data from May 1 through September 30, 2004 within Los Angeles, Orange, Riverside, and San Bernardino Counties, which contained 664 (85.2%) of 779 statewide cases in 2004.
The DYCAST procedure was run at the center of every cell for which a minimum of 15 birds (the analysis threshold) was reported within the spatiotemporal domain, to minimize statistical instability that otherwise occurs at lower numbers of birds (14). Clustering was evaluated by comparing the observed number of pairs of dead birds that were close in both space and time (based on aforementioned values of proximity), with the expected number of pairs given a random spatiotemporal distribution of these reports (15). The resulting p values were assigned to individual cells, which were considered to indicate high risk for WNV transmission to humans at p<0.1 (15).

Evaluation
Model evaluation was conducted by analyzing the relationship between the location of human cases and the ability of DYCAST to predict their occurrence in both space and time. Prediction was defi ned as the identifi cation of a cell as high risk before or on the date of illness onset (15) of the earliest case located within a cell. Sensitivity was calculated as the number of high-risk cells classifi ed as predicted (true positives) divided by the total number of cells in which a human case occurred. Specifi city was calculated as the proportion of low-risk cells without cases (true negatives) to the total number of cells without cases. Because agreement between model predictions and cases can occur by chance, a spatiotemporal implementation of the κ statistic was used to provide a measure of chanceadjusted agreement (15,24).

Implementation
An initial pilot phase and subsequent prospective implementation occurred through a cooperative agreement with the Center for Advanced Research of Spatial Information at Hunter College, City University of New York. The Center for Vectorborne Diseases (CVEC) at the University of California Davis provided server infrastructure (Microsoft SQL Server, Microsoft Corporation; ArcIMS, Environmental Systems Research Institute, Inc.) for data exchange and implementation of interactive online risk maps, in collaboration with CDPH and the Mosquito and Vector Control Association of California. The Center for Advanced Research of Spatial Information calibrated and ran the DYCAST procedure and exported data to the CVEC map server. During the pilot phase, animations of daily risk from June 1 through June 23, 2005, were retrospectively generated for 3 study areas that were selected based on high numbers of dead bird reports: the south Sacramento Valley region (Sacramento, Placer, and Yolo Counties), the central San Joaquin Valley region (Fresno, Kings, and Tulare Counties), and the greater Los Angeles area.
Prospective modeling began on June 17, and on July 1 the system was fully implemented and integrated into the CDPH WNV Surveillance Program. This implementation involved running the DYCAST procedure for analysis regions every weekday through November 1, 2005; daily risk maps ( Figure 3) were generated and made available in real time to mosquito control agencies via the CVEC password-protected website, the California Vectorborne Disease Surveillance Gateway (www.calsurv.org). These interactive maps were overlaid with county boundaries,  (16,17) implemented at the center of an individual ≈0.44 km 2 grid cell. The 2.4-km (1.5-mi) radius of the spatial domain represents twice the daily feeding distance (14) of Culex spp. mosquitoes in California (18) and is equivalent to the effective fl ight range of these vectors (19,20). The 21-day temporal domain accounts for the extrinsic incubation period of West Nile virus (21) and 2 avian infection cycles of 7 days each (5,14,22). These bounds defi ne the spatiotemporal domain, within which reports of dead birds (asterisks) are evaluated for proximity in space (0.40 km) and time (3 days) (small white cylinder). Statistical signifi cance of dead bird report pairing is assessed by using random simulations (p<0.1) (15). Procedure is repeated at other cell centers to create a continuous surface of risk.
streets, and locations of reported and WNV-positive dead birds.

Implementation 2006-2009
Beginning in 2006, DYCAST was implemented for the entire state of California and adopted as a formal component of the California Mosquito-Borne Virus Surveillance and Response Plan (8). Addresses of where dead birds were found were automatically geocoded in real time by using the Yahoo Maps application programming interface (Yahoo! Inc., Sunnyvale, CA, USA), which allowed hotline staff to validate location data while callers remained on the line; birds not automatically geocoded were omitted from DYCAST analysis. Interactive DYCAST risk maps were made available online to local mosquito control agencies and integrated with dead bird, mosquito, and sentinel chicken surveillance data from May 1 through October 31, 2006, May 1 through August 31, 2007 and 2008, and March 1 through August 31, 2009. Statewide reports of DYCAST activity, including maps and animations of high-risk areas over time, were sent to local agencies on a routine basis. A real-time alert system was also introduced in 2006 to provide custom DYCAST reports and interpretations for counties experiencing rapidly increasing or elevated levels of high-risk areas (8).
In December of 2006 and 2007, links to web-based surveys regarding the DBSP were provided by email to 64 local mosquito control agencies in 47 counties, in part to assess which agencies used DYCAST to assist mosquito larviciding or adulticiding activities each year. For agencies that participated in the 2005 DYCAST program, the 2006 survey also asked if DYCAST results were used "to assist public education or to promote dead bird reporting" in 2005 (control activities were not surveyed for this year). Rate ratios (RRs) were used to compare annual DYCAST prediction rates of reported human WNV cases between  agencies that did and did not use DYCAST to assist each mosquito control activity (25).

Results
During 2005, a total of 124,876 calls were placed to the DBSP hotline, >3 million hits were made to the website, and 109,358 dead birds were reported in California ( for >4 weeks before onset of illness ( Figure 4). Overall, 289/354 (81.6%) of cases were predicted (Table 3), with cells identifi ed as high risk before onset of illness by a mean of 37.2 days (range 0-126, median 34, SD 20.9 days). A total of 252/354 (71.2%) of cases were predicted 15 days before onset of illness, and >50% of cases (179/354, 50.6%) were predicted 30 days before onset of illness.

Discussion
Results from prospective implementation of the DYCAST system in California indicate that the risk model provided accurate and early identifi cation of areas at high risk for WNV transmission to humans during a statewide epidemic in 2005, and was used by local agencies to assist public education campaigns, surveillance, and mosquito control programs. Our fi ndings indicate that DYCAST yielded high levels of sensitivity and specifi city for predicting human cases and that relative risk for a WNV case was >39× higher in high-risk cells than in low-risk cells (this value should be considered somewhat infl ated, however, because not all low-risk cells contained populated areas). Given the low prevalence of cells containing cases (0.45%), the dynamic nature of DYCAST, and the (>1 cell) spatial scale of WNV transmission and mosquito control (8), positive predictive value is considered inferior to other metrics such as κ for evaluating model predictions. κ values >0.50 indicate that DYCAST correctly identifi ed >50% of cells expected to be misidentifi ed by chance alone, which is considered high because WNV causes symptoms in only ≈20% of infections (28). Values maintained a moderate strength of chance-adjusted agreement for >4 weeks before onset of illness, which indicates temporal robustness of model predictions.
Cells containing predicted cases were identifi ed as high risk before onset of illness by a mean of 37.2 days; given the 2-14 day range of the human WNV incubation period (28), this identifi cation preceded transmission to humans and provided suffi cient time to respond and potentially reduce the number of infections (Figures 4, 5). Indeed, 252/354 (71.2%) of cases were predicted 15 days before onset of illness, before the maximum range of the incubation period. Additionally, because the DYCAST procedure only analyzes dead bird reports, it provided for more timely results than did active systems relying on the collection and testing of bird carcasses.
Results from Sacramento County in 2005 demonstrate the practical application of DYCAST for conserving and directing public health resources, such as targeting surveillance efforts that detected the county's fi rst positive mosquito pools that year. During subsequent months, Sacramento County was the location of the largest WNV epidemic in the United States, with 163 reported human cases (30) and an incidence rate of 14.5 infections per 100,000 population (31). DYCAST results played a key role in SYMVCD's decisions for implementing and targeting emergency aerial mosquito control in the county (D. Brown, pers. comm.; 31), which ultimately reduced human illness and potential death from WNV infection (32).
Notably, prediction rates during 2006-2009 were substantially lower than in 2005, which has implications for the robustness of the model in nonepidemic years or regions. The fairly prevalent use of DYCAST results to assist mosquito control activities in 2006 and 2007 may have played a role in reducing the model's prediction rates in circumstances in which WNV transmission was successfully interrupted before human infection occurred (31,32). However, while DYCAST could have helped to reduce the absolute number of cases, relatively higher prediction rates were generally observed for agencies that used DYCAST results compared with agencies that did not (Table 4). One explanation is that these areas may have had higher rates of WNV transmission initially, which in turn may have increased agencies' likelihood of utilizing DYCAST for directing control activities or of simply conducting control activities in general. Furthermore, higher rates of WNV transmission may yield greater numbers of subsequent cases within high-risk cells or clusters, thereby increasing the prediction rate. This phenomenon could have also contributed to the higher prediction rates observed during the more severe epidemic in 2005, as could have the self-selecting nature of agencies that participated in the DYCAST program that year, which may have included areas with higher rates of WNV transmission compared with the rest of the state or to subsequent years.
Effi cacy and sustainability of the DYCAST system may be compromised by declines in dead bird reporting, which could be caused by public fatigue or apathy, reductions in reporting infrastructure, or declines in bird deaths caused by herd immunity (33). Potential approaches for ameliorating these effects could include recalibration of DYCAST parameters (e.g., lowering the analysis threshold), strategic timing and targeting of press releases and media campaigns, and technologic solutions such as mobile phone application software and text messaging to disseminate information and facilitate the reporting of dead birds. Furthermore, it is uncertain how DYCAST results are affected by spatial and temporal heterogeneities of WNV transmission, including inter-and intraspecies variability in the competence (21,34), pathology (6), and distribution of vector and host populations (35,36). Other confounding factors may include demographic and socioeconomic composition of human populations (37) as well as environmental (38) and meteorologic variation. Regardless, DYCAST proved to be a timely and effective early warning system during a severe WNV epidemic. The use of such prospective measures enable the conservation and focus of valuable human and fi nancial resources, which in some cases could be the difference in making an otherwise chaotic epidemic situation tractable. More responsive and effi cient surveillance and control can prevent additional human disease, decrease reliance on more substantial control activities later in the season, and reduce indirect costs from medical expenses and productivity loss. The total cost of the 2005 WNV epidemic in Sacramento County alone has been estimated at ≈$3 million (39). Furthermore, dynamic monitoring of risk throughout the season may inform decisions for redirecting and triaging resources and may also provide a means for evaluating effi cacy of mosquito control efforts. Ultimately, the DYCAST system illustrates the utility of establishing a biologically relevant, spatiotemporal framework for disease surveillance, and adaptation of the DYCAST method may be useful for detecting other infectious diseases and clustering phenomena.
This study also highlights the benefi ts of interdisciplinary and interagency collaboration; synergies between 2 academic institutions and a governmental public health agency shortened the time from research to implementation, and engagement with local mosquito control agencies enabled the practical application of results in real time. Furthermore, our fi ndings demonstrate the potential of harnessing the public's ability to provide timely and useful surveillance data through telephone and internet communications. The leveraging of similar sociotechnologic infrastructure, from mobile phones to internet search queries and social networks, may play a major role in the success, scalability, and cost-effectiveness of predicting and preventing emerging diseases in the future. *DYCAST, Dynamic Continuous-Area Space-Time; CI, confidence interval; NA, not applicable. Agencies were asked whether they used DYCAST to assist larviciding, adulticiding; agencies that did not respond to survey, as well as answers of "Don't know"