Predictive Modeling for West Nile Virus and Mosquito Surveillance in Lubbock, Texas

West Nile virus (WNV), a member of the family Flaviviridae, is the causative agent of West Nile fever and West Nile neuroinvasive disease. West Nile virus, a globally dispersed zoonotic pathogen, was first detected in the USA during 1999 and has since spread throughout the contiguous USA (Nash et al. 2001, Kunkel et al. 2006). Relatively quickly, WNV emerged as a significant public health threat in the USA (Kunkel et al. 2006, Kilpatrick et al. 2007, Rossi et al. 2010) and continues to be monitored by local, state, and federal agencies.

Transmission of WNV to humans is most commonly through the bite of an infective mosquito (OIE 2013, CDC 2017) and in the southern High Plains of Texas is primarily transmitted by Culex tarsalis Coq. and Culex quinquefasciatus Say.

Vector control programs (VCPs), commonly associated with local and regional governmental public health entities, monitor and control mosquito populations to reduce nuisance populations and potential public health threats that mosquitoes pose as vectors of disease. Oftentimes VCPs do not have the needed information or data relative to mosquito abundance. As a result, often outbreaks of mosquito-borne disease cannot effectively be managed within those jurisdictions (Winters et al. 2008). Treatment of mosquito populations varies by jurisdiction and individual VCPs (i.e., some VCPs are more proactive with chemical mosquito control approaches than others). For instance, in Lubbock, TX, the city-based VCP faces both nuisance and disease risk from the local mosquito populations, with the latter issue historically of greater concern (Ward 1964).

Within the city of Lubbock, information on drivers of risk for WNV and other arboviruses of concern has been limited, thus mosquito control operations are typically based on complaints from citizens and systematic larval sampling results. Additionally, “blanket” spraying of the city during the assumed peak mosquito season has often occurred to mitigate perceived risks. However, since higher mosquito populations are not necessarily proportional to higher mosquito-borne disease risk, this approach to treatment can lead to wasted resources and man-hours as well as negative ecological affects from overspraying. This inefficiency is important because VCPs are commonly restricted in their capacity due to limited availability of resources. Because the potential for outbreaks of arboviral human illness due to unmanaged mosquito populations weighs heavily in VCP decision making, tools that aid VCP decision making regarding when to conduct mosquito control operations based on available data would be of high value.

The overall goal of this effort was to provide information for facilitating informed mosquito management practices by VCPs in the southern High Plains of Texas during conditions when WNV is most likely to occur in vector populations. Our objectives included 1) developing a user-friendly, spatially implicit statistical model that uses weather information to estimate the probability of detecting mosquitoes that are positive for WNV infections within local mosquito populations and 2) implementing a model in coordination with the Lubbock County VCP to estimate the most practical time to focus resources on mosquito abatement efforts. The model developed during this project will improve our understanding of certain weather variables that drive the probability of viral infectivity in Lubbock County mosquitoes. This project may serve as a guide for developing future predictive statistical models in areas with similar climatic conditions, with additional mosquito species, and for other arboviruses. Minimal knowledge exists on mosquito/climate dynamics in western Texas and our findings provide a preliminary statistical model using survey data to predict the probability of WNV-infected mosquitoes in the southern High Plains of Texas.

As part of an ongoing seasonal WNV vector surveillance program, we trapped mosquitoes from in and around the city of Lubbock, TX, using CO₂-baited encephalitic vector survey (EVS) traps (Fig. 1). Mosquitoes were also provided by the Lubbock County VCP. Traps were deployed once or twice each week from April to November, and were typically operated from approximately 1 h prior to sunset to approximately 1 h after sunset. From 2009 through 2017, traps were operated a total of 1,748 trap nights, over a period of 219 wk, and ranged from 2 to 10 trapping sites per week. More than 101,000 mosquitoes were collected, of which 37,904 were tested for WNV. Each week, female mosquitoes were pooled by species and trap site, and a reverse transcription–polymerase chain reaction (RT-PCR) assay was utilized to determine WNV infection in select mosquito species: Cx. tarsalis and Cx. quinquefasciatus, random pools of Aedes vexans (Meigen), and any unidentifiable mosquitoes (based on mangled specimens or missing body parts).

Pooled mosquitoes were lysed using a 1× sodium chloride-Tris-EDTA (STE) grinding buffer (1:50 proteinase K:1× STE) and ground using a motor pestle. Samples were centrifuged and 125 μl of the supernatant was transferred into a tube containing 1 ml TRI reagent®, vortexed, and incubated at room temperature for 10 min. Chloroform (200 μl) was added, then samples were incubated at room temperature for an additional 8 min. Samples were centrifuged and 300 μl of the aqueous clear top layer was transferred into a tube containing 500 μl of 2-propanol and incubated at room temperature for 8 min. Samples were centrifuged and the 2-propanol was decanted, leaving an RNA pellet in the tube. Samples were washed by adding 1 ml of 75% ethanol, centrifuging, and decanting. Samples were dried of any remaining ethanol, and the RNA pellet was resuspended with 100 μl Tris-EDTA (TE) buffer.

Each RT-PCR assay was performed using positive (WNV RNA extract obtained from the Arbovirus Research Laboratory, The University of Texas Medical Branch, Galveston, TX), negative (WNV RNA template, no reverse transcriptase enzyme), and blank (water) controls. Samples were initially tested using Flaviviridae consensus primers (mFU1: 5′-TAC AAC ATG ATG GGA AAG CGA GAG AAA AA-3′; CFD2: 5′-GTG TCC CAG CCG GCG GTG TCA TCA GC-3′ [Chien et al. 2006]), and those determined positive for Flaviviridae were further tested for WNV using specific primers (FP3: 5′-TTG TGT TGG CTC TCT TGG CGT TCT T′3′; RF4: 5′-CAG CCG ACA GCA CTG GAC ATT CAT A-3′ [Lanciotti et al. 2000]). Reverse transcription-PCR products were separated using electrophoresis and a 2% agarose gel. Image capture of the agarose gel was accomplished using an Alpha Innotech FluorChem system (ProteinSimple, San Jose, CA).

Samples were considered positive for WNV if both sets of primers were reactive with the samples. Mosquitoes were determined positive for WNV during 16.9% (37/219 wk) of the duration of the project (Table 1).

Weather variables were selected based on ease of data collection (Table 2), as well as for their potential influence on mosquito biology and WNV prevalence. Weather data were acquired from Weather Underground (www.wunderground.com). All weather data were pooled by variable, and averaged by week. Due to uncertainty of weather factor influence on probabilities of infection over time, we included for evaluation the weather variables from 1–6 wk prior to the sampling period (Table 2). The model also includes 2 terms to account for seasonal variation in arbovirus activity, including week (of the year) and week of the year squared. These terms worked together to impose a hyperbolic shape on the effect that time of year had on the probability of infection, providing an estimate of the average peak time of viral activity over the survey period.

To assess likely weather predictors of WNV infection in mosquitoes, Spearman's rank correlations were calculated between weather variables and positive WNV results. Variables were considered for inclusion in model development if they were significantly correlated (P < 0.05) to a positive WNV outcome. If multiple offset variables within a single variable (i.e., average dew point at various time points [DA, DA2, DA3, and DA5]) were selected for model inclusion, then all selected variables were tested for multicollinearity and only the most correlated to a WNV positive were used. Years 2009–16 of our data set were used to generate our model while year 2017 was withheld to test our model results against for validation. Model selection was accomplished through a backwards stepwise selection procedure based on the lowest Akaike information criterion (AIC_c) value (Diuk-Wasser et al. 2006). The model is described with the following general equation using logit link:

where Y indicates whether an infected mosquito was captured (Y = 1) or not (Y = 0), β₀ is the intercept, and β₁…β_k represent the coefficients associated with each independent variable x₁…x_k.

Final model outputs were converted into a probability (P) using the following equation (Diuk-Wasser et al. 2006):

where P is the probability of mosquitoes trapped during a given week testing positive for WNV. Lastly, as a measure of model utility, the sensitivity (i.e., true positive rate) and specificity (i.e., true negative rate) of the models to predict infection were assessed at a probability cutoff of 0.5 (i.e., model predictions >0.5 were deemed infectious while those <0.5 were deemed not infectious). Sensitivity was calculated in the following way:

while specificity was calculated using the following equation:

The locations of the highest mutual sensitivity and specificity were also determined (Dohoo et al. 2012). The accuracy of the model was also calculated using the following equation:

and compared between the years used to generate the model (2009–16) and the year 2017, which was withheld to test the model against. All statistical analyses were performed using SAS^® software (SAS Institute Inc., Cary, NC).

Twelve of the original 52 variables were initially selected for inclusion in model development based on their statistically significant correlation to a positive WNV outcome. Backwards stepwise selection produced the following final model (Table 3):

where Y is the log-odds of capturing an infected mosquito, Week is the week of year (numeric) the model is calculated for, Week² is the week of year (numeric) squared the model was calculated for, HA2 is the average humidity from 2 wk prior to sampling, VA is the average visibility from the week the model is calculated, WA3 is the average wind from 3 wk prior to sampling, DA4 is the average dew point from 4 wk prior to sampling. Three of the parameters (Week, Week², and DA4) in the model were statistically significant (Wald's chi-square, P < 0.05).

The final model was applied to each of the 195 wk used to generate the model where mosquitoes were tested. The highest sensitivity and specific values (approximately 86%) coincided at a probability cutoff of approximately 0.28. At a probability cutoff of 0.5, the sensitivity of the model was 66%, while the specificity was 96% (Fig. 2). The accuracy of the model (years 2009–16) at a probability cutoff of 0.5 to predict the probability of mosquitoes testing positive for WNV was 91% and the accuracy of the model results from the withheld 2017 data was 92%.

By 2002, WNV had emerged as a full-scale epidemic. Within 5 years postintroduction into the USA, the virus had spread throughout the country. West Nile virus is now considered the leading cause of arboviral disease in the USA (CDC 2017). As no specific WNV treatments exist for those infected with the virus, and no vaccines are currently available to aid in the prevention of human disease (OIE 2013, CDC 2017), the only effective public prevention must be accomplished through VCP management of mosquito populations (Petersen et al. 2002, Sampathkumar 2003). To be effective, VCPs must have the knowledge of the risk of disease posed by mosquito populations.

We constructed a statistical model to predict the probability of WNV occurring in mosquito populations. For development of the model we used weather variables as predictors due to the ease of this type of data collection by VCP employees, and the biological dependency of mosquito populations on climate. Our final model included variables that are likely related to aspects of both mosquito and viral ecology, as well as factors influencing capture rates.

The WNV model had an average seasonal peak occurrence in week 34 (mid-August) (Fig. 3), with a probability (11.3%) of infection due to week alone, when all other variables were held at their average values. This suggests that seasonal factors not explicitly included here, such as day length, may be important predictors of infection status of mosquitoes. We overlaid the average weekly minimum infection rate on this graph for comparison and we see that at the first sign of increased infection, the average seasonal occurrence starts to trend upwards as well. This is a great indicator of when the first anticipated detection of WNV will be. This graph can be used by the Lubbock County VCP as a tool to aid their abatement efforts.

Our model had a negative relationship with average wind, which is intuitive and consistent with previously reported studies (Reisen et al. 2003, Karki et al. 2016). Our model also had a positive association with average dew point, humidity, and visibility. Soverow and colleagues (2009) found a similar positive association with dew point and humidity, with the exception that they found their results most significant 3 wk prior to be the best indicator whereas our results used dew point 4 wk prior to sampling and humidity 2 wk prior to sampling. Soverow and colleagues (2009) suggest that the effect of humidity is through mosquito breeding rather than biting, while Shaman and Day (2007) suggest increased humidity enhances mosquito flight activity and host-seeking behavior.

Surprisingly, our model development did not select for temperature or precipitation as both are commonly reported as potential drivers, both related to mosquito biology and epidemiologic aspects of WNV (Landesman et al. 2007, Degroote et al. 2008, Winters et al. 2008, Soverow et al. 2009, Ruiz et al. 2010, Lockaby et al. 2016). However, the nature and magnitude of specific variables may vary based on the location, as well as spatial and temporal scales. For example, when considering the spatial scale of the USA, between 2001 and 2005, Soverow and colleagues (2009) reported that both short-term higher temperature and increased precipitation were positively associated with WNV prevalence in humans. In contrast, Degroote and others (2008) reported that in a study of WNV prevalence in Iowa from 2002–06, relationships with climatic variables varied between years and geographic units, and that in at least 2 years, WNV transmission was associated with drier conditions. These differences point out the difficulty of generalizing the prediction of WNV transmission between study regions, as well as the links between the factors that drive mosquito and human infection dynamics. As far as the authors are aware, few previous efforts have been reported that characterize such dynamics in the southern High Plains of Texas (Stephens 2010). This study represents an initial step toward informed decision making for mosquito management, and contributes to a better understanding of WNV dynamics in the southern High Plains, though clearly more work is required.

The current model has a moderate sensitivity (66%) at a predicted probability of infection of 0.5. Thus, it offers a 16% improvement for predicting positive cases over chance alone. On the other hand, the specificity of the model at the 0.5 probability level is quite high (96%), meaning that there is high confidence in model prediction of noninfection. The accuracy (91%) of this model is also respectable. A relatively large sample size and sampling period (9 years) also lends robustness to model predictions. Although likely improved by additional predictor classes, a strength of the model presented here is that its use by VCPs in the southern High Plains region of Texas is relatively easy due to its few parameters, and the readily available nature of the weather information needed to solve it. In addition, it can also be implemented and solved with widely available spreadsheet computer software (i.e., MS Excel), with only minimal training needed to interpret outputs. Thus, this model provides a needed contribution to mosquito control efforts in the region, and may serve as the basis for future, similar efforts. In particular, the incorporation of ecological, socioeconomic, and treatment factors would likely strengthen the accuracy of predictions and provide a greater understanding of factors driving WNV infection in mosquitoes in Lubbock County.

The authors would like to thank Anna Hoffarth, Juliet Kinuya, Misty Moriarty, Sam Hawkins, Richie Erickson, Galen Austin, and Christina Stephens for all of their assistance with data collection and/or processing, as well as the City of Lubbock Health Department and Vector Control Team.