Adjusting for sampling variability in sparse data: geostatistical approaches to disease mapping

A Bayesian maximum entropy model for predicting tsetse ecological distributions

BACKGROUND

African trypanosomiasis is a tsetse-borne parasitic infection that affects humans, wildlife, and domesticated animals. Tsetse flies are endemic to much of Sub-Saharan Africa and a spatial and temporal understanding of tsetse habitat can aid surveillance and support disease risk management. Problematically, current fine spatial resolution remote sensing data are delivered with a temporal lag and are relatively coarse temporal resolution (e.g., 16 days), which results in disease control models often targeting incorrect places. The goal of this study was to devise a heuristic for identifying tsetse habitat (at a fine spatial resolution) into the future and in the temporal gaps where remote sensing and proximal data fail to supply information.

METHODS

This paper introduces a generalizable and scalable open-access version of the tsetse ecological distribution (TED) model used to predict tsetse distributions across space and time, and contributes a geospatial Bayesian Maximum Entropy (BME) prediction model trained by TED output data to forecast where, herein the Morsitans group of tsetse, persist in Kenya, a method that mitigates the temporal lag problem. This model facilitates identification of tsetse habitat and provides critical information to control tsetse, mitigate the impact of trypanosomiasis on vulnerable human and animal populations, and guide disease minimization in places with ephemeral tsetse. Moreover, this BME analysis is one of the first to utilize cluster and parallel computing along with a Monte Carlo analysis to optimize BME computations. This allows for the analysis of an exceptionally large dataset (over 2 billion data points) at a finer resolution and larger spatiotemporal scale than what had previously been possible.

RESULTS

Under the most conservative assessment for Kenya, the BME kriging analysis showed an overall prediction accuracy of 74.8% (limited to the maximum suitability extent). In predicting tsetse distribution outcomes for the entire country the BME kriging analysis was 97% accurate in its forecasts.

CONCLUSIONS

This work offers a solution to the persistent temporal data gap in accurate and spatially precise rainfall predictions and the delayed processing of remotely sensed data collectively in the - 45 days past to + 180 days future temporal window. As is shown here, the BME model is a reliable alternative for forecasting future tsetse distributions to allow preplanning for tsetse control. Furthermore, this model provides guidance on disease control that would otherwise not be available. These 'big data' BME methods are particularly useful for large domain studies. Considering that past BME studies required reduction of the spatiotemporal grid to facilitate analysis. Both the GEE-TED and the BME libraries have been made open source to enable reproducibility and offer continual updates into the future as new remotely sensed data become available.

Progression of a large syphilis outbreak in rural North Carolina through space and time: Application of a Bayesian Maximum Entropy graphical user interface

In 2001, the primary and secondary syphilis incidence rate in rural Columbus County, North Carolina was the highest in the nation. To understand the development of syphilis outbreaks in rural areas, we developed and used the Bayesian Maximum Entropy Graphical User Interface (BMEGUI) to map syphilis incidence rates from 1999-2004 in seven adjacent counties in North Carolina. Using BMEGUI, incidence rate maps were constructed for two aggregation scales (ZIP code and census tract) with two approaches (Poisson and simple kriging). The BME maps revealed the outbreak was initially localized in Robeson County and possibly connected to more urban endemic cases in adjacent Cumberland County. The outbreak spread to rural Columbus County in a leapfrog pattern with the subsequent development of a visible low incidence spatial corridor linking Roberson County with the rural areas of Columbus County. Though the data are from the early 2000s, they remain pertinent, as the combination of spatial data with the extensive sexual network analyses, particularly in rural areas gives thorough insights which have not been replicated in the past two decades. These observations support an important role for the connection of micropolitan areas with neighboring rural areas in the spread of syphilis. Public health interventions focusing on urban and micropolitan areas may effectively limit syphilis indirectly in nearby rural areas.

Exposure to industrial hog and poultry operations and urinary tract infections in North Carolina, USA

An increasing share of urinary tract infections (UTIs) are caused by extraintestinal pathogenic Escherichia coli (ExPEC) lineages that have also been identified in poultry and hogs with high genetic similarity to human clinical isolates. We investigated industrial food animal production as a source of uropathogen transmission by examining relationships of hog and poultry density with emergency department (ED) visits for UTIs in North Carolina (NC). ED visits for UTI in 2016-2019 were identified by ICD-10 code from NC's ZIP code-level syndromic surveillance system and livestock counts were obtained from permit data and aerial imagery. We calculated separate hog and poultry spatial densities (animals/km²) by Census block with a 5 km buffer on the block perimeter and weighted by block population to estimate mean ZIP code densities. Associations between livestock density and UTI incidence were estimated using a reparameterized Besag-York-Mollié (BYM2) model with ZIP code population offsets to account for spatial autocorrelation. We excluded metropolitan and offshore ZIP codes and assessed effect measure modification by calendar year, ZIP code rurality, and patient sex, age, race/ethnicity, and health insurance status. In single-animal models, hog exposure was associated with increased UTI incidence (rate ratio [RR]: 1.21, 95 % CI: 1.07-1.37 in the highest hog-density tertile), but poultry exposure was associated with reduced UTI rates (RR: 0.86, 95 % CI: 0.81-0.91). However, the reference group for single-animal poultry models included ZIP codes with only hogs, which had some of the highest UTI rates; when compared with ZIP codes without any hogs or poultry, there was no association between poultry exposure and UTI incidence. Hog exposure was associated with increased UTI incidence in areas that also had medium to high poultry density, but not in areas with low poultry density, suggesting that intense hog production may contribute to increased UTI incidence in neighboring communities.

Bayesian maximum entropy-based prediction of the spatiotemporal risk of schistosomiasis in Anhui Province, China

Background

“Schistosomiasis” is a highly recurrent parasitic disease that affects a wide range of areas and a large number of people worldwide. In China, schistosomiasis has seriously affected the life and safety of the people and restricted the economic development. Schistosomiasis is mainly distributed along the Yangtze River and in southern China. Anhui Province is located in the Yangtze River Basin of China, with dense water system, frequent floods and widespread distribution of Oncomelania hupensis that is the only intermediate host of schistosomiasis, a large number of cattle, sheep and other livestock, which makes it difficult to control schistosomiasis. It is of great significance to monitor and analyze spatiotemporal risk of schistosomiasis in Anhui Province, China. We compared and analyzed the optimal spatiotemporal interpolation model based on the data of schistosomiasis in Anhui Province, China and the spatiotemporal pattern of schistosomiasis risk was analyzed.

Methods

In this study, the root-mean-square-error (RMSE) and absolute residual (AR) indicators were used to compare the accuracy of Bayesian maximum entropy (BME), spatiotemporal Kriging (STKriging) and geographical and temporal weighted regression (GTWR) models for predicting the spatiotemporal risk of schistosomiasis in Anhui Province, China.

Results

The results showed that (1) daytime land surface temperature, mean minimum temperature, normalized difference vegetation index, soil moisture, soil bulk density and urbanization were significant factors affecting the risk of schistosomiasis; (2) the spatiotemporal distribution trends of schistosomiasis predicted by the three methods were basically consistent with the actual trends, but the prediction accuracy of BME was higher than that of STKriging and GTWR, indicating that BME predicted the prevalence of schistosomiasis more accurately; and (3) schistosomiasis in Anhui Province had a spatial autocorrelation within 20 km and a temporal correlation within 10 years when applying the optimal model BME.

Conclusions

This study suggests that BME exhibited the highest interpolation accuracy among the three spatiotemporal interpolation methods, which could enhance the risk prediction model of infectious diseases thereby providing scientific support for government decision making.

Development of the Australian Cancer Atlas: spatial modelling, visualisation, and reporting of estimates

BACKGROUND

It is well known that the burden caused by cancer can vary geographically, which may relate to differences in health, economics or lifestyle. However, to date, there was no comprehensive picture of how the cancer burden, measured by cancer incidence and survival, varied by small geographical area across Australia.

METHODS

The Atlas consists of 2148 Statistical Areas level 2 across Australia defined by the Australian Statistical Geography Standard which provide the best compromise between small population and small area. Cancer burden was estimated for males, females, and persons separately, with 50 unique sex-specific (males, females, all persons) cancer types analysed. Incidence and relative survival were modelled with Bayesian spatial models using the Leroux prior which was carefully selected to provide adequate spatial smoothing while reflecting genuine geographic variation. Markov Chain Monte Carlo estimation was used because it facilitates quantifying the uncertainty of the posterior estimates numerically and visually.

RESULTS

The results of the statistical model and visualisation development were published through the release of the Australian Cancer Atlas ( https://atlas.cancer.org.au ) in September, 2018. The Australian Cancer Atlas provides the first freely available, digital, interactive picture of cancer incidence and survival at the small geographical level across Australia with a focus on incorporating uncertainty, while also providing the tools necessary for accurate estimation and appropriate interpretation and decision making.

CONCLUSIONS

The success of the Atlas will be measured by how widely it is used by key stakeholders to guide research and inform decision making. It is hoped that the Atlas and the methodology behind it motivates new research opportunities that lead to improvements in our understanding of the geographical patterns of cancer burden, possible causes or risk factors, and the reasons for differences in variation between cancer types, both within Australia and globally. Future versions of the Atlas are planned to include new data sources to include indicators such as cancer screening and treatment, and extensions to the statistical methods to incorporate changes in geographical patterns over time.

Modelling local uncertainty in relations between birth weight and air quality within an urban area: combining geographically weighted regression with geostatistical simulation

In this study, we combine known methods to present a new approach to assess local distributions of estimated parameters measuring associations between air quality and birth weight in the urban area of Sines (Portugal). To model exposure and capture short-distance variations in air quality, we use a Regression Kriging estimator combining air quality point data with land use auxiliary data. To assess uncertainty of exposure, the Kriging estimator is incorporated in a sequential Gaussian simulation algorithm (sGs) providing a set of simulated exposure maps with similar spatial structural dependence and statistical properties of observed data. Following the completion of the simulation runs, we fit a geographically weighted generalized linear model (GWGLM) for each mother's place of residence, using observed health data and simulated exposure data, and repeat this procedure for each simulated map. Once the fit of GWGLM with all exposure maps is finished, we take the distribution of local estimated parameters measuring associations between exposure and birth weight, thus providing a measure of uncertainty in the local estimates. Results reveal that the distribution of local parameters did not vary substantially. Combining both methods (GWGLM and sGs), however, we are able to incorporate local uncertainty on the estimated associations providing an additional tool for analysis of the impacts of place in health.

Exploring spatial patterns of sudden cardiac arrests in the city of Toronto using Poisson kriging and Hot Spot analyses

INTRODUCTION

Our study looked at out-of-hospital sudden cardiac arrest events in the City of Toronto. These are relatively rare events, yet present a serious global clinical and public health problem. We report on the application of spatial methods and tools that, although relatively well known to geographers and natural resource scientists, need to become better known and used more frequently by health care researchers.

MATERIALS AND METHODS

Our data came from the population-based Rescu Epistry cardiac arrest database. We limited it to the residents of the City of Toronto who experienced sudden arrest in 2010. The data was aggregated at the Dissemination Area level, and population rates were calculated. Poisson kriging was carried out on one year of data using three different spatial weights. Kriging estimates were then compared in Hot Spot analyses.

RESULTS

Spatial analysis revealed that Poisson kriging can yield reliable rates using limited data of high quality. We observed the highest rates of sudden arrests in the north and central parts of Etobicoke, western parts of North York as well as the central and southwestern parts of Scarborough while the lowest rates were found in north and eastern parts of Scarborough, downtown Toronto, and East York as well as east central parts of North York. Influence of spatial neighbours on the results did not extend past two rings of adjacent units.

CONCLUSIONS

Poisson kriging has the potential to be applied to a wide range of healthcare research, particularly on rare events. This approach can be successfully combined with other spatial methods. More applied research, is needed to establish a wider acceptance for this method, especially among healthcare researchers and epidemiologists.

Spatiotemporal approaches to analyzing pedestrian fatalities: the case of Cali, Colombia

OBJECTIVE

Injuries among pedestrians are a major public health concern in Colombian cities such as Cali. This is one of the first studies in Latin America to apply Bayesian maximum entropy (BME) methods to visualize and produce fine-scale, highly accurate estimates of citywide pedestrian fatalities. The purpose of this study is to determine the BME method that best estimates pedestrian mortality rates and reduces statistical noise. We further utilized BME methods to identify and differentiate spatial patterns and persistent versus transient pedestrian mortality hotspots.

METHODS

In this multiyear study, geocoded pedestrian mortality data from the Cali Injury Surveillance System (2008 to 2010) and census data were utilized to accurately visualize and estimate pedestrian fatalities. We investigated the effects of temporal and spatial scales, addressing issues arising from the rarity of pedestrian fatality events using 3 BME methods (simple kriging, Poisson kriging, and uniform model Bayesian maximum entropy). To reduce statistical noise while retaining a fine spatial and temporal scale, data were aggregated over 9-month incidence periods and censal sectors. Based on a cross-validation of BME methods, Poisson kriging was selected as the best BME method. Finally, the spatiotemporal and urban built environment characteristics of Cali pedestrian mortality hotspots were linked to intervention measures provided in Mead et al.'s (2014) pedestrian mortality review.

RESULTS

The BME space-time analysis in Cali resulted in maps displaying hotspots of high pedestrian fatalities extending over small areas with radii of 0.25 to 1.1 km and temporal durations of 1 month to 3 years. Mapping the spatiotemporal distribution of pedestrian mortality rates identified high-priority areas for prevention strategies. The BME results allow us to identify possible intervention strategies according to the persistence and built environment of the hotspot; for example, through enforcement or long-term environmental modifications.

CONCLUSIONS

BME methods provide useful information on the time and place of injuries and can inform policy strategies by isolating priority areas for interventions, contributing to intervention evaluation, and helping to generate hypotheses and identify the preventative strategies that may be suitable to those areas (e.g., street-level methods: pedestrian crossings, enforcement interventions; or citywide approaches: limiting vehicle speeds). This specific information is highly relevant for public health interventions because it provides the ability to target precise locations.