A systematic review of spatial decision support systems in public health informatics supporting the identification of high risk areas for zoonotic disease outbreaks

Selection of studies

Inclusion and exclusion criteria

Assessment

For each article that met our inclusion criteria, we reviewed the entire text in addition to tables, figures and supplementary material. For the review process, we first evaluated the quality of each publication using a scoring methodology (MMAT) approved for use in systematic reviews. This tool first applies screening criteria for any type of study to confirm whether there are clear objectives and whether the collected data allowed the researcher(s) to address the objective. Studies which passed the initial screening question were then further evaluated based on study type (quantitative, qualitative, or mixed methods) using a ranking system to answer specific questions (see Fig. 2), with either: yes, no, unable to tell, or not applicable similarly to Fournet et al. [39] (each response was ranked as 4, 3, 2, 1 respectively). The Template for intervention description and replication (TIDieR) checklist was then used to evaluate each of the full texts that passed the MMAT screening for completeness in addressing the intervention proposed as a means of qualitative synthesis [40]. The TIDieR was published in 2014 to extend the CONSORT and SPIRIT statements, and can be applied to all evaluation study designs [41]. This checklist covers topics including naming the specific intervention, why, what, who, how, where, and when.

General descriptions of the final set of articles are documented in excel, including the target disease or diseases, the target population at risk, and the geographic coverage, and data sources and types used for identifying areas at risk for outbreak. To compare each SDSS in terms of modeling capability, we documented and described eight necessary characteristics of a SDSS as identified by Sugumaran et al. [29]. These include spatial data management/analysis, visualization, report generations, interactive problem solving, spatial modeling, semi structured problem solving, scenarios evaluation, and easy user interaction.

Screening results

After duplicates were removed from search results and inclusion/exclusion criteria were applied to title and abstract screening, 75 articles were subject to full text review. We then screened each full article for inclusion/exclusion criteria and reduced the final set to 34 articles. Of those remaining, all articles selected described SDSS development intended for public health officials as end users which focused on identifying at risk areas of infectious disease outbreaks. The full screening process is summarized in Fig. 3. The final set of articles were published evenly throughout the selected time-period of 2008–2018, with a decreasing trend towards the end of the study period. See Additional file 1 for a basic description of all articles [8, 23, 24, 31–34, 42–109] that underwent full text review, with reasons for exclusion.

Qualitative assessment

Quality filtering

Of the final 34 articles identified after full text screening, 22 did not meet the MMAT screening criteria by addressing a specific objective or research question that could be readily quantitatively or qualitatively assessed (see Additional file 2 and Fig. 4). The articles which could not be fully evaluated using the MMAT generally reviewed the development and implementation of an SDSS, while a few also described a case study (n = 4), a simulation study (n = 3), and a pilot study (n = 1). Of those studies fully evaluated using the MMAT (n = 12), three were qualitative, seven were quantitative, and 2 were mixed methodologies. The average MMAT score for the final set of 12 articles was 87.5%.

Quality of interventions description

The papers that passed the MMAT screening questions were also assessed using the TIDieR checklist to determine if the proposed intervention was adequately described (see Fig. 5 and Additional file 3). All studies addressed most items on the checklist, all with a similar intervention to develop a spatial decision support system, though with some variations to address high risk areas for outbreaks and methodology. Rationales behind development varied from visualization of space time events, detection, monitoring or prediction of disease clusters or epicenters, or more specific tasks related to targeting regions for disease elimination or management. All but one [73] article described the intervention as effective, either through sensitivity (n = 3), significance (n = 3), or percentage (n = 2) evaluations on ability to identify at risk areas for outbreaks, in addition to user access traffic (n = 2), and user feedback on usability or usefulness (n = 5).

Evaluation of selected SDSS

Summary of core SDSS characteristics

General features

In our review, we found a large amount of diversity in design and scope regarding study areas targeted by recent SDSS. The systems observed here were developed by smaller research teams at individual institutions, as well as by government divisions throughout the world to address the concerning zoonotic disease developments and spread which have increased in recent years. Several SDSS built for smaller regions relied on local data sources and reporting, whereas those developed to be more generalizable used more public data sources. These findings highlight a common problem wherein systems are developed in isolation and rely on locally collected and curated data.

Fortunately, the trend in the data presented here indicates most systems are developed to address numerous diseases over large regions, entire countries, or the entire globe. Those systems that are more geographically inclusive have the potential to address a range of decision making needs of public health officials who are often faced with a variety of outbreak threats from surrounding environments. This universal approach reduces the necessity to have multiple tools that perform similar or redundant tasks.

Core SDSS functionalities

For each of the core functionalities, studies often provided enough detail to determine specific details regarding architectural design that allow visualization, reporting, problem solving etc. and implementation. Throughout the history of SDSS development, there has been a heavy reliance on proprietary, standalone GIS [111]. Beginning with open source GIS distributions such as GRASSGIS, the last 10–15 years have seen a remarkable expansion in the availability of open source software such as statistical packages, web mapping services, and others which have shifted the focus on client–server architecture to a web services model in newer SDSS [111]. The results here indicate that the availability of new tools to aid SDSS development are diversifying what software is used to produce the final product, ranging from desktop applications integrating one of many GIS with spatial database management to the implementation of web application frameworks that utilize open source database management systems (PostgreSQL, MySQL, etc.) with mapping services such as OpenLayers. Also reflected are a large variety of analytical tools or algorithms that can be packaged for use in a variety of systems for a similar purpose, or complementary analysis. For instance, Kulldorff’s spatial scan statistic software [112], k-nearest neighbor and support-vector machine learning models were implemented in just one SDSS reviewed here.

Many of the reviewed studies were less consistent in describing or performing model evaluations, obtaining stakeholder requirements, pilot studies, or usability testing. Several studies described their system as user friendly [23, 48, 49, 51, 64, 106, 107, 113], though not all performed an evaluation or even approached the topic in the same way. Most studies utilizing this term referred to usability testing, which was often followed with robust evaluation methods. However, other studies used the term in reference to ease of configuration with other software, or visual displays that were esthetically appealing without providing details of user participation to evaluate these claims. Despite this, it is noteworthy that usability metrics mentioned within the selected articles do not necessarily evaluate all aspects that potentially contribute to ease of use.

Current data sources and future developments

Integration of disparate data sources has been a common theme in surveillance systems and decision support systems for public health surveillance [114]. To accomplish this, collections from local health practitioners, state departments, or government agencies may be joined such that reportable disease data are housed within the same framework that integrates sources such as media, genetic sequencing, socio-economic data, or other environmental characteristics. However, integration of multiple data sources was not universal in recently developed SDSS reviewed here. Availability and utilization of social media, news, landscape, climactic, socio-economic data are reflected here, as several systems covered are inclusive of these increasingly readily available data types, providing even more data to better understand what contributes to disease spread and outbreaks. There are still many inequities in accessibility to resources globally that make collection and availability of data difficult, stratifying health status and potential interventions by socio-economic means and access to urban areas [115]. Limited resources to collect and organize data to develop effective surveillance systems is further impacted with difficulties in combining datasets recorded for varying formats, quality standards, and reporting requirements [116].

Genetic sequencing and related bioinformatics tools and resources are an emerging data source which are not well represented in recent SDSS development. Open source genetic data repositories are increasingly more common, such as GenBank [19] and the IRD [20] which contain an abundance of pathogen sequences and relevant metadata that present opportunities to enhance current practices. Utilization of sequencing to support epidemiological and disease prevention efforts have been demonstrated in recent epidemic events including the Ebola outbreak in west Africa from 2014 to 2015, in which genetic sequencing provided valuable information on the genetic diversity, to estimate how fast the disease spread and predict future transmission through phylogenetic analysis [117]. Likewise, Fraser et al. found that transmission potential estimates derived from genetic sequencing were comparable to traditional epidemiological estimates based clinical attack rates during the H1N1 pandemic of 2009 [118]. While these developments show great potential for the use of bioinformatic tools and resources in public health, the public sector has yet to take full advantage of genetic data driven tools. Despite a substantial amount of literature that documents the value of understanding the genetic variability of disease [119–121], only three potential articles included in the screening process were inclusive of sequencing data but did not pass the inclusion criteria or quality filtering process [53, 89, 95]. Complications for including sequencing data for zoonotic surveillance are largely due to limited resources or laboratory capacity to rapidly sequence emerging infections [121], and a unifying framework from which to integrate the phylogenetic patterns of an infectious agent and potential hosts [122], in addition to drivers of disease.

Understanding the phylogenetics of a disease in context with the spatial distribution, environmental characteristics, and potential hosts, assists in the identification of geographic areas that drive zoonotic disease spread and circulation. Ge et al. [123] compared methodologies for H5N1 avian influenza outbreak identification derived from three different fields including phylogenetics, spatial statistics, and epidemiological analysis of socio-ecological determinants of disease such as human and avian population density, migration routes, railways, and inland water, which found that integrating all three serves to better corroborate observed occurrence and estimate ability to spread. Changes in human population, industrialization, global trade, and travel also play a key role in facilitating the introduction or discovery of novel pathogens throughout the world [124], and phylogenetic analysis facilitates tracing the spread of disease through human interactions [125]. Health agencies are now better able to respond to emerging zoonotic diseases through understanding how changes in the genome of a pathogen impacts disease risk and spread by identifying transmission routes, mutation rates, in conjunction with other epidemiological parameters of interest [3]. Substantial concern for potential pandemics due to transmission of zoonotic disease from animal to human populations such as avian influenza [126] has also prompted integration of animal phylogenetic analysis to better understand interaction and transmission which can aid management and preventative measures [127–129]. The need to develop an integrative approach necessary to understand the emerging spatial patterns of zoonotic disease outbreak and spread has long been recognized, with the advent of movements such as “One Health” which call for a global strategy that is multidisciplinary to address the health of humans, animals, and the ecosystem [130]. Future development should focus on assisting in the ease of access and integration of multiple data sources to achieve this goal.

Future directions

Data integration and availability still pose several challenges in the development of SDSS for zoonotic disease outbreak management, including variability among available sources such as scale, completeness, and timeliness. Potential data sources range from the microscopic to the global scale, necessitating data transformation and projection for appropriate analysis and visualization. While georeferenced datasets describing environmental and climactic phenomena are more readily available, emerging genetic data sets present some challenges, as location of isolation are generally extracted manually from public records or publications. To address visualization of genetic sequence data, there have been efforts to extract geospatial metadata such as location and host from GenBank records, to ease automation of linking relevant sequence data for spatial modeling of disease [131]. Efforts to model outbreaks within a decision support environment which integrate data collected on different spatial scales need to address automated data extraction and transformation such as aggregation of case reports, host population densities, and locations from which isolates were sequenced for a region under study such that visualization of a multifaceted scenario is possible.

Sparsity in collected data across large geographic regions could also introduce uncertainty in results, leading to imputation or summarization methods to account for unobserved data. This is particularly a problem for genetic sequence data currently available, which does not cover regions comprehensively and geographic metadata may vary by location specificity, ranging from coordinates to country [132]. Machine learning algorithms have been utilized in clinical decision support systems to address missing patient data to aid medical professionals with decision making [133], and may also prove useful in spatial decision support systems to address this limitation. Other approaches to documenting disease case reporting that have been increasingly utilized are social media, new sources and various forms of citizen science. One SDSS reviewed here incorporated public reports of disease morbidity to enhance surveillance, provided through a hotline or online submission. The advent of the mobile phone in an increasingly interconnected digital age has given rise to a form of citizen science aptly described by the phrase ‘Wikification of GIS by the masses’ (WGM) coined by Kamel Boulos [134], and later described as volunteered geographic information (VGI) [135]. WGM is the specific embodiment of crowdsourcing approach in GIS [136]. WGM encompasses data contributions from citizens to bring local knowledge and spatial awareness on matters of interest into better focus. WGM is commonplace now in applications such as Google maps, and similar platforms have arisen to assist in disease surveillance such as Flu Near You [137, 138]. Recent literature has questioned the utility of WGM for surveillance purposes, suggesting that crowd sourced reporting at the local level does not correlate well with CDC records [138]. However, this citizen reporting through crowd sourcing can potentially complement traditional data sources.

Timeliness of emerging genetic data sources are currently primarily used for retrospective study due to inability to perform sequencing locally during epidemiological investigations, especially in regions of the world with limited resources and infrastructure. However, advances in near real time genetic sequencing may address this problem in the near future with the arrival of a portable sequencing instrument, which can produce sequencing results within a day of receiving a sample during on ongoing epidemic [139].

Future SDSS also need to better support existing public health surveillance practices, skills, resources and methodologies [140]. Previous work indicates that LHD’s utilize a wide variety of information systems for surveillance and identifying at risk areas in the United States alone, with little understanding of effectiveness [141]. Incorporating user centered design principles such as designing for the user tasks would allow more effective, efficient work within a user-friendly interface. Sutcliffe et al. has previously described a design framework for visualization decision support tools for epidemiology which incorporates user centered design [101]. They argue that to effectively translate data into actionable policy or prevention measures, the software designers and end-user analysts must collaborate extensively. This approach has been implemented more consistently in clinical informatics projects such as the design of electronic health records [142], though rarely in public health informatics [101]. The successful implementation of decision support tools is often limited to evaluations of the proposed system in which potential users provide valuable feedback, and frequently suffer from user dissatisfaction due to a mismatch in expectation, user knowledge of spatial analysis and other contributing factors [143]. While several SDSS here did evaluate usability after system design, often without prior user input, our finding here indicate that improvements can still be made to better include potential users throughout the design process.

Limitations

To limit the possible number of comparisons to more dated methods and technology, we focused on more recent work and thus excluded articles prior to 2008. We also did not search all available databases, opting for the breadth of Google scholar which might have resulted in unique publications that were not available in individual databases.