In the 1940s, a malaria epidemic began when economic pressures and poverty led the agriculture sector of Trinidad into the cacao industry, an activity that requires a large human labor force and shade trees. These shade trees supported a bromeliad tank species, a family of plants that contain water-holding structures, creating an ideal breeding site for Anopheles bellator mosquitoes and new niches for different mosquitoes within the forest canopy (19,20). During the epidemic, medical doctors first noted an increase in the prevalence of splenomegaly, an indication of malaria, among schoolchildren correlated with areas cultivating cacao, and that A. bellator mosquitoes were only found near the cacao farms. Removal of the bromeliads reduced A. bellator mosquito populations and returned malaria rates to prior endemic levels (20). In hindsight, the outbreak resulted from the convergence of poverty, commerce, and agricultural and land-use changes; lack of malaria prevention and treatment services; and resulting shifts in local mosquito ecology. Building on our accumulating knowledge of drivers and the digital streams of data that now exist, the public health sector can be provided with early warning tools to recognize when conditions are ripe for disease emergence events, tools that could not have existed in the mid-20th century.

Driver monitoring requires access to data from multiple domains at sufficient resolution and scale to correlate with known disease events and overcome potential biases. To build a driver surveillance system, issues with data availability and compatibility must be addressed in conjunction with temporal resolution and spatial scale.

Data availability categorized by drivers of interest is shown in Figure 2. The data were collected for the HealthScapes Project (http://healthscapes.io), an initiative launched in 2009 to assemble and collaboratively curate global open data of relevance to the global health research community. Of note, datasets were not readily found for certain drivers (i.e., human wildlife interaction and breakdown of public health measures), and few datasets were found for other drivers (i.e., human susceptibility to infection and poverty and social inequality).

Digital media reports already provide a rich source of information on disease outbreaks (24) and, likewise, may be useful in monitoring disease drivers to identify periods of increased risk for outbreaks (e.g., news articles about breakdown of public health infrastructure, conflict, and vaccine programs). However, this driver surveillance resource presents a unique mix of opportunities and challenges. On the opportunity side, only a handful of languages account for most content on the Internet, resulting in a distilled set of global languages (25). If this pattern continues, only a subset of languages will need surveillance. Furthermore, the ability of natural language processing tools to parse reports for contextual data, such as location information, is improving. On the challenge side, digital media is fleeting and curated datasets are limited. For example, Google News, a popular digital media aggregator, does not provide a list of the sources used in its feed, and, thus, its content could change with time (26). Digital media can also miss nondigital disease-related news reports and details due to language or source (e.g., print and audio) limitations (27).

An objective evaluation of digital media for driver surveillance will ultimately require an unbiased dataset that links media reports of disease drivers to actual outbreaks that enable determination of the specificity and sensitivity of the system. Notably, it will be difficult to develop such a system because of the circular nature of these data (i.e., the system would be evaluated based on the system used to collect the data). In the absence of independently collected gold standard data, one cannot determine the degree to which unknown unknowns are missed. In summary, more research is needed in evaluating digital media systems and the influence of these potential limitations.

Numerous projects have been established as hubs for collecting, cataloging, and sharing global data on a wide variety of potential drivers. Underlying these streams is an increasingly open-data ethic. Numerous repositories of existing data specialize in distributing accessible data formats (e.g., World Organisation for Animal Health’s World Animal Health Information Database, Global Biodiversity Information Facility, Center for International Earth Science Information Network, Google Earth Engine, Gapminder World, European Union BioFresh project, and HealthScapes). Recent Digital Disease Detection conferences have highlighted numerous new data sources, including mobile phone and Internet technology that has enabled real-time syndromic and resource monitoring from social media (Twitter and Facebook), active surveillance (e.g., Flu Near You and Influenzanet), and participatory surveillance (Opendream, SMS for Life, and Saude Na Copa 2014). On the ground and at the source, data loggers and sensors are decreasing in cost and are supported by open-source platforms (e.g., Mataki, Arduino, and Onset HOBOs).

HealthMap is one example of emerging digital surveillance technology that uses a flexible architecture and an automated processing system to target ID events and the wildlife trade (13,28). This framework could be applied to identify any driver of interest as dictionaries for different languages and subject areas are used; the platform is fundamentally subject- and language-agnostic. This structure enables for inclusion of data not distributed digitally, resulting in the ability to expand overall media coverage. In the near future, developments in optical character and audio language recognition offer the potential for the rapid processing of nondigital media, such as photographs of print media and audio news sources.

Coinciding with the growing number of data sources has been an explosion of visualization, analytical, and hybrid tools, often fueled by the open-source community. Once data are collected, these tools will provide new ways for analysts to convey complex data and ideas, such as driver surveillance trends, to decision makers in a clear and logical format (e.g., Hans Rosling’s Trendalyzer, Technical Appendix Table 2). Cloud-based platforms have evolved from desktop-based Esri software and leverage the open-source framework, social coding, and collaboration to provide researchers new opportunities to combine complex multisource big data with newly created virtual forums (e.g., Google Earth Engine and social modeling platforms). In addition to new analytic techniques, digital driver surveillance can build on advances in the way science is conducted. One example is the Open Science Framework, which has pioneered online project spaces that enable others to make changes or add on to the project without affecting the original (https://osf.io/). Another example is rOpenSci (https://ropensci.org), an effort funded by the Sloan Foundation to wrap scientific application programming interfaces in a manner that is readily accessible to analysts and researchers. This new approach sets the groundwork for collaborative, iterative, and adaptive management of ID driver surveillance models.

The emergence of Nipah virus (NiV) on the global scene can be linked to several drivers, all of which could have been monitored and potentially recognized, resulting in quicker containment of outbreaks. The first human NiV cases were reported in 1998 in Malaysia. This initial spillover event has been linked to agricultural intensification where mango trees were planted in close proximity to pigsties (16,29). This change in practice led to increased contact between pigs and the natural reservoir for NiV, the flying fox, and subsequent introduction of NiV into the pig population (29). The continued contact between the bats and pigs resulted in NiV establishment within the pig population and increasing exposure of pig workers to NiV (29,30).

Perhaps NiV emergence could have been anticipated if a system existed that raised warnings about the coincident land-use change and agricultural intensification. Scientists were eventually able to match the unknown disease affecting humans with NiV and identified risk factors and modes of transmission, but the key connections might have been recognized sooner. Although an isolated change in agricultural practices would not typically warrant an intervention, the emergence of a new porcine respiratory disease in conjunction with the change could have led to closer monitoring. When the initial spillover into the human population occurred, the knowledge of a new porcine disease, changes in driver activity, and the identification of a new virus in bats all occurring within the same region could have individually triggered heightening alerts.

After emergence, driver surveillance would have provided valuable situational insight. In the early stages of the outbreak, there was a mass sell-off of pigs in response to the cluster of cases and the perceived link among farmers between the sick pigs and the cluster (29,31,32). This sell-off resulted in the spread of NiV to southern regions of Malaysia, where there is a dense pig population (29). As the outbreak spread southward, the international pig trade moved the virus to Singapore (32). Market monitoring could have potentially identified the increase in pig sales, as well as the regional sources and endpoints of traded pigs. If supply chains and the locations that sit at high-risk critical nodes could be identified, increased active surveillance or even preemptive interventions, such as animal culling, could have been deployed.

A successful epidemic surveillance platform should leverage the latest web technologies to support global communication among health workers with near–real-time information on the drivers of disease emergence. It is important that the platform have the capability to capture ground-truthed data from field teams to effectively crowd-source data curation to amplify efforts in the global response to disease emergence. Such open data curation and provisioning services would further help democratize data access to those often most affected by IDs. Core features of the system would include the ability to identify relevant drivers for a disease or region of interest as well as the capacity to monitor spatiotemporal trends for these drivers. A user could use this information to make decisions about allocating funding for health control measures or prioritizing health effect assessments of development initiatives. In the field, a mobile interface to the platform could be used to identify future field sampling sites, intervention options, or changes in drivers. It is essential the platform have the capacity to monitor and overlay multiple drivers, in light of known interactions. The utility of the system would be enhanced by being able to set notifications on drivers of interest that could initiate a response, follow up, or further investigation. A truly dynamic web interface would enable the user to simulate different health outcomes under different scenarios, such as economic policies, control strategies, or climate futures. In anticipation of future ID outbreaks, a well-designed system could be used strategically by decision makers to allocate or predeploy limited human and material resources to risk hotspots. Globally, the system could be used to help orchestrate multinational responses to emerging ID threats, identify priorities, and raise awareness of the role of antecedent drivers in global health.

For a first-generation analysis platform to prove useful, it must provide insight regarding specific drivers and the drivers’ subsequent effect upon current risk. We envision a country-based analysis framework that could comprise 3 components. The first component will measure a nation’s risk for an emerging ID event by establishing driver baselines through the use of historical data or published ratings. The hotspots approach used by Jones et al. (4) illustrates the potential of this component, which could be extended from using historical events to using historical driver data to identify geographic locations at risk for ID emergence. Together these baselines would provide a Bayesian prior distribution of risk. The second component will introduce existing and new digital alerts as new evidence for a model. The addition of these alerts would build a risk map according to baseline levels; that is, a driver with a low baseline will not contribute much to changing the overall risk with the addition of a few alerts, but for a driver with a high baseline, the addition of a few driver alerts could push it into a higher risk category. By overlaying the digital disease alerts upon the baseline prior distribution of risk, we can then analyze these data for trends and biases to anticipate future emerging ID events. The third component will involve continually updating beliefs about the distribution of risk that is determined on the basis of new evidence.

Despite our positive outlook on digital driver surveillance, we note several caveats, such as the complexity of causal human–environment interactions (33) and the use of models to confront imperfect knowledge. For instance, false-positive signals, when a driver system identifies a strong signal but no disease occurs, are inevitable. Without appropriate user sensitization, this result could lead to misallocation of resources and erosion of trust. Conversely, some disease events might lack any warning signals, as may have been the case with the 2009 pandemic of influenza A(H1N1)pdm09 virus infections. Despite the existence of these problems, we believe they will be alleviated through a better understanding of the disease–driver and driver–driver associations. However, some problems will be unavoidable. For example, a potentially strong limitation would be that infrequent updating of driver data could result in either missing a brief signal or not identifying the signal until it is too late to implement control measures. Ultimately, we must recognize that digital surveillance is one tool among many that can help supplement, but not replace, traditional public health surveillance.

A digital driver surveillance platform that improves situational awareness by active monitoring of ID events and associated drivers is an obtainable goal for the public health community. We believe that future surveillance platforms will be able to not only describe driver activity in space and time but also indicate driver thresholds, severity, and likely interactions among drivers. Digital media reporting offers tremendous potential to contribute to existing datasets because of its automated ability to scrape news sources for alerts and provide real-time driver data. Channels of communication need to be opened between the data producers, who may not see the wider utility of what is collected, and data consumers, who may not understand barriers with data collection. In addition, funding agencies that support driver-associated data collection efforts need to require that recipients follow best data practices. We recognize the existence of obstacles to developing a digital driver surveillance platform, but as data become more available, compatible, and refined, our ability to overcome systematic biases and sources of error to identify driver activity will become easier. This report is a call to action to improve collection of driver and ID event data to rapidly develop the science and our understanding of relationships between drivers and emerging ID events, and move toward driver-based ID surveillance systems.

References for data examples in the Table and URLs for web platforms mentioned in the text.