The Evolution of Big Data in Science

All this new dependency on data, and especially on big data, may seem like an abrupt change from business as usual in risk communication, but in fact it represents a natural tipping point from a trend that began decades earlier. Its origins can easily be traced to the early days of computing, when “mainframe” information systems first made it possible to input thousands of individual data points (from tapes or punch cards) into electronic memory, against which a complex set of data-processing commands—also known as programs—could be executed. It was this early capacity in large-scale data processing that revolutionized business and expanded the horizons of science. The social sciences naturally began to adapt and expand their capacity to perform multivariate statistical analysis to accomplish in minutes what would have taken weeks to calculate by hand. Then, with the introduction of the silicon-based semiconductor, this capacity to process large amounts of data in increasingly shorter amounts of time began to expand. Intel cofounder Gordon Moore famously predicted that with all the advances that he was observing in the semiconductor industry, computing power would continue to grow at an exponential rate for the foreseeable future—essentially doubling every two years. This audacious prediction, known colloquially as “Moore’s Law,” turned out to be surprisingly prescient. In the 1980s, much of the power of mainframes had migrated to scientists’ desktops through personal computers.

While the 1980s saw the rise of the personal computer, the 1990s and early 2000s saw the emergence of mass connectivity through diffusion of the Internet and its user-friendly architecture for collaboration, which became known as the World Wide Web. In this newest phase of the data revolution, technology publisher Tim O’Reilly predicted that data would become the new “Intel Inside” for the next generation of Internet technologies (O’Reilly, 2005). These new technologies would effectively turn the Internet into a large platform for collective participation, he argued, enabling a new era of “open science.” In 2007, Steve Jobs introduced the smartphone to the global market. Expanding on Moore’s Law even further, smartphones funneled the computing power of the old computational mainframes into a device that could sit in a scientist’s pocket as it opened new channels of data flow into remote servers through what would come to be known as the “cloud” (i.e., placing data stores and processing speed into collectively accessible remote servers rather than depending solely on local storage or processing).

Data flows would burgeon further with the introduction of tablet computers, wearable sensors, smart devices, and now—in the age of the “Internet of Things”—smart objects. These converging data flows allow a new paradigm in analytic techniques, allowing high-tech companies such as Google, IBM, Microsoft, and Amazon to improve the intelligence of their respective business systems in real time by harvesting the flow of user data to improve the performance of their search algorithms and personal digital assistants.

The emerging abundance of data streams—catalyzed by Moore’s Law and widened exponentially by the parallel storage capacity of millions of data-recording devices—led Microsoft data architect Jim Gray to conclude that science was entering a “fourth paradigm” in its approach to discovery and knowledge generation (Hey, Tansley, & Tolle, 2009). The first paradigm, he argued, was systematic adoption of the experimental method, which allowed focused, disciplined investigation based on a priori logic of inferred causality. The second was the evolution of theoretical science, emphasizing coherent frameworks for the interconnections among empirical observations. The third was the introduction of computing, which expanded the reach of empirical observation and enabled the application of advanced statistical techniques through the efficiencies afforded by computational systems. The fourth paradigm was the introduction of connected data streams into a data-rich environment, essentially developing science from its historical roots as an individual endeavor into a contemporary approach enabled by cutting-edge platforms for mass participation.

It has been within this context that the popularized notion of “big data” came about. Like any term de jour, the concept has often been misunderstood and subject to hyperbole (Parks, 2014). Engineers at IBM systems characterized the concept in terms of four dimensions, something they referred to as the “four V’s” of big data. The first was “volume,” or simply the unprecedented amount of data stored in digital format. By some estimates, the amount of digital data generated by the world’s computing capacities may exceed 2.5 exabytes (2.5 x 10¹⁸ bytes) being added to the global store daily. The second was “variety,” with data inputs available for analysis from software logs, remote sensing devices, cameras, radio-frequency identification, satellite telemetry, microphones, and health informatics infrastructures. A third dimension was “velocity,” with a recognition that broad streams of data can be made available for analysis in real time. The fourth and final dimension was “veracity,” as data scientists struggle with the uncertainty and value of information computed from imprecise sources. These four aspects aptly characterize the tenure of data scientists in public health and medicine who have begun to exploit the power of a data-rich environment to accelerate success in reducing the impact of disease and trauma in the information age.

Big Data in Public Health and Medicine

When the archetypal epidemiologist John Snow gathered systematic observations on outbreaks of cholera in the SoHo district of London in 1854, he was setting into place the beginnings of modern public health surveillance systems (Johnson, 2006). During the 20th century, governments utilized local and national surveillance data sources to inform policy, to stay ahead of disease outbreaks, to identify population trends, to explore hypotheses for research, and to monitor businesses for regulatory adherence to public health safety standards. Some of the data from these sources were extracted directly from public records and compiled through census techniques to gain a better understanding of the incidence and prevalence of disease. These archival sources were soon complemented by data extracted from national probability samples to gain a more comprehensive understanding of how physical environment, social context, lifestyle choices, and behavioral predilections might influence the onset, course, and sequelae of disease (Hesse et al., 2013). Initially, these surveillance programs tended to be run as sequestered efforts, collecting and processing data for the primary benefit of a single agency or sovereignty. Efforts have been underway internationally through the World Health Organization (WHO) to harmonize and share data collected through these systems to enable a more coordinated system of global health tracking. With troves of public health data being collected within the United States alone, suggestions have been made to utilize the power of big data analytics to expand the breadth of analyses across complementary datasets for the purposes of improving health situational awareness among all stakeholders (Hesse, 2011; Thacker, Qualters, & Lee, 2012).

In 2001, the U.S.-based National Committee on Vital and Health Statistics took stock of the data flows connecting decision-makers in public, clinical, and personal health. From their evaluations, it concluded that work would be needed to turn these traditionally paper-based data systems into machine-readable, electronic streams that could be integrated for the public good. The committee outlined an ambitious blueprint for creating an electronic, national health information infrastructure in the United States (National Committee on Vital and Health Statistics, 2001). In a follow-up to the report, U.S. president George W. Bush announced a national goal in his 2004 State of the Union address to Congress and the public to connect the majority of Americans to electronic health records (EHRs) within a decade. In support of that goal, he appointed a National Coordinator for Health Information Technology to oversee the effort.

By 2009, some progress had been made toward achieving that goal, but because of lack of market incentive, adoption was slow. To boost adoption, Congress passed the 2009 Health Information Technology for Economic and Clinical Health (HITECH) Act, featuring provisions for offering financial incentives to hospitals and individual medical practices that demonstrated the meaningful use of Health Information Technology (Health IT) for the improvement of patient care. Adoption rates began to rise precipitously. By 2015, data suggested that 96% of nonfederal acute care hospitals had adopted a certified EHR (Henry, Pylypchuk, Searcy, & Patel, 2016).

Although adoption rates of data-driven health information technologies had risen by 2015, the ability to pool data across healthcare systems remained obstructed. In testimony before Congress, professional medical societies pointed to data blocking—the practice of limiting the export of data to shield proprietary value—as the main culprit. In December 2016, Congress passed the 21st Century Cures Act, which prohibited data blocking and authorized penalties of up $1 million per violation. On the technological front, the Office of the National Coordinator pointed to the value of application programming interfaces (APIs), to enable the flow of data between legacy EHR systems and fit-for-purpose applications to support clinical care and patient self-management. Investments in demonstration projects illustrated how data tools and infrastructures could be built on top of the existing EHR systems to catalyze a thriving ecosystem of consumer-focused apps, much in the same way as mobile smartphone and tablet operating systems allowed development within their respective app stores (Mandl & Kohane, 2016).

One of the ultimate targets of this interoperability development is to improve data liquidity among practice components, and by so doing, to enable the creation of a learning healthcare system (Olsen, Aisner, McGinnis, & Institute of Medicine Roundtable on Evidence-Based Medicine, 2007). The vision of a learning healthcare system is to create an environment in healthcare, fueled by data, that could (a) be leveraged for quality improvement, (b) be utilized for comparative effectiveness research, (c) inform real-time decision-making, and (d) serve as the springboard for new biological and behavioral discovery. Inputs would come from structured fields within EHRs, unstructured fields analyzed through natural-language-processing algorithms, imaging data, laboratory data, genomic-sequencing data, sensor data, and even data generated by patients equipped with mobile accelerometers, blood glucose sensors, cardiologic sensors, and the like. The resulting flows created by such interoperable systems would naturally be of high volume, high velocity, and high variety, and—to serve medical needs—would have to be evaluated for high veracity as well. In 2013, the director of the U.S. National Institutes of Health (NIH) appointed the agency’s first associate director for data science to supervise the research needed to harmonize these data streams and extract scientific value through advanced analytic techniques. One of the first initiatives funded by the associate director’s office was the “Big Data to Knowledge” program, designed to create centers of excellence in data science.

A Big Data Toolbox for Risk Communication Scientists

With these developments as prelude, communication scientists now have an expanded toolbox with which to exploit the value of big data for new knowledge in health and risk communication. Hindman (2015), for example, noted that for years, the overriding analytic paradigm in the social sciences has been dominated by a use of ordinary least squares (OLS) regression techniques. OLS approaches, he points out, are subject to weaknesses and limitations that can be overcome with some of the newer machine-learning approaches being tested within the field of computer science. To begin with, standard OLS techniques are poor compared to other techniques when it came to predicting values beyond the sample from which the regression coefficients were generated. Because the models tend to be poor predictors, they also tend to be less replicable than other techniques. Many of the machine-learning approaches coming out of the big data revolution call for segmenting a robust, large scale dataset into two sets. One set, called the training set, would be used to build the predictive model. The other set, called the test set, would be used to evaluate the performance and replicability of the predictive model. Computer scientists would then use an ensemble of complementary modeling techniques on the training set to ferret out which model, or combinations of models, was best at predicting quantitative relationships reliably and efficiently in the test set. The overall goal is to improve the explanatory power of the resulting statistical model.

Detractors of these machine-learning approaches often argue that data-mining techniques constitute an atheoretical, computational exercise that needs little in the way of deep subject matter understanding of the data being modeled. Nothing could be further from the truth, observed Hindman (2015). Contemporary data-mining techniques require as much of an understanding of the underlying data for interpretation of the resulting models as do linear regression or logit techniques. Moreover, because big data methodologies necessarily value parsimony as a cardinal principle, the resulting models can be integrated back into a conceptual understanding of the underlying phenomenon in straightforward ways—something that stands in contrast to the overly complex OLS models built from a kitchen-sink approach to variable inclusion. On the other hand, it remains true that for big data approaches—as in all empirical approaches—the same principles of rigorous scientific understanding must apply. The following caveats are worthwhile to bear in mind:

Open-science and data-sharing advocates have reasoned that under the right circumstances, big data approaches could help address some of these issues. Preregistration of hypotheses and variables can serve to limit a posteriori fishing, while data sharing can inform meta-analyses and replicability studies (Ioannidis, Munafo, Fusar-Poli, Nosek, & David, 2014; Nosek et al., 2015). Machine-learning techniques, as described earlier, emphasize replicability as a critical dimension of analysis (Hindman, 2015).

Big Data and the Public Web

Another area in which big data have come into play relevant to risk communication is through the open environment of the public Web. To be sure, marketing companies have been utilizing big data collected from search logs, clickstream data, persistent cookies, and automated Web crawlers to anticipate the needs and predict the behavior of their customers for years. In 2008, the online search giant Google made headlines when it demonstrated how an analysis of spikes in its search data for flulike symptoms could serve as an early warning system for localized flu outbreaks. Although the algorithm for predicting flu outbreaks underperformed in comparison to official statistics released by the Centers for Disease Control and Prevention (CDC) in subsequent years, many public health scientists have continued to explore improvements to the underlying statistical models. A promising area of improvement has been through the use of Autoregressive Integrated Moving Average (ARIMA) approaches to control for temporal dependencies. Public health scientists have successfully utilized big data approaches to improve predictions of outbreaks in influenza, Ebola, dengue, and Zika (Yang, Santillana, & Kou, 2015).

Big data analytics also have been applied to the task of mining public Web resources about noninfectious diseases. Tourassi and her colleagues at the Oak Ridge Laboratories, for example, utilized Web-mining software and advanced natural-language-processing methods applied to online obituaries to replicate incidence and mortality data from the National Cancer Institute’s Surveillance Epidemiology and End Results (SEER) program (Tourassi, Yoon, Xu, & Han, 2016). Soliman, Nasraoui, and Cooper (2016) used an advanced text-mining approach to extract information about gene interactions related to glaucoma from the National Library of Medicine’s online PubMed database, and then used that data to benchmark an interaction network associated with the disease. Yoon and Bakken (2012) demonstrated how Web-mining techniques could be applied to the microblogging context of tweets to uncover mental models related to physical activity, and Chan, Lopez, and Sarkar (2015) investigated the nonmedical use of opioids using similar efforts. Cappella, Yang, and Lee (2015) modeled a recommendation system for tailoring messages to smokers based on the high-density data algorithms used by commercial content providers such as Amazon and Netflix.

The public Web has also served as a platform for mass participation in knowledge creation and data generation. The public site PatientsLikeMe is one such platform, where patients with a specific diagnosed disease can go online to compare the progress of their treatments with those of patients who are similarly diagnosed. With more than 300,000 registered users in 2015, the site encourages its members to share personal data over the course of their treatments so that biomedical researchers and pharmaceutical manufactures can collect the data that they need to uncover side effects and develop new hypotheses on targets for treatment. In a similar way, the U.S. Food and Drug Administration (FDA) opened its Sentinel database to record adverse events from patients through social media feeds, and by so doing, was able to expand its coverage of reported side effects (Ball, Robb, Anderson, & Dal Pan, 2016). Each of these examples represents a type of crowdsourcing, a term coined to describe the multiplying effect of asking large groups of people to contribute solutions to a problem through parallel inputs on the Web. The Amazon Mechanical Turk platform, which allows individuals to perform online tasks for negotiated amounts of money, has become a mainstay of crowdsourcing for many academic researchers.

To understand more about how large groups of individuals can help accelerate science by contributing data in parallel, the National Science Foundation has invested heavily in citizen science projects. In one such project, laypersons with asthma volunteered to use specially equipped smartphones to monitor air quality in their local environment. Geopositioning data affixed to the transmitted data streams allowed researches to locate readings in real time on a virtual map. The volunteers benefited immediately by learning more about the environmental quality of locations in which they might find themselves, while the donated data streams could be analyzed through big data analytics for geospatial and temporal trends.

The Robert Wood Johnson Foundation also has invested in understanding how crowdsourcing principles can be brought to bear on accelerating discovery in the context of health. The foundation has coined the phrase “data for good” to describe how individuals imbued with a sense of altruism can contribute data from personal monitoring devices with the purpose of furthering research on health promotion. Data altruism lies at the core of many of the large biomedical research initiatives being funded by the NIH. One of these initiatives, termed the “All of Us Research Program,” will solicit data contributions from patients over their lifespans using patient portals, EHRs, and mobile monitoring devices. The thrust will be to collect the data necessary to calibrate treatments more precisely to patients’ genomic and phenotypic profiles. In this sense, the program is intended to be the platform upon which NIH will pursue its goals in the much-publicized area of precision medicine (NIH, 2016).