Digital Pharmacovigilance and Disease Surveillance: Combining Traditional and Big-Data Systems for Better Public Health.

The digital revolution has contributed to very large data sets (ie, big data) relevant for public health. The two major data sources are electronic health records from traditional health systems and patient-generated data. As the two data sources have complementary strengths-high veracity in the data from traditional sources and high velocity and variety in patient-generated data-they can be combined to build more-robust public health systems. However, they also have unique challenges. Patient-generated data in particular are often completely unstructured and highly context dependent, posing essentially a machine-learning challenge. Some recent examples from infectious disease surveillance and adverse drug event monitoring demonstrate that the technical challenges can be solved. Despite these advances, the problem of verification remains, and unless traditional and digital epidemiologic approaches are combined, these data sources will be constrained by their intrinsic limits.

Traditional disease surveillance has been a key ingredient in any public health portfolio for many decades. Disease surveillance is widely recognized as one of the most important tools to assess, predict, and mitigate infectious disease outbreaks. Traditional disease surveillance is based on data collected by health institutions, and the data typically consist of information such as morbidity and mortality data, laboratory reports, individual case reports, field investigations, surveys, and demographic data. They are generally collected by physicians, public health laboratories, hospitals, and other health providers and institutions. The computer revolution that began in the 1970s has affected traditional disease surveillance systems by improving the accessibility of data and by increasing the speed at which data are transmitted between institutions. However, the ongoing Internet and mobile phone revolution has a qualitatively distinct effect: in addition to making epidemiologic data available faster and more broadly, new data are generated directly by the public, often on platforms not primarily designed for health purposes. These data streams of user-generated data are almost always bypassing traditional public health channels. They are the data streams on which digital epidemiology is generally based [1, 2].

One of the first and certainly the most prominent examples of digital disease surveillance was Google Flu Trends [3]. Google Flu Trends was essentially an analytical estimate of the level of weekly influenza activity based on the search queries that Google received. The analytical estimate was derived by a model selected by generating the best fit to the Centers for Disease Control and Prevention's (CDC's) influenza-like illness (ILI) data from a number of different US regions. The original model results obtained a mean correlation of 0.9 with the CDC data. A few years later, in summer 2015, Google decided to shut down the public website of Google Flu Trends and instead opted to give select academic and public health institutions access to the data. This announcement followed numerous reports [4-6] that systematically assessed Google Flu Trends' overestimation of influenza activity, attributing it to a combination of a phenomenon termed “big-data hubris” and algorithm dynamics. The first refers to the assumption that the novel big-data streams are a substitute, rather than a supplement, to traditional data collection efforts. The second refers to the observation that, while the Google search algorithm receives updates on a weekly or even daily basis, the Google Flu Trends model received updates only rarely. This led to a situation where the model did not keep in sync with the changing nature of the data from which it was supposed to generate predictions.

Despite the problems of Google Flu Trends, the system was an important example of the promises of digital epidemiology: to use novel data streams, often generated for purposes quite distinct from public health, to extract additional public health signals, such as those relevant for disease surveillance. But while Google makes some search pattern data available through an interface called Google Trends, the raw search-query data that Google Flu Trends was based on is not publicly available. In recent years, two other digital data sources have attracted the attention of digital epidemiologists: Twitter, the popular microblogging services, and Wikipedia, the world's largest open-access encyclopedia. Twitter data are openly accessible through an application programming interface, which allows any third party to stream Twitter data in real time to their own application. Twitter has been extensively used to assess influenza activity [7, 8], but it may principally suffer from the same problems of overfitting [9] and static algorithms. Wikipedia access logs, a public data source, have recently attracted the attention of the research community as a proxy of search engine query logs because Wikipedia pages are often ranked highly in the search engine results following disease-related queries. Early analyses of Wikipedia access logs have shown great promise in providing real-time estimates (so called now-casting) of the prevalence of a number of infectious diseases [10, 11].

While search engine logs, social media posts, and Wikipedia access logs are a few examples of big-data sets that have emerged following the ongoing Internet penetration worldwide, there is also another source of data that is increasingly relevant for disease surveillance—the public itself. In contrast to classical surveillance that reports on a patient's health status once they have accessed the health system, participatory surveillance asks patients to report symptoms and other data directly online. Web-based participatory surveillance systems have shown great promise in the case of influenza, for example. In Europe, the Influenzanet project has been successfully collecting data on ILI activity in a number of European countries [12], and in the United States, Flu Near You has emerged as a leading crowdsourced influenza surveillance system, and there are more in other parts of the world [13]. Given the widespread use of smartphones with broadband Internet access worldwide, we can expect many more participatory public health applications in the near future, complementing traditional surveillance systems.

Importantly, because many of the data streams of digital epidemiology have not been generated for the disease surveillance niche, much broader insights can be gained from these data sources. While much of the earlier work on digital epidemiology has focused on user-generated description of symptoms, later work has increasingly focused on the analysis of health behaviors and sentiments/opinions, particularly as they relate to infectious diseases. For example, Twitter data have been mined for signals of vaccine sentiments to estimate vaccine uptake rates. During the 2009 influenza A(H1N1) pandemic, vaccination sentiments measured on Twitter correlated positively with prospectively reported vaccination uptake rates across US states [14]. This indicates that these new data streams can help in the public health decision-making process, because sentiments expressed on Twitter can be measured in real time, giving those in public health practice early warning signals of possibly undervaccinated populations. Later work on the same data set investigated how negative and positive sentiments about vaccination spread across the social network, suggesting that negative sentiments are more susceptible to social contagion than positive sentiments [15]. Last but not least, data from most of these services are increasingly generated on mobile phones and other devices, increasing the probability that high-resolution geographic information is associated with the data, a phenomenon that will become increasingly important, given the spatial dynamics of disease spread.

DIGITAL PHARMACOVIGILANCE

The widespread use of the Internet and of social media in particular has had a dramatic effect not only on infectious disease surveillance, but also on the surveillance of drug use and related events. Perhaps even more so than traditional infectious disease surveillance, traditional surveillance of adverse drug reactions (ADRs) after drug use is slow and patchy. When reported by patients or healthcare professionals, ADRs are typically assessed by drug experts and pharmaceutical companies, and the results are then passed on to government agencies. This leads to substantial data loss and delays. A recent study in the United States showed that hospital staff did not report 86% of ADRs among patients [16]. The rate of underreporting in nonclinical settings is arguably even higher. Once government agencies receive the reports, they often release them with a delay of months or even years. The lack of speed and broad coverage has multiple causes, including the fact that a proper assessment of ADR data is both imperative and time-consuming; it is nevertheless in direct contrast with the public health importance of ADRs. In the European Union alone, ADRs are the cause of 5% of all hospital admissions and are responsible for an estimated 197 000 yearly deaths [17].

Public ADR reporting systems are largely unknown to the public, despite long-term governmental support. A recent study in Australia reported that only 10.4% of the general population was even aware of the national ADR reporting system [18]. This low awareness was comparable to the results reported in an earlier study in the United Kingdom, where only 8.5% of the adult population was aware of the United Kingdom ADR reporting system [19]. Among physicians, ADR reporting has been declining over time in both countries. Such declining reporting by physicians has been linked to ignorance, diffidence, lethargy, and insecurity (sorted here by decreasing frequency associated with not reporting ADRs, as identified elsewhere [20]), leading some to suggest that physicians should get paid to report ADRs [21].

While consumers rarely use official ADR reporting systems, they increasingly use online platforms to investigate potential ADRs. Health-related interests are now a major driver of Internet use [22]. When experiencing a potential ADR, consumers can now easily search the web to look for information about a potential connection between their symptoms and the drugs they are taking. Indeed, for the purpose of mining digital consumer data for an ARD signal, the patient does not even need to be conscious of the link between drug intake and symptoms, as long as they can be correlated in the data. Social media are also increasingly used to share ADRs with others. Both digital traces left behind as a consequence of these online activities can be used for digital pharmacovigilance. By mining and analyzing search logs or social media posts for ADRs, signals may be detected much faster than through the traditional ADR reporting systems.

A recent study by White et al [23] exemplifies the idea of pharmacovigilance through search logs. Using a 2011-reported adverse event (hyperglycemia) due to a previously unknown interaction between the drugs paroxetine, an antidepressant, and pravastatin, a cholesterol-lowering drug, the question was whether the adverse event could have been detected earlier by using search-log analysis. By mining through millions of search queries on Google, Bing, and Yahoo Search from 2010 (provided to the researchers anonymously by users who opted in to share their search history), White et al found that people who searched for both drugs were also more likely to search for terms related to the adverse event than those who searched for only one of the drugs. The study was done after the ADR had been identified, and using this approach for the identification of unknown ADRs will remain a challenge. A later study by some of the same authors [24] demonstrated that jointly leveraging data from the Food and Drug Administration's (FDA's) Adverse Event Reporting System (FAERS) and search logs could improve the identification of ADRs by 19%, compared with use of each data source alone. This improvement corresponded to the proportion of error reduction gained by using the combined signals over the better-performing individual data source, as measured by the difference in area under receiver operating characteristic curve.

Social media services are increasingly becoming online places where people share possible ADRs. Freifeld et al [25] used Twitter, the popular microblogging service, as a data source to assess the feasibility of digital pharmacovigilance through social media. Using Twitter posts (termed “tweets”) in English language mentioning medical products, they identified possible ADRs with a combination of manual and semiautomated techniques. The aggregate frequency of possible ADRs was then compared to FAERS. From 6.9 million tweets collected between November 2012 and May 2013, Freifeld at al identified 4401 possible ADRs, and although the comparison of possible ADRs from Twitter to those from FAERS at the preferred level was not possible because of the “Internet vernacular on Twitter,” [25, pp 347] the rank order correlation by system organ class was relatively high, with a Spearman rank correlation coefficient of 0.75 (P < .0001).

Focusing on a more specific drug type, Adrover et al [26] analyzed ADRs with respect to drugs for human immunodeficiency virus (HIV) infection, using a data set of >40 million tweets containing HIV drug names collected over 3 years. They used a combination of crowdsourced human assessment and machine-learning algorithms to identify the tweets of individual reports about ADRs with HIV drugs such as Atripla (Gilead and Bristol-Myers Squibb) and Truvada (Gilead, Foster City, California). The remaining 1642 tweets represented ADRs from single drugs or drug combinations and captured well-recognized toxicities known from clinical practice. For example, efavirenz-containing treatments (eg, Sustiva [Bristol-Myers Squibb, New York City, New York] and Atripla) were often reported in conjunction with sleep-related problems, such as nightmares or lack of sleep, a phenomenon well-documented in the clinical literature [27]. The study also analyzed the sentiment expressed in these tweets, which was mainly but not always negative, highlighting a benefit of social media studies over search query analysis. Tweets, despite their limitation of 140 characters, can still convey much more information than a search query, containing valuable information that can put potential ADRs in a specific and possibly relevant context.

Another source of user-generated content on the Internet are health forums. Leaman et al [28] mined data from the website DailyStrength and manually annotated 3600 posts relating to 4 drugs, carbamazepine, olanzapine, trazodone, and ziprasidone. The ADR incidence rates for these drugs is well established by the FDA, and the study showed that there was a strong correlation between those well-established incidence rates and the rates derived from the annotations generated from the user-generated posts. The authors also developed an automated system to identify adverse reactions by means of a primary lexical method, using 450 of the 3600 posts. When they evaluated the system against the 3150 posts not used for system development, they found that it performed well, with a precision of 78.3% and a recall (sensitivity) of 69.9%. Chee et al [29] conducted a study with a less constrained data set consisting of 27 290 public health and wellness groups on Yahoo. They used a natural language processing (NLP) approach to identify drugs that were withdrawn from the market. The identification was based on an NLP classifier trained on forum posts, allowing for further prediction of drugs that may be candidates for market withdrawal.

These studies are only a few examples of a growing literature aiming to detect ADRs through nontraditional data streams of patient-generated data. Such digital pharmacovigilance has the potential to strongly supplement pharmacovigilance based on traditional ADR reporting systems. At the same time, this new approach comes with its own set of challenges. First, access to data is often difficult or at times impossible. Given the widespread global use of Facebook, for example, ADR reports on Facebook would likely be a tremendous resource for pharmacovigilance, but the data are not accessible to the public or to researchers. Even with easier data access as provided by Twitter or scrapable websites, the terms of service of these platforms often prohibit access to the full data set and the sharing these data with others to verify and replicate results.

The challenge posed by the question “Are privately held data accessible for public health research?” is one of many ethical challenges surrounding the use of big data for public health. Vayena et al [30] have identified a number of challenges surrounding digital epidemiology that are directly applicable to digital pharmacovigilance, as well. For example, issues of methodologic validation are highly pertinent in the context of ADRs: false predictions of potential adverse events may drive substantial spending of limited public health resources. Algorithmic claims of undocumented adverse effects may quickly sway public opinion in one way or another and, if the claims turn out to be wrong, might potentially taint otherwise safe drugs with a bad reputation for a long time. Of course, this problem is not limited to algorithmic suggestions alone: the now widely debunked claim that the measles, mumps, and rubella vaccine may cause autism, for example, was based on a fraudulent study whose failings were entirely noncomputational. Nevertheless, today's availability of cheap computational power substantially reduces the ease with which algorithmic claims can be made, and we therefore need to think about a system that can weigh these claims in a way that is both scientifically sound and remains open to anyone.

CONCLUSIONS

The emergence and subsequent public withdrawal of Google Flu Trends has illuminated two potential key problems of digital epidemiology: big-data hubris and algorithm dynamics [6]. As the examples mentioned above have shown, there is tremendous potential for epidemiology in these novel data streams that have emerged during the growth of the Internet and the widespread use of smartphones. Nevertheless, these data streams are conducive to big-data hubris, a situation where these new data streams seek to supplant traditional data streams, rather than supplement them. In this context, it is worthwhile to note that the original authors of Google Flu Trends warned that “this system is not designed to be a replacement for traditional surveillance networks or supplant the need for laboratory-based diagnoses and surveillance” [3, pp 1013]. Despite traditional epidemiology's shortcomings, it is ultimately the generator of ground-truth data against which novel, digital systems need to be validated. It will be prudent of the public health community to build on the strengths of both systems—veracity in traditional epidemiology and velocity and variety in digital epidemiology—in conjunction [31]. At the same time, traditional public health systems need to integrate novel data streams into their work flow and provide the corresponding infrastructural investment. Algorithmic intelligence in the public health domain needs to adjust to changing conditions all the time. The incentive structures in most of academia (frequent publication of novel findings) are at odds with the requirements of building long-term systems with dynamic algorithms that need to be maintained and updated regularly. The call for leveraging digital epidemiology intelligence for public health is therefore strongest when it comes from within the existing public health institutions.