The progression of genomics in the last few decades has been remarkable. Since 2001, when the Human Genome Project mapped and sequenced virtually every gene in the human genome, genetic sequencing technology has advanced rapidly in both the public and private domains. Next-generation sequencing costs have plummeted by almost 100%, and research opportunities have grown exponentially as a result [1]. For example, a simple search in the medical database, PubMed, shows that research on genomics has more than quadrupled since 2000, from around 340,000 published articles on this topic growing to 1.5 million by 2020. This increase has translated into quicker diagnoses, better outcomes, and more effective health care for patients [2-4]. Great strides have been made in cancer research, for example, where patients are now being treated according to their own or the tumor’s genomic data [5].

Being able to use genomic data in conjunction with routinely collected data holds even greater potential to advance knowledge by including factors wider in scope. Precision medicine requires that novel correlations of genotype, phenotype, and the environment be identified to inform new methods for diagnosing, treating, and preventing disease in a way that is responsive to the individual [5]. Knowledge of gene-environment interactions can also contribute on a population level by informing health and public health services in areas such as service planning, population genetic testing, disease prevention programs, and policy development [6]. Routinely collected data, electronic health records (EHRs) in particular, already hold vast amounts of clinical and environmental information on large numbers of people and preclude the need for lengthy and expensive data collection. Adding these phenotypic data to knowledge about a person’s genome can elucidate new knowledge, such as that about gene-environment and gene-drug interactions, and can thus provide a richer understanding of health and disease [7].

Increasing the use of genomic data for health research is a high government priority to understand and address the causes of disease. The potential of integrating genomic and routine data sets has been recognized in the United Kingdom by the Welsh Government [8] via their Genomics for Precision Medicine Strategy and by Genomics England [9] with the inception of the 100,000 genome project in 2018, both of whom are investigating ways to link genomic data and EHRs. From a more international perspective, the former president of the United States, Barack Obama, launched the Precision Medicine Initiative to improve individualized care by combining genomic data and EHRs with diet and lifestyle information from US citizens [10]. The UK Chief Medical Officer, Dame Sally Davies’s “genomic dream” of mainstreaming genomic medicine into National Health Service (NHS) standard care is becoming ever closer, which means that data from a person’s genome will likely be directly recorded into EHRs, making this type of research far more accessible [11].

As this field prepares to move forward, it is important to take stock of the current state of play in order to highlight new avenues for development, identify challenges, and ensure that adequate data governance models are in place for safe and socially acceptable progress. Previous work has examined the benefits and logistical challenges of integrating genomic and routinely collected data in health care practice, but less is known of this specifically in a research setting [6,12,13]. Therefore, our objective is to add to this literature by conducting a narrative literature review and a series of interviews that would provide an outline of the use of genomic data alongside routinely collected data in health research to date. It focuses on the types of data that have been used, the role of routinely collected data in these studies, the data sources, how researchers access the data, and the challenges surrounding their use. This will inform further work in developing a framework for working with genomic and routinely collected data [14].

First, we conducted a literature search of research that used genomic data in conjunction with routinely collected data. We define routinely collected data as data collected as a matter of course and not specifically for research [8]. Genomic data refer to the data generated after processing a person’s genome, in full or in part, for example, by sequencing [7]. Studies were eligible for inclusion if they had used both types of data in combination to answer a health research question. We included studies of any design published in either peer-reviewed journals or gray literature in the English language.

We searched the following databases to identify these studies: PubMed, Ovid, CINAHL, OpenGrey, CENTRAL, LILACS, and Web of Knowledge from inception until January 31, 2019. We also searched for books, gray literature, and websites. We used a piloted search strategy that included keywords representing genetic and routinely collected data, and the following is the search strategy used for PubMed and modified for use with the remaining databases:

This search in PubMed resulted in more than 50,000 hits. After initial pilot screenings, for example, by restricting to only publications within the last 10 years, it was clear that, given the number and heterogeneity of potentially relevant articles, neither a systematic review nor a meta-analysis would be practical. We took a pragmatic approach by scanning these articles to retrieve information on the following items until we reached data saturation, when no new information appeared in the text: types of genomic data; types and roles of routinely collected data; data sources; and data access models, that is, how researchers access the data. We chose examples from each criterion to ensure that we included a range of health conditions and presented them in a narrative format, which followed the format of the criteria given above.

To understand the use of genomic and routinely collected data in context, we recruited a purposive sample of individuals who have been involved in leading research projects using a combination of genomic and routinely collected data. We identified potential participants from the literature search outlined above and sent 19 interview invitations via email. Reminders were sent after 2 weeks, and if there was no response, we made no further contact. In total, 11 individuals agreed to participate in either an in-person or teleconferencing interview, depending on their geographical location. Participants were involved with the following projects: the Swansea Neurology Biobank (Swansea, Wales); Dementia Platform UK (Oxford, England); PsyCymru (Swansea, Wales); UK Biobank (Stockport, England); BC Generations Project (British Columbia, Canada); the Province of Ontario Neurodevelopmental Disorders (POND) Network, IC/ES (Ontario, Canada); Electronic and Medical Records and Genomics (eMERGE) Network (Vanderbilt, United States); and the Sax Institute’s 45 and Up study (New South Wales, Australia). We omitted any further information on the participants to maintain their anonymity. We developed interview questions with our advisory board: a group of UK geneticists and data scientists who were interested in using genomic data and our discussions centered around these:

We followed up interviews by email if any answers needed clarification.

Examples found of the types of genomic data used in these studies included single nucleotide polymorphisms (SNPs), gene activity scores, and DNA methylation status. The most frequently used were SNPs [16,17,19,23-29], which represent a single base-pair change in the DNA sequence and are highly granular; hence, they are popular in health research [33]. Research studies tend to refer to SNPs by their reference SNP number, a unique identifier given to a SNP, or a cluster of SNPs by the National Center for Biotechnology (NCBI) [34].

Polygenic risk scores are closely related to SNPs and were used in 2 of the studies included in this review [15,16]. These predict a person’s likelihood of developing a particular disease based on the cumulative effect of a number of genetic variants [15]. Our literature search found other instances of quantitative measures of gene activity combined with routine data, including the 21-gene Recurrence Score, which measures the activity of 21 genes (16 cancer-related and 5 reference) for patients with breast cancer [33] and the enzyme activity score for the CYP2D6 (cytochrome P450 family 2 subfamily D member 6) gene, which codes for an important drug-metabolizing enzyme, and is highly polymorphic in humans [34]. Rusiecki et al [29] measured changes in DNA methylation status in their participants to investigate whether this predicted posttraumatic stress disorder in US military service members.

Conducting genome-wide association studies (GWAS) [15,17-19,21], phenome-wide association studies (PheWAS) [24,25], and a combination of both [26,27] were the main purposes of combining genomic and routine data. Both these methods use powerful statistical techniques to find associations between genetic variants (SNPs) and phenotypes, which can then be used to predict the genetic risk factors of disease, levels of gene expression, and even social and behavioral characteristics such as educational attainment, impulsivity, and recreational drug experimentation [35,36]. Given the number of tested associations, these studies require large sample sizes to yield enough statistical power to detect differences in genetic variation between cases and controls [37]. Other study designs included longitudinal studies [16,18,30-32], a candidate gene study [23], case control studies [19,29], a gene profiling study [22], and an exploratory study [28]. Each of these studies was designed either to identify genetic risk factors of the disease or to investigate drug safety or efficacy (Multimedia Appendix 1 Table S1).

EHRs appear to be, by far, the most common form of routinely collected data used in the studies identified in this review. For the type of research discussed here, data in EHRs have been used to identify eligible participants, for phenotyping, and to provide long-term follow-up on specific health outcomes. These data are collected as a matter of course in health care systems, and depending on their country of origin, EHR content can vary, although usually this digital record will include the patient’s name, address, demographics, medical history, care preferences, lifestyle information (such as diet, exercise, and smoking status), and free-text notes [38]. An example of EHR data used in the identified studies was the International Classification of Disease (ICD) coding. This allows clinicians to record the status of a patient in a standardized way, whether it is for disease, disorder, injury, infection, or symptoms [39]. Hebbring et al [26] used ICD codes in EHRs to identify appropriate cases and controls for their PheWAS study of the HLA-DRB1*150 gene involved in immune regulation. Rusiecki et al [29] used ICD codes to identify cases with a postdeployment diagnosis of posttraumatic stress disorder in their study on the link between gene expression and posttraumatic stress disorder in US military service members.

EHRs were also used to describe and validate a phenotype of interest, which is particularly important for PheWAS studies. The large number of phenotypes used in these analyses need to be clearly defined to ensure that any associations made with genetic variants are precise and replicable [40]. For example, Breitenstein et al [23] used EHRs to define the type 2 diabetes phenotype needed for their candidate gene study based on diagnoses, medications, and laboratory tests. The algorithm used to achieve this was developed by the eMERGE Network (see below for more details on this organization) and has been used successfully many times since [41].

In addition to providing a baseline snapshot of the patient, EHRs are also longitudinal in nature, and this makes them ideally placed to provide long-term follow-up to study participants. An example of this is the Genetics and Psychosis (GAP) study, which looked at a purported association between a variant of the ZNF804A gene and poor outcomes after first-episode psychosis [20]. Using individual-level linkage between EHRs and genotype data, this study followed the clinical outcomes for 291 patients over a period of 2 years, and subsequently found strong evidence for their hypothesis.

Other examples of routine data used in genomic research include disease registry morbidity and mortality records [20,22]. Disease registries collect information on clinical outcomes and care for a specific patient population over time. EHRs often feed data into these registries, but registry data can also include patient-reported outcomes and other biometric data, and therefore provide a more holistic view of the patient than an EHR would in isolation, for example, the UK MS Register [42]. Routine data sets need not be individual-level or person-based to prove useful in genetic research. A case in point is the Scottish Index of Multiple Deprivation, which gives an indication of a geographical area’s socioeconomic deprivation based on data about employment, income, health, education, housing, crime, and access to services [43]. Clarke et al [15] were able to link participants’ postcodes to the Scottish Index of Multiple Deprivation and generate socioeconomic deprivation variables to investigate their association with polygenic risk scores for alcohol dependence.

Combining genomic and routine data does not come without its challenges. Our participants, who are currently working with these data, spoke to us about these different challenges during their interviews, and their experiences are summarized below.

Using genomic and routinely collected data holds great potential for health research. The genomic data we identified in our literature review included SNPs, polygenic risk scores, and gene activity scores. Routine data primarily consisted of EHRs, but we did find other routine data types including registry data and deprivation indices that had been combined with genomic data. This paper shows how genomic research has progressed in recent years—from basic GWA and PheWA studies—to more complex methodologies in which health records are linked to genomic data at the individual level. Associations between genetic variants and phenotypes, identification of drug targets, knowledge about drug toxicity, and effectiveness can all be studied from the combination of genomic and routine data and leveraged for public benefit.

The EHRs created during routinely collected care provide the large sample sizes needed for GWAS analysis without the need for costly and time-consuming prospective data collection. These larger samples allow for greater validity, generalizability, and yield adequate statistical power, while also minimizing participant burden and reducing attrition during follow-up [64]. In addition, these data sets are not limited to a circumscribed number of phenotypes in the way that a traditional research study might be [40]. This richness and diversity of EHR data means that the large number of phenotypes used in PheWAS can be very clearly defined (referred to as deep phenotyping), which enhances the accuracy of phenotype-genotype associations [40].

Most importantly, with regard to EHRs, perhaps, is the additional information that routinely collected data provide about an individual’s environment. This can include lifestyle factors, education, work history, pollution, and even traumatic events. Genetic determinism has long been rejected, and we now widely accept the powerful influence that lifestyle and the environment have on the way that our genes are expressed [65-67]. Precision medicine requires novel correlations not only between genotype and phenotype to be made but also with an individual’s environment to inform new methods for diagnosing, treating, and preventing disease in a way that is tailored to that individual. Linking genomic data to routine data promises to elucidate important findings for precision medicine research, which will enable researchers to understand the relationship between an individual’s genome and their complete life course [68]. More of these types of studies are needed for precision medicine to reach its full potential, to make the intricate genotype-phenotype associations needed to advance the understanding and treatment of disease.

From our interviews, we identified 3 main ways that researchers can access genomic and routine data: publicly available on the web, released to researchers, and via a data safe haven. We also observed that biobanks and databanks seem to be moving toward renouncing a data release model and instead favoring a data safe haven approach. Aside from potentially solving data storage issues, this will also help assuage privacy and governance concerns. Reidentification and disclosure from genomic data is possible [69,70] and can lead to many undesirable consequences: discrimination by health or life insurance companies, societal stigma, and the discovery of a genetic predisposition to a condition when one does not want to be told. This is complicated further because of the familial nature of genes, and disclosure could cause some devastating effects for biological relatives as well [71-73].

Despite this, we must keep in mind that simply because genomic data are unique, this in itself does not render it identifiable. Rather, the reidentification risk of genomic data depends on the way they are accessed, the type of analyses that are conducted, and the format in which the results are finally published [70]. This means that decisions to use certain data access models based on practicality and decreased costs are not sufficient. Genomic research relies heavily on human participation, and the public should be consulted to inform the way in which their data are accessed [74]. There is a plethora of activity taking place to consult the public about health research in general, and the success and acceptability of large-scale data research is owed, in part, to the extensive public engagement activities that have been taking place [75,76]. Organizations such as the Global Alliance for Genomics and Health [77] are engaging with the public on the use of genomic data in research, but there is still more work to be done regarding the use of genomic and routinely collected data [78].

This review was not intended to be exhaustive or systematic; therefore, it does not include all health research studies that have used genomic and routinely collected data. We also excluded biobanks and databanks that do not house routine data; therefore, we only include examples here. This may have resulted in inadvertently excluding some countries and institutions where this research takes place and the types of genomic and routinely collected data that have been used. Not implementing systematic methods in study selection may have introduced bias in this review’s conclusions; however, the pragmatic approach used here was deemed sufficient to meet our objectives.

We included 19 studies in this review, only one of these involved data and research from a non-Western country (China) [21]. We did not come across any research that had taken place in low- to middle-income countries, although as this review was not conducted systematically, we may have unintentionally overlooked these. However, the absence of any studies from such countries may be, as Tekola-Ayele and Rotimi [79] explain, a result of having a limited number of well-trained genomic scientists and poor research infrastructure, and due to a less well-established routine data collection infrastructure and procedures such as EHRs [80].

Qualitative interviews may be subject to recruitment bias, which means that some viewpoints and experiences were excluded. In addition, there is limited information about participants; however, as this is a relatively small field of research, it was deemed necessary to maintain participants’ privacy.

Given the projected increase in the availability of genomic data, the potential to be obtained from its combination with routine health data is vast. This review has shown examples of what has been done in this field so far with, for example, GWAS and PheWAS plus other study designs. For fields such as pharmacogenomics, these methodologies need to be used further, where using routinely collected data will simplify the process of tracking longer-term outcomes of personalized medical treatments, and elucidate new findings on the effects of the environment on drug-gene interactions. Our take away from this study is that there are several challenges involved in using these data, particularly surrounding privacy. Therefore, it is imperative that appropriate data governance be documented and that public engagement activities take place to ensure socially acceptable practices. Overcoming these barriers will ensure that the use of these data to progress health research can be exploited to its full potential.