Main

The use of diagnostic gene tests within clinical care has risen rapidly in the United States as the cost of genotyping drops precipitously1 and new research supports the value of testing.2 Pharmacogenomics is poised to experience similar growth as many routinely prescribed drugs now have increasingly well-validated relationships to adverse events or reduced efficacy when gene variants are present.3,4,5 In addition, genotyping technologies have advanced to the point that panel assays involving hundreds of genes are economical, raising the prospect of testing patients once and using stored genomic data repeatedly over a lifetime. With 119 US Food and Drug Administration–approved drugs currently including germline or tumor pharmacogenomic information in their labels, the potential for a patient to be exposed to a drug with published pharmacogenomic associations is significant. We have previously demonstrated that the opportunities to use variants from a pharmacogenomic panel test are high, with 65% of ambulatory-care patients followed longitudinally at our institution exposed to at least one medication with an established pharmacogenomic association within a 5-year time frame.6

The promise of translating pharmacogenomics to clinical practice is highly dependent on the ability to communicate the value of genomic data to practicing clinicians and to manage genomic data across a fractured care-delivery system.7 The use of health information technology (HIT), including electronic health records (EHRs) and clinical decision support (CDS), is considered indispensable. However, there is little published experience on how to best apply these technologies to clinical pharmacogenomics.8,9 Several consortia funded by the National Institutes of Health are filling the gaps. The Clinical Pharmacogenetics Implementation Consortium has defined and published best practices for knowledge management and CDS. Clinical Pharmacogenetics Implementation Consortium recommendations are extensively annotated, supported with graded evidence, and freely available.10,11,12,13,14 In addition, two multi-institute consortia, the Translational Pharmacogenomics Project of the Pharmacogenomics Research Network and the Electronic Medical Records and Genomics Network (known as the eMERGE Network), are actively piloting efforts to integrate genomic information with EHRs, both to facilitate translation of pharmacogenomics to the clinical setting and to capitalize on the wealth of clinical data contained in the EHR for research.

Vanderbilt University Medical Center (VUMC) has established a large pharmacogenomic program known as PREDICT (Pharmacogenomic Resource for Enhanced Decisions in Care and Treatment).15 PREDICT is based on the principles that pharmacogenomic testing should be preemptive and harness HIT to facilitate ordering, storage, and timely dissemination of genetic results at the point of care. The design and implementation model presented here arose out of a 4-year development process to adapt a largely locally developed EHR16 to enable the maintenance, interpretation, and distribution of panel-based pharmacogenomic data to a broad base of providers and patients ( Figure 1 ). For this article, the term EHR is inclusive of all clinical information systems that manage or manipulate genomic information while serving clinician information needs. In addition, we include a brief description of the connection between the EHR and the personal health record (PHR). We believe our experiences can inform adaptations of both locally developed and commercial EHRs for pharmacogenomics.

Figure 1
figure 1

PREDICT EHR development timeline. PREDICT has undergone a 4-year process of design, implementation, and iterative refinement. Several milestones, including new drug–genome interaction implementation as well as high-impact EHR design features, are highlighted. CDS, clinical decision support; DGI, drug–genome interaction; EHR, electronic health record; PHR, personal health record; PREDICT, Pharmacogenomic Resource for Enhanced Decisions in Care and Treatment.

Full size image

EHR Design Principles for a Pharmacogenomics Implementation

PREDICT was established as a quality-improvement program in 2010 to apply clinically significant gene variants designated by the US Food and Drug Administration as pertinent to decisions involving drug selection and dosing.15 EHR features were developed with the expectation that panel-based pharmacogenomic testing will become pervasive, and genomic considerations will routinely influence prescribing. Accordingly, the design of supportive EHR functions has followed 10 objectives ( Table 1 ), which seek to give universal, comprehensible, and timely access to clinically significant genetic variants. Displays of pharmacogenomics results were created to be highly visible, in an effort to prevent priority results from being “buried” among other laboratory data. Preemptive identification of patients who were expected (based on statistical prediction models) to benefit from panel-based gene variant data to tailor future therapies was incorporated into outpatient workflow. All gene variant data were stored long term, but selective, clinically actionable drug–gene combinations that met the burden of evidence for a significant drug–genome interaction (DGI) and attained institutional approval for release, and for which we had developed CDS logic to guide the physician, were promoted to the EHR. The design for disseminating results features a single source for both genetic variant data and genotype to drug phenotype interpretation, reinforcing the consistency and reliability of genotype reporting. Knowledge and data sources were constructed using service-based software architecture such that both genetic variant data and the DGI knowledge base could be easily updated and the updates would propagate to all linked systems. Finally, the EHR mechanisms for reporting the results and delivering CDS were initially designed to serve a small set of targeted DGIs but easily scale to support a large quantity of pharmacogenomic variants.

Table 1 Design objectives for a pharmacogenomics-enabled electronic health record
Full size table

A Locally Developed EHR Perspective

Early in the course of designing the translation of genomic medicine to clinical practice, the biomedical informatics and genomic professions projected the need to store, manage, interpret, present, and share genetic results.17 Institutions with locally developed clinical information systems are well suited for pharmacogenomics implementation because they wield greater control over the underlying architecture and interoperability of their HIT as compared with institutions with vendor-installed systems. Historically, “homegrown” EHRs have been recognized for providing a test bed for new HIT ideas, for evaluation of clinical effectiveness, and for providing proof-of-concept implementations for the wider informatics community.18,19,20,21 In addition, locally developed EHRs have the advantage of access to an engaged user base with whom on-site developers can work directly to obtain feedback and produce iterative improvements that can refine usability and features. However, as clinical information needs have expanded, the financial and human capital required to create, maintain, and certify locally developed EHRs can become daunting, even for large academic medical centers and integrated health systems; few centers have maintained this capability, even with decades of previous investment in technical infrastructure and programming. A second potential disadvantage is that development work is not easily shared or exported, reflecting the current monolithic model of EHRs, in which large ecosystems of HIT from a single vendor or institution are interoperable internally but have limited facility to interact externally. For example, three of the eMERGE pharmacogenomics implementation sites have partial or full development efforts on site, yet these are proceeding relatively independently because of the difficulty in standardizing EHR implementations. The eMERGE and Translational Pharmacogenomics Project sites do share and disseminate best practices in design and knowledge management, giving other sites or vendors a path to follow. In the remainder of the article, we describe the EHR and related functions designed and implemented to support pharmacogenomics from ordering to clinical use ( Figure 2 ).

Figure 2
figure 2

EHR development and operational processes. Pharmacogenomics implementation requires preimplementation research and assessment, technical development of informatics infrastructure, and integration with laboratory and clinical operations. Accessibility to users, both patients and providers, is integral. CDS, clinical decision support; EHR, electronic health record; P&T, pharmacy and therapeutics; PGx, pharmacogenomics; PHR, personal health record; PREDICT, Pharmacogenomic Resource for Enhanced Decisions in Care and Treatment; Rx, prescription.

Full size image

Implementation of Preemptive and Indication-Based Pharmacogenomic Test Ordering

Pharmacogenomic variant data are ideally incorporated into the initial drug selection and dosing; after the patient has achieved a stable dose or drug selection through experience or sequential drug trials, the genomic information contributes diminishing returns to clinical outcomes for the majority of prescribing scenarios currently covered by PREDICT, including the drugs warfarin, simvastatin, clopidogrel, tacrolimus, and thiopurines.3,4,5,22,23,24,25 As an example, warfarin dosing is stochastically adjusted in response to serial international normalized ratio measurements, and thus the clinical impact of genetic data, such as VKORC1 and CYP2C9 variant status, is thought to wane considerably after a stable international normalized ratio is achieved, an event which generally occurs within the first 2 weeks of therapy. Similarly, the risk of in-stent thrombosis in CYP2C19 variant patients who are prescribed clopidogrel is highest in the first 30 days after the placement of a stent. Therefore, the program has prioritized testing in advance of or concurrent with drug initiation to maximize the impact of the genotype data on clinical care.

Consequently, two pharmacogenomics ordering strategies were created—preemptive and indication-triggered testing. For preemptive genotyping, the EHR was modified to display an alert when a statistical risk score meets or exceeds the threshold for all patients scheduled for clinical encounters in primary care or cardiology. The risk score predicts the probability of receiving simvastatin, warfarin, or clopidogrel over a 3-year time horizon, and the trigger score was set to 40%—a threshold that saturated the capacity of the molecular diagnostics laboratory. When a patient’s chart is flagged, the system creates a draft order for the PREDICT test within the outpatient order-entry system, which requires confirmation by the treating clinician. For indication-based testing, the PREDICT panel test was incorporated into order sets or preprocedure planning before cardiac catheterization (to capture catheterization patients who receive intracoronary stents and antiplatelet therapy such as clopidogrel) and certain orthopedic procedures (e.g., joint replacements) for which warfarin-based anticoagulation is standard. Notably, preemptive genotyping eliminates delays in obtaining the genotype, which has a minimum of 2 days and a median of 5 days of turnaround time.

We suggest significant cost savings using a preemptive panel-based genotyping strategy as compared with serial single-gene tests, given the decreasing cost of genotyping, possible exposure to multiple different medications with pharmacogenomic indications, and very high cost of severe adverse events.6,26,27 Use of multiplexed gene tests over a patient’s lifetime is likely to be less expensive relative to the potential benefit, particularly in patients with a common set of cardiovascular risk factors likely to need associated therapies. However, no health economic studies have determined the value of panel-based genetic tests outside of oncology, and there is a paucity of evidence relating panel-based genetic tests to health-care spending. VUMC has supported the PREDICT program costs with institutional funds, including assay costs, reagents, labor, instrumentation for processing, empiric research among patients and providers, development of patient informational materials, decision-support tools that provide point-of-care interventions and drug/dosing guidance based on test results, and education and training given the associated dearth of knowledge and familiarity among prescribers.15 A key goal of this investment is to catalyze further pharmacoeconomic analyses of this approach.

EHR Storage Model for Sequestration and Repository

National data standards for genetics are in early stages; a model to exchange genetic testing results is proposed by Health Level 7,28 with contributions by Pharmacogenomics Research Network–affiliated academic groups29,30 and EHR vendors.31,32 In the absence of established standards in 2010, and to meet the immediate needs of the program, PREDICT developers created a coded storage model to meet local requirements for CDS and distribution to multiple clinical information systems. Future adaptation to emerging standards such as Health Level 7 is planned to support communication with external systems. Genetic variant data produced by the Illumina VeraCode Absorption, Distribution, Metabolism, and Excretion Core Panel for PREDICT are provided either as a Portable Document Format or as plain text. As the former does not provide computable results, automatic parsing of the text format is required to extract the gene name, variant result in star nomenclature, and a call rate, which indicates the ability of the panel to yield a result at a specific variant. In the event of a call rate <98.7%, the test result is manually reviewed and generally retested by Molecular Diagnostics Laboratory staff; otherwise, it is released to an Oracle database, which initially sequesters all results from the main EHR storage.

The Oracle database is exposed to downstream systems through a filtered view limited to actionable approved variants. An automated script queries the filtered database view hourly to extract new or updated entries and, if discovered, creates a new or updated entry in the genotype section of the Patient Summary Service, a central Web service that is available to all components of the EHR and CDS (see Supplementary Figure S1 online). Examples of four components of the EHR that use Patient Summary Service are shown in Figure 3 . Patient Summary Service serves as a single source of patient-specific knowledge for medications, diagnoses, allergies, and other significant family and social history, and this infrastructure was expanded to manage genomic variants and their interpretations.

Figure 3
figure 3

Task-specific views of genomic results present in the EHR. (a) The patient summary, which serves as the front page of each patient’s record, includes a drug–genome interaction section detailing the patient’s genotype in star allele nomenclature as well as phenotype and implications for prescribing. (b) Genomic results and phenotypes are also available in the lab results section of the EHR. When a drug is ordered for a patient with an actionable genotype, clinical decision support (CDS), such as the representative (c) outpatient substitution adviser, is presented to the ordering clinician. Similarly, parallel mechanisms offer CDS in the (d) inpatient setting.

Full size image

Genotype-To-Phenotype Translation

Although the advantages of multiplexed genetic testing are becoming increasingly apparent, there are clear challenges associated with managing panel-based genetic data. Raw genotype output is not typically delivered in a standardized format and does not include phenotypic interpretations, which may be drug and patient specific. In order for the genetic results to be useful for clinical implementation through PREDICT, results were individually categorized to create a translation layer, which assigns a coded phenotype category and generates the DGI text string used for display in the EHR and CDS, when triggered (see Supplementary Figure S1 online). The assigned phenotypes are drawn from a translation table, which relates the raw genotype text string to drug and metabolism effect categories ( Table 2 ). Translations are made based on actionable variants, defined as variants that have been reviewed and approved for clinical implementation by the VUMC Pharmacy and Therapeutics Committee; however, a large proportion of variants on the PREDICT platform are not actionable due to insufficient evidence. For CYP2C9, for example, only 2 of the 13 variants tested on the platform have been approved for implementation. Genetic variants that are not deemed actionable are sequestered within a separate database, outside of the EHR, and are not accessible to patients or providers. The genotype data will only be released into the EHR as new genotypes are deemed actionable and new DGIs are incorporated into clinical care.15

Table 2 Example genotype-to-phenotype translations
Full size table

The model for the current genotype-to-phenotype translation table is to assign a value to every result produced by the Absorption, Distribution, Metabolism, and Excretion platform, even if rare. For variants without sufficient evidence to be deemed actionable, a category labeled “indeterminate” was created ( Table 2 ). For purposes of CDS implementation, no change to usual care is recommended for indeterminate genotypes. Other pharmacogenomics implementation sites have used similar approaches,33 and several consortia have been established to develop and maintain consistent guidelines for translation of genotype test results, including the Clinical Pharmacogenetics Implementation Consortium and the Translational Pharmacogenomics Project.34 The translated interpretations are viewable by providers via the EHR and incorporated into the EHR advisers; however, they are not tailored to background level of provider pharmacogenomic knowledge. Therefore, developing phenotype interpretations that are meaningful and clinically useful for providers presents its own set of challenges.

EHR Representations of Genotype and Phenotype

The centralized service architecture of the genotype-to-phenotype translation layer allows simultaneous population of multiple clinical information systems, supporting the clinician through EHR views and patients through their access to a PHR hosted on a patient portal (see Supplementary Figure S2 online). For each client system, the service responds to requests for new or updated genomic results. Whenever a phenotype assignment is changed (such as when CYP2C19*3 heterozygotes are added to an actionable “poor metabolizer” status for clopidogrel), the translation table within the service is updated manually, which triggers automatic revision of the results displayed in the EHR and PHR. Following the principle of high visibility and universal access, four task-specific views of genomic results are supported in the EHR ( Figure 3a – d ). First, the program team created a space for genomic variants to be visible within the patient summary that serves as the “front page” of the electronic chart and adjacent to the medication list. Much like an “allergy” section, this space is intended to communicate significant genomic variant information when a target medication is contemplated and before initiating a prescription. During review of the design, clinicians and the Pharmacy and Therapeutics Committee required the display of any pharmacogenomic result whether indicating a variant or not, such that there was a quick method of determining if a patient had already been tested. This current presentation format does not scale to many implemented DGIs; therefore, a redesign is in progress.

Second, the phenotype delivered by Patient Summary Service triggers CDS within the outpatient e-prescribing environment as well as the inpatient computerized physician order entry environment when a prescription or medication order conflicts with the phenotype status (see Supplementary Figure S1 online). For example, providers prescribing clopidogrel for a patient with an intermediate-metabolizer or poor-metabolizer phenotype will receive therapeutic guidance to switch to an alternative antiplatelet therapy (see Supplementary Table S1 online). Finally, new pharmacogenomic information is released from the laboratory. This mechanism (along with the patient summary) supports reconsideration of patient therapy whenever new DGIs are released. Among the challenges encountered, EHR designers must decide how to represent risk; the potential impact of phenotype labeling and the utility of adding quantitative-risk measures to these brief interpretations are currently unknown.

Display of Genomic Results in PHRs

PREDICT genetic results are released into the patient’s EHR to guide therapy and clinical decision making. In addition, given the burgeoning body of literature suggesting the importance of empowering patients with health information and increased efforts surrounding the Health Information Technology for Economic and Clinical Health Act,35 PREDICT genetic results have also been made available to patients through VUMC’s patient portal, My Health at Vanderbilt, a resource that allows patients to view EHR data, message their health-care providers, and read general health information tailored to their medical history. Through PREDICT, we have added content in My Health at Vanderbilt related to a patient’s genetic test results (see Supplementary Figure S2 online). The first release of genomic results contained a simplified copy of what was displayed to providers in the EHR: the genetic test result with a brief interpretation, e.g., “CYP2C19, one copy of the variant, poor metabolizer of clopidogrel.” Feedback from focus groups overwhelmingly indicated that patients preferred detailed, descriptive background information related to drug side effects and how genetics may affect a patient’s risk for adverse events. On the basis of this feedback, more comprehensive narratives with graphics are being developed and provided at a seventh-grade reading level.

Evolution of PREDICT Since Launch

PREDICT was launched in September 2010 with genotype-tailored dosing guidance for clopidogrel.15 The decision to focus on clopidogrel was made following the Food and Drug Administration black box warning alerting physicians and patients to the role of CYP2C19 variants in medication response.3 The Food and Drug Administration did not indicate how to incorporate CYP2C19 variants into clinical decision making; however, an efficacious alternative, the antiplatelet drug prasugrel, was not affected by CYP2C19 genotype.36,37,38,39 Therefore, the initial clopidogrel adviser was designed to activate when patients were homozygous for CYP2C19*2 or *3 allele and displayed recommendations to increase clopidogrel maintenance dose to 150 mg daily or switch to prasugrel barring any contraindications.

Since launching the program, over 75 articles have been published with the potential to influence genotype-to-phenotype mappings or the content of the clopidogrel CDS. Following publication of a large meta-analysis3 and our internal analysis,40 which both showed significant reduction in clopidogrel efficacy in individuals heterozygous for CYP2C19 variants, we added such individuals to the program. Moreover, new, rare CYP2C19 variants were determined to impair clopidogrel metabolism,14 and new, effective alternatives to clopidogrel were released on the market. These advances warranted modifications to both the genotype–phenotype translations and the clopidogrel CDS recommendations. Updating the knowledge base and changing the user interface for the CDS to add additional choices required comparatively less effort than the initial development, partially because of the separation of these components into Enterprise Services (see Supplementary Figure S1 online). However, modifications to the phenotype map often changed the risk status of patients who were already genotyped, requiring providers to reconsider the initial drug selection or dosing. For each of these scenarios, we organized a communication plan, identifying affected patients and manually notifying providers using secure electronic messaging within the EHR.41

The program continues to expand and incorporate CDS for additional DGIs into the EHR, including recommendations for warfarin, simvastatin, thiopurines, and tacrolimus. Two of the released DGIs are relevant to pediatric populations and required the development of guidelines applicable to both adult and pediatric populations, as well as DGI-specific suppression of genetic results and EHR advisers for those DGIs that were not applicable to a pediatric population (e.g., warfarin advisers). Infrastructure available at the time of these deployments allowed for a simple, alternative set of text for adult and pediatric patients. This required changes in both the database model and the presentation layer to determine, on the basis of the age of the patient, which text was appropriate for display.

Discussion

The design and implementation of EHR features to support a large multi-DGI pharmacogenomics program required iterative refinements, in part because there is little published guidance on how to leverage HIT to translate genomic medicine to clinical practice. We described our initial design choices and subsequent changes in an effort to inform other institutions that are contemplating or have initiated a similar effort. One of the major successes in the past 5 years is the formation of cooperative efforts from pioneering institutions associated with the Pharmacogenomics Research Network to organize and curate the pharmacogenomics knowledge base relating genomic variation to therapeutic decision making in the form of clear, accessible guidelines.10,11,12,13,14 Similar efforts to share implementation practices among members of the Translational Pharmacogenomics Project and the eMERGE Network have made substantial progress.34 Overall, the gap between the conceptual model of personalized medicine and actual clinical implementation is closing but remains wide for most health systems.42 The PREDICT implementation approach is distinct because of the scope of drug–genome interactions that are targeted for adult and pediatric populations, the duration of the program, and the emphasis on preemptive testing. In addition, the ability to leverage on-site developers familiar with the locally developed EHR allowed efficient implementation. Although the specific form of this implementation is institution specific, the abstracted challenges described in this article are generalizable.33

We found the major challenges for incorporating PREDICT relate to the complexity of raw genotype data and the lack of existing standards to store and transmit genomic data. Genotyping platforms do not output results in a coded reference standard and are not accompanied by interpretations. Integrating with downstream EHR tasks required parsing of the gene result report and a translation layer able to contend with undefined variants. Manufacturers of genotyping instruments can improve the ease of implementation by adhering to coded standards (as they are developed) and providing more detailed documentation of potential genomic output. Second, we sought to preemptively map all variants but discovered rare variants that were undefined; an automated process within the EHR infrastructure to track and examine new, undefined variants would be valuable to ensure the timely updating of a translation table and could eventually serve as a tool for discovery of potential variant function. Third, EHR integration of genomic data requires a process to manage the release of new or materially updated drug–genome data as thousands of patient records are affected. Such releases also require significant communication and education efforts to inform providers of emerging or changing evidence. Finally, the scalability of EHR integration is challenged by several technical factors, including limited screen “real estate” to display significant variants and inflexible models of displaying results that may not yet be pertinent to patient care.

Limitations of the PREDICT EHR Model

The application of pharmacogenomic testing to clinical care is complex and requires established and comprehensive infrastructures to support implementation. With quickly advancing genotyping (and genome sequencing) technologies, emerging evidence, and changes in therapies, these infrastructures must be prepared to accommodate rapid modifications and an explosion in genetic variants. Although PREDICT represents one viable model for implementation of pharmacogenomic information into the EHR, there are limitations and challenges that offer opportunity for improvement and fine-tuning of the program. Despite attention to the succinct and understandable interpretation of genomic results, the EHR displays may not be sufficient for providers without specific pharmacogenomics training. The brief interpretations provided presume a baseline knowledge of pharmacogenomics and are not intended to be educational. Furthermore, PREDICT affects providers in multiple specialties, creating even greater provider education challenges. The provider EHR displays are not currently customizable by specialty, health-care role, or baseline knowledge, but such flexibility may be needed as the number of implemented DGIs increase. Moreover, results may be returned outside of the context of a clinical encounter, for example, when a DGI is released into the EHR many years after the patient’s initial genetic testing. Similarly, although significant effort has been made to develop understandable and meaningful PHR displays, further research is warranted to elucidate more effective methods of communicating complex genomic information to patients. In addition, there is currently no infrastructure in place to automatically and reliably deliver genetic results to providers outside of Vanderbilt’s EHR system; thus, some patients may be tested through PREDICT but not benefit from future decision support after they return to their primary providers outside of the Vanderbilt network. Although PREDICT recommendations are based on the most up-to-date evidence and expert opinions, incorporating genomic information with clinically relevant nongenomic factors in CDS recommendations is currently outside of the scope of the program.

Pharmacogenomic Adoption: The Way Forward

The challenges and lessons learned from PREDICT implementation highlight the need for improved EHR integration and interoperability. For patients not receiving care exclusively at VUMC, improved communication and transfer of genetic results to external providers is the first step toward this integration and is necessary to advance genotype-tailored decision making. Clinical notification of high-priority genetic results (e.g., those associated with life-threatening adverse events or with prolonged clinical utility) could be achieved by leveraging national electronic messaging infrastructures and will pave the way for full EHR integration. Pharmacogenomic adoption is limited by provider knowledge and usability of EHR-displayed genomic information. Maintaining awareness of evolving pharmacogenomic evidence and emerging therapies and incorporating this information into clinical practice require procedures for systematic evidence review and an informatics infrastructure that enables prompt modifications of genomic advisers within the EHR system.15 Improved advisers and information displays that can be modified easily and incorporated within the EHR with very little informatics support will be vital as existing DGIs are updated and additional DGIs continue to be implemented. Moreover, portability of internally developed CDS across EHR systems will be critical for dissemination of clinical pharmacogenomics. We believe that use of Internet-based Web services to encapsulate genetic results and securely communicate relevant guideline-based recommendations and knowledge across institutional boundaries will compel efficient and widespread clinical adoption of pharmacogenomic evidence in real-world medical practice.

Disclosure

The authors declare no conflict of interest.