Background
Genetic variants can influence drug metabolism, transport and receptor response and thereby lead to reduced drug activity or increased toxicity [
1‐
3]. Prominent examples are the anticoagulants clopidogrel and warfarin that are metabolized by CYP2C19 and CYP2C9, respectively. Variants in these enzymes can alter the plasma levels of the anticoagulants and thereby lead to insufficient anticoagulation or increased risk of bleeding. The influence of genetic variants on drug activity led to the development of pharmacogenomic tests and drug dosing guidelines which incorporate pharmacogenomic data into the drug prescription process [
4,
5]. An example for the development of pharmacogenomic recommendations and best practices guidelines is the publicly available web-based knowledge base PharmGKB (
https://www.pharmgkb.org/). It includes dosing guidelines by the Clinical Pharmacogenetics Implementation Consortium (CPIC), the Royal Dutch Association for the Advancement of Pharmacy - Pharmacogenetics Working Group (DPWG), the Canadian Pharmacogenomics Network for Drug Safety (CPNDS) and other professional society. Other examples of pharmacogenomic knowledge bases are the OncoKB (
oncokb.org/#/) by the Memorial Sloan Kettering Cancer Center and the PMKB (
https://pmkb.weill.cornell.edu/) by the Weil Cornell Medical College.
In prospect of whole genome sequencing, the discovery of new gene-drug interaction pairs is very likely and will further increase the pharmacogenomic knowledge base. However, translating this pharmacogenomic knowledge into clinical routine has been slow and is hindered by the lack of the physicians’ knowledge and experience in pharmacogenomic testing [
1,
6‐
8].
In recent years, informatics has gained crucial relevance for improving patient care. This includes a considerable amount of published literature which describes the current efforts on developing and implementing pharmacogenomic clinical decision support systems (CDSS) [
9‐
11]. Pharmacogenomic CDSS might help overcome some of the barriers of implementing pharmacogenomic knowledge into clinical routine [
7,
10].
Pharmacogenomic CDSS are computer-based systems which support health care providers in prescribing drugs at the point of care. These systems provide physicians and other health care providers with reasonably filtered pharmacogenomic information such as gene-drug interaction alerts or patient-specific treatment recommendations. A pharmacogenomic CDSS can either be integrated into the local hospital information system (HIS) or used as a separate program such as a web service or mobile application [
10]. Furthermore, pharmacogenomic CDSS can provide passive or active clinical decision support (CDS). Active CDS include rules and alerts. An alert, for example, might be triggered because a high-risk drug is prescribed and pharmacogenomic testing prior to the drug application would be indicated. Passive CDS require the user to actively search for the information, e.g. clicking on a button or opening a case report [
10,
12].
To develop a prototype it is crucial for developers to know which pharmacogenomic CDSS features and user-system interactions have been developed, implemented and tested in previous pharmacogenomic CDSS efforts and if they were successfully applied. Welch and Kawamoto systematically reviewed the literature on pharmacogenomic CDSS including manuscripts from 1990 to 2011 [
13]. Given the recent rise of omics technologies, the findings of their systematic review cannot include the most recent developments of pharmacogenomic CDSS. In addition to that, Welch and Kawamoto did not compare the designs of user-system interactions (e.g. passive vs. active CDS, displaying pre-testing vs. post-testing alerts, or the contents of such alerts presented to the user).
Dunnenberger et al. and Hicks et al. previously reported on recent developments of pharmacogenomic CDSS since 2012 [
14,
15]. They also mentioned some potential concepts for implementing pharmacogenomic CDSS into clinical routine. However, they did neither analyze designs of user-system interactions nor did they describe whether or not such designs have been evaluated. Furthermore, they limited their scope to concepts involving an EHR. For developers, it is crucial to also know which potential designs of user-system interactions exist without involving an EHR.
To our knowledge, there is no systematic or scoping review to date which provides an overview of the recent developments of pharmacogenomic CDSS and their designs of user-system interactions. Given the clinical importance of pharmacogenomic CDSS, we address this topic by comparing the functionalities and the designs of user-system interactions of published pharmacogenomic CDSS since 2012. The objective of this paper is to provide an overview of the recent developments of pharmacogenomic CDSS and their designs of user-system interactions.
Methods
To minimize bias in the selection of included studies, we followed the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines as far as appropriate for this scoping review [
16,
17]. We achieved a high degree of completeness by providing information on 22 out of the 27 points recommended (see Additional file
1). A study protocol was written prior to the investigation, but has not been registered.
We searched PubMed for pharmacogenomic CDSS published between January 1, 2012 and November 30, 2016. We used the keywords “pharmacogenetic*”, “pharmacogenomic*”, “decision support” and “medical decision making” for the search query as shown in Table
1. The final search was conducted on December 01, 2016. The inclusion criteria for the scoping review were as follows: English article; manuscript in peer-reviewed journal; research article; describing a clinical prototype or a fully developed pharmacogenomic CDSS in clinical routine; describing the functionalities and the design of user-system interactions of a pharmacogenomic CDSS.
Table 1
PubMed search query
1. Pharmacogenetics[MeSH Terms] 2. pharmacogenetic*[tw] 3. pharmacogenomic*[tw] 4. “decision support systems, clinical”[MeSH Terms] 5. “decision support”[tw] 6. “decision making”[tw] 7. (“2012/01/01”[PDAT] : “2016/11/30”[PDAT]) 8. #1 OR #2 OR #3 9. #4 OR #5 OR #6 10. #7 AND #8 AND #9 |
Two authors (MH and MS) screened the titles, index terms, and abstracts for all identified publications to determine, if all inclusion requirements were met. This was done independently by both authors (MH and MS) and for all identified articles. In this way, potential differences in the judgment of including or excluding certain articles could be spotted. If no clear decision could be made on the basis of this information, the article was obtained in full-text and a decision on the inclusion was based on information from the full-text. The full-texts of all included articles were obtained via institutional library access or the authors’ user profile on the ResearchGate platform.
Data abstraction
For all publications meeting the inclusion criteria listed above, the following data items were extracted by both authors (MH and MS) independently and for all included articles: system or project name; users and study location; CDSS development status; EHR integration; web-based access; active or passive CDS and the CDS features for user-system interactions (categorized as “alerts”, “reports”, “EHR data”, “inbox messages”, “search engines” and “others”). Differences were discussed among co-authors (MH, MB, MS) in order to resolve disagreements and to achieve a consensus. Following this, all abstracted data were reviewed and revised by all co-authors, which led to the final version of data abstraction of all articles.
The system or project name was either the specific pharmacogenomic CDSS name (if applicable) or (if no system name existed) the related project name. If neither a system nor a project name was available, we created a surrogate pharmacogenomic CDSS name (marked with s/n for “surrogate name”). Users were defined as physicians, pharmacists, other health care providers or patients. The study location was the main institution, where research was conducted. The CDSS development status was either prototype or fully developed. The EHR integration category distinguished between pharmacogenomic CDSS which were integrated into an electronic health record (EHR) or a computerized physician order entry system (CPOE) (“yes”) and those designed as stand-alone systems (“no”). A stand-alone system exists in parallel to the EHR or the CPOE. If the pharmacogenomic CDSS was accessible through the internet (e.g. online portals of laboratories), it was marked as “web-based access”. The alert category defined whether the pharmacogenomic CDSS provided pre-testing or post-testing alerts and which information these alerts contained. A report was defined as a summary report or patient letters including the patients’ pharmacogenomic information. Within the EHR data category every kind of pharmacogenomic information stored in the EHR for clinical decision support was included. Furthermore, inbox message designs and search engine designs used for pharmacogenomic CDSS were documented if applicable.
The risk of bias for individual studies was not systematically assessed.
Data analysis and presentation
Extracted and categorized data were used and the data items were grouped into two logical domains. The first domain included the categories “system or project name”, “users and study location” and “CDSS development status“. Whereas the second domain comprised the categories specifying the user-system interaction: “EHR integration”, “web-based access”, “active or passive CDS”, “alerts”, “reports”, “EHR data”, “inbox messages”, “search engines” and “others“. We summarized the articles in the form of tables and narrative discussion. Differences were discussed among co-authors (MH, MS) in order to resolve disagreements and to achieve a consensus. Following this, the results by MH and MS were reviewed and consented by all co-authors. This led to the final interpretation and presentation of the abstracted data of all articles.
To provide an unbiased overview of all pharmacogenomic CDSS including those mentioned in this scoping review, we grouped all articles which used a common pharmacogenomic CDSS. The articles were grouped by the abstracted system or project names and by the attributes of the first domain.
Furthermore, we identified and analyzed the designs of user-system interactions of recently published pharmacogenomic CDSS. Therefore, we used the abstracted system or project names and the results of the tabulated attributes of the second domain.
Discussion
Within the last 5 years, several pharmacogenomic CDSS have been developed which have the potential to support the incorporation of pharmacogenomic testing into clinical routine. They comprise different forms of active and passive CDS targeting both physicians and pharmacists. However, most of the included articles are solely describing prototypes of pharmacogenomic CDSS rather than evaluating them.
We performed a literature research to conduct a scoping review of the designs of user-system interactions of pharmacogenomic CDSS. We limited our PubMed search to the articles published between January 1, 2012 and November 30, 2016 to focus on pharmacogenomic CDSS which were recently developed and published. Nevertheless, we might have neglected relevant designs of user-system interactions published before 2012.
We found pre-test and post-test alerts to be amongst the most cited user-system interactions within recently published pharmacogenomic CDSS. As mentioned by Bell et al. pharmacogenomic test results remain relevant to the medical treatment of a patient over his/her lifetime, and need to be stored in a way that they will not be forgotten or lost [
12].
Recommendations, which are presented in a pharmacogenomic alert, need to be formulated carefully for several reasons. First, other variables besides genetics such as “comorbidities” or “co-medications” should be considered when prescribing drugs [
10]. Therefore Manzi et al. refrained from using words such as “should” or “must” and from dictating exact dosing adjustment recommendations [
28]. Second, presenting specific drug alternatives or dose adjustment recommendations in an alert might also raise concerns about the liability in case of an adverse drug event, especially when different guidelines are displayed for the same drug [
10,
33].
In this context, alert fatigue has been mentioned as another main challenge for using alerts within pharmacogenomic CDSS [
25,
28,
38]. To avoid over-alerting caused by repetitive alerts Ji et al. [
25] included exclusion criteria in the rules and Manzi et al. [
28] designed their alerts to only notify if an action is recommended by the physician. Both strategies should be considered and combined when implementing alerts into pharmacogenomic CDSS.
Overcoming technical barriers seems to be another main challenge in designing alerts for a pharmacogenomic CDSS. For instance, Nishimura et al. intended to link physicians directly from the alert to the patient’s pharmacogenomic lab results in the local EHR. In their study, such a link was considered to be useful for user-system interactions. However, due to restrictions of the vendor-based CDSS it was impossible for Nishimura et al. to integrate such a link into the post-test alerts [
33].
Only four alert designs of pharmacogenomic CDSS have been evaluated since 2012 and were described in seven articles. Three studies relating to two pharmacogenomic CDSS demonstrated that alerts can lead to a change of the initial prescription and therefore prevent severe adverse drug events effectively [
12,
22,
31]. These order changes can be seen as the physician’s acceptance of the pharmacogenomic CDSS alert. Furthermore, the physicians’ acceptance of the UW-PowerChart prototype alerts has been high [
11,
33,
38]. In contrast, using alerts to drive providers towards online pharmacogenomics education might be ineffective [
35].
The acceptance of pharmacogenomic CDSS reports was high amongst those physicians and pharmacists who participated in the four studies that evaluated four different pharmacogenomic CDSS [
10,
22,
40,
41].
Delivering pharmacogenomic information not only to physicians but also to patients might be a crucial feature of pharmacogenomic CDSS. Pharmacogenomic information is usually gathered within the environment of a particular clinic or health care provider. Via patient letters [
18,
31,
39,
42] and online portals [
25], which contain the pharmacogenomic test results, this information can be transferred to other health care providers. For instance, patients can be advised to discuss their pharmacogenomic test results with physicians whenever a medication is prescribed which is affected by these pharmacogenomic results [
42]. In the evaluation study by Peterson et al. the online patient portal of the PREDICT project was deemed to be useful and was accepted by participating patients [
20].
Another essential function of many pharmacogenomic CDSS was storing pharmacogenomic information in the EHR. For instance, if override reasons for alerts are passed on to the order tracking field of a medication, they can be communicated to other care providers when they reorder, verify, dispense or administer this medication [
9]. Furthermore, the storage of all pharmacogenomic test results (not only of the relevant ones) in a patient’s medical record offers physicians an overview of all pharmacogenomic tests which have already been carried out. As a result, reordering the same pharmacogenomic test might be avoided [
20]. Collecting all pharmacogenomic information in a separate problem list within the laboratory section might also serve as an option for a quick overview of all potential gene-drug interactions known for a particular patient [
28‐
30,
34]. 11 out of 22 physicians who participated in the evaluation study by Overby et al. believed that the genetic test results within the EHR laboratory section were useful [
44].
Via inbox messages physicians can be informed about new pharmacogenomic test results, which are available within the laboratory section of a patients’ medical record [
9,
29,
31,
35,
36,
39,
42,
43]. If the physician needs any kind of advice regarding the pharmacogenomic test results, he/she might request a consultation by sending a message to a clinical pharmacist [
22,
30]. This provides the physician with an active decision support before he/she orders a particular medication at the point-of-care.
With a search engine physicians are enabled to search for a disease indication and compare various medical treatment options based on a patient’s pharmacogenomic information. The benefit of such a user-system interaction is that physicians can get the decision support before they decide which medication they want to prescribe. In contrast, the use of alerts requires the physician to first select a medication before getting the pharmacogenomic information in response [
18,
22,
23].
Recommendations for the implementation of a particular design of user-system interaction can only be made on a very high level since the recently developed pharmacogenomic CDSS have not been sufficiently evaluated yet within a clinical setting. Implementing pre-test or post-test alerts seemed to be the most popular approaches. However, such active CDS tools require a comprehensive and well-curated pharmacogenomic knowledge base. Such a knowledge base has to be both developed and maintained by physicians with sufficient knowledge of pharmacogenomic CDSS prior to the implementation of pharmacogenomic alerts. We recommend the establishment of both the necessary group of medical experts and the corresponding knowledge base in order to further evaluate pharmacogenomic alerts.
However, we recommend the implementation of a passive CDS in the form of structured pharmacogenomic reports in the first instance. Clinical reports for clinicians have been well-established in clinical environments over many years [
45‐
47] even though such reports might be unstructured or incomplete in some cases [
48,
49]. Therefore, we believe that the implementation of a structured pharmacogenomic report would most likely fit into the working habit of a clinician.
Future pharmacogenomic CDSS will likely include prediction models to recommend pre-emptive genotyping for patients exceeding particular risk thresholds. A patient’s diagnosis might be, contain or induce a risk factor, which will likely require a medication with known gene-drug interactions within the next few years. Whenever such a diagnosis is entered into a patient’s medical record, an alert might be set off, which recommends pre-emptive genotyping [
50,
51]. This will further enhance the usefulness of integrating alert and inbox message options into a pharmacogenomic CDSS. Prediction models have previously been used, but only to identify potential patients for a pharmacogenomic CDSS study [
20,
25]. Nevertheless, risk thresholds or risk scores should be defined carefully and in consensus with the medical staff in charge. Otherwise, inadequate risk thresholds might lead to over-alerting and alert fatigue.
A limitation of this study is its methodological rigor as compared to a full systematic review. This is especially relevant to the selection of the source databases. We only used the MEDLINE Database, which might limit the completeness of our search. However, we believe that our adherence to high methodological standards throughout this review as defined in the PRISMA statement helps to minimize bias on study selection and reporting of evidence.
We used the keywords “pharmacogenetic*”, “pharmacogenomic*”, “decision support” and “medical decision making” for the search query. These are the most common terms in literature to describe a pharmacogenomic CDSS. It is possible that these keywords did not cover all pharmacogenomic CDSS published since 2012. In our opinion, using the most common terms was sufficient to provide a broad range of the designs of user-system interactions which were used in recently published pharmacogenomic CDSS. To conduct a systematic review of this topic, an advanced search query with further relevant terms might be preferable
We did not assess the risk of bias. The reader therefore cannot assess the validity of the individual studies included in this scoping review.
As this review is intended to describe the functionality and designs of user-system interactions of pharmacogenomic CDSS, it is also limited in collecting evidence for the effectivity and the overall medical usability of CDSS for pharmacogenomics in the clinical setting.
Evaluating the users’ acceptance of designs of user-system interactions was part of only a few articles in our scoping review. Further evaluation efforts addressing this topic will be necessary will be necessary to support the development of prototypes.
Acknowledgements
The present work was performed in fulfillment of the requirements for obtaining the degree “Dr. rer. biol. hum.” from the Friedrich-Alexander-Universität Erlangen-Nürnberg (M Hinderer). This study was conducted within the MIRACUM consortium.