We identified a minimum starter set of DDIs which should be classified in all medication KBs as high severity and implemented for decision support in all EHRs. The DDIs identified by the panel represent a clinically important group because they have a high potential for patient harm, and are agents that are contraindicated for co-administration. The list suggested here may not be complete but represents a high proportion of DDIs that fall into this category. Further, some interactions that are clinically important and deserve a high severity rating may have been ruled out by the panel because they did not meet the strict criteria of drug pairs that are contraindicated to be prescribed together and where the risk definitely outweighs the benefit offered by co-prescription. The DDIs that are in the final list identified by the panel represent a very small proportion, probably <0.2%, of DDI alerts that are generated in clinical practice.7
There would be hardly any increase in workload since these alerts would come up rarely. From an implementation perspective, in a scenario where alert fatigue is at a maximum and a clinician chooses to ignore all alerts, this list should probably constitute an additional layer of response from the clinician before being overridden. For example, if clinicians are allowed to override all alerts irrespective of severity levels, then one way of making sure they have not ignored these interactions is to require them to provide a reason for overriding an alert related to co-prescription of the drugs contained in this list. This list identifies interactions that meet the stringent criteria of being both clinically severe and drugs that should not be concurrently prescribed. It is not a comprehensive list but in turn an attempt to describe what could be included in such a starter set and how it can be developed. Since the vast majority of DDI alerts that are generated in most EHRs may not be on this list, alert fatigue will still occur if the remainder of DDI alerts are presented to prescribers. To manage this, there is a need to determine which of the remaining DDI alerts should be presented in a non-interruptive manner and which can be removed from the KB.
Previous studies have identified lists of critical DDIs for specific classes of agents or drugs exhibiting specific mechanisms of interaction.18–21
More broad based efforts, by Malone et al
, have previously resulted in a list of high severity interactions intended for use in the outpatient setting.13
The effort described here differs from these efforts in scope and the intended use of the proposed DDI list. Malone et al
derived their initial list of DDIs from four drug compendia and not from medication KBs currently employed for decision-support by EHRs. This may be because the focus of their study was limited to community and ambulatory pharmacy settings rather than all EHRs, which is the focus of the effort described here. This limitation resulted in the exclusion of interactions related to drugs, such as halothane or dopamine, which are not routinely dispensed in the ambulatory pharmacy setting, but are commonly administered to inpatients. We elected to consider all types of drugs because we wanted to develop a set of interactions that could be generalized across EHRs used in both the inpatient and ambulatory settings. Malone et al
utilized a small panel of experts, including two physicians, two clinical pharmacists, and an expert on drug interactions, to vet these interactions. Twenty-one panelists participated in this study; they represented diverse perspectives on the use, development, and implementation of medication-related decision support. In addition, this effort benefited from the long experience and commitment at our own institution to the development and maintenance of the Partners MKB and its use in driving CDS in diverse clinical settings—inpatient and outpatient, community and academic medical centers, and Computerized Provider Order Entry (CPOE) and other clinical information systems. A previous study employed a similar panel-based approach to identify critical interactions for DDI checking and duplicate therapy checking within CDS systems. This study was limited in assessing only those interactions that are available in the Partners MKB, without consideration of the additional knowledge sources described here.22
A recent evaluation compares leading approaches to critical DDI lists and suggests that the generation of the list described here has the potential to represent an important step forward in standardizing critical DDIs across EHRs and in ensuring that the most clinically important interactions are being seen by all providers.23
Utilization of a large review panel with diverse expertise in medication-related decision support provided credibility to this list which could serve as a minimum set of critical DDIs. This list and the methods employed here could also serve as a starting point for additional DDI work that would have more impact on both alert fatigue and consequent patient safety outcomes.
Such a list, if implemented in EHRs, can standardize the process of identification of critical DDIs and save others the effort of doing this work. Further, by mandating alerting against the most critical DDIs in addition to tailoring the list for only the most significant ones that warrant interruption, we can prospectively change practice and prevent unintended consequences at the bedside.
In addition to vetting the list of contraindicated drug pairs for implementation in EHRs, the expert panel also provided insights into the pragmatic challenges that might encompass the development and implementation of such a list. These are detailed below.
Gaps in assigning drug class membership for pharmacokinetic interactions
The panel suggested an objective assessment of drug membership for a DDI pair, especially those interactions that are pharmacokinetic in nature, based on their effects on the human cytochrome (CY) P-450 system, for example 3A4 inhibitors or 1A2 substrates. Specific modifications have been outlined in . The incorporation of drug classes based on the CYP-450 system represents a more pragmatic approach to representing and maintaining drug interaction knowledge in providing medication-related CDS. For pharmacodynamic interactions however, the existence of pharmacologic variability within a class can cause a DDI to affect only some but not all the drugs in a class.
A key issue from the informatics perspective is that an important gap exists in representing these classes using Federal Medication Terminologies, such as RxNorm and the National Drug File–Reference Terminology. Neither of these terminologies is currently capable of representing concepts such as “CYP-450 3A4 inhibitors” or “CYP-450 1A2 substrates”. This corroborates the findings of Bodenreider et al
, who evaluated the mapping of drug classes used in representing DDIs to the Unified Medical Language System (UMLS) concepts and found that 17% of the names could not be mapped to any particular class. This implies that these classes are not represented in any of the UMLS source vocabularies, including SNOMED and National Drug File–Reference Terminology. A large majority of the unmapped concepts referred to the metabolism of the drugs, related to a particular enzyme of the CYP-450 family. In addition, there was lack of representation of UMLS concepts that could be used for drug classes, such as “drugs known to prolong the QTc interval”, etc.24
This is another concept that is employed in the high-priority DDI list which cannot be represented using current medication terminologies. Appropriate implementation would require bridging these gaps to allow adequate representation of drug classes.
Defining accurate and complete membership of drugs encompassed within pharmacokinetic drug classes is another critical gap. We examined a number of references seeking a complete list for determining the membership under these drug classes. None of the widely used sources, even some of the most authoritative such as the FDA,25
table on CYP-450 drug interactions, or Hansten and Horn,15
provided a complete list of drug members for these classes. Future work should focus on deriving consensus on membership underlying these drug classes so as to promote adequate representation of the drug class concepts.
Need for specific criteria to assess critical interactions
Previous studies have identified the lack of standardized criteria for evaluating the severity and clinical significance of DDIs. Two recent studies, one by Wang et al17
and the other by Olvey et al
have identified the low rate of overlap (5% and 13%, respectively) among drug compendia for even the most clinically significant interactions. The authors concluded that the lack of overlap existed due to differences in criteria for assessing the severity and level of documentation, among medication KBs, for even the highest severity interactions.
The lack of uniformity across KBs makes it difficult to identify a single list of DDIs with high clinical significance. Panelists cited the MKB developed in the Netherlands by the Royal Dutch Association for the Advancement of Pharmacy as an example of a nationally implemented DDI database. The panel recommended the development and maintenance of explicit editorial guidelines to facilitate a standardized severity assessment process based on criteria of evidence supporting the DDI, clinical relevance of potential adverse reaction resulting from the DDI, assessment of risk factors, and probability of the interaction. The methodology of evaluation described here paves the way for such a consensus process.
Lack of primary literature supporting the evidence for DDIs
Gathering empirical evidence from the literature was a barrier in being able to assess the likelihood of an interaction, a major criterion for assessing the significance of a DDI. We observed that most of the available literature was in the form of documented case reports or clinical studies with drugs belonging to other closely related drug families, or package inserts. The information contained in these package inserts is often incorrect or too conservative for use to assign DDIs in a KB. Moreover, these interactions are not updated regularly to reflect current knowledge. The reason for lack of proper clinical literature contradicting the manufacturer's information is that if the label identifies a certain agent as contraindicated with another agent or in a particular disease state, it is hard for researchers to justify doing clinical trials to empirically test these interactions. Another problem was that while theoretically an interaction may be significant, empirical evaluation may suggest otherwise. Thus, more often than necessary, KB providers have to rely on package inserts to make determinations on DDIs, which may result in overestimating the severity of a large majority of drug interactions.27
Poor use of predisposing patient risk factors
Consideration of patient characteristics and co-morbidities is needed to improve the specificity of DDI alert logic. KB vendors pointed out that utilization of these characteristics in conjunction with the drug interactions logic would have a large effect on improving the specificity of alerts. However, despite the knowledge on risk factors that predispose a patient to particular interaction, this information is seldom employed. This is because in the current state, EHR implementations typically do not employ logic that uses both patient characteristics in conjunction with the medication profile of the patient to generate alerts. Information that is routinely present in EHRs, such as patient age, gender, and co-morbidities that mitigate or increase the risk of an interaction, should be used to contextualize the alert for a specific patient to reduce the generation of clinically insignificant alerts. Greater collaboration is needed between EHR vendors and KB vendors to improve alert logic based on patient context information and provider behavior in response to alerts. This information sharing can be beneficial to both parties. The EHR vendors will see greater user satisfaction due to a possible decrease in alert fatigue, and the KB providers will be able to improve their products and prune the alerts based on how they are actually used in clinical practice.
Resource intensive process
Developing and maintaining a list of high priority DDIs is a resource intensive process. Panelists expressed concern regarding maintenance of two separate KBs, should such a list be implemented in addition to what is employed by EHRs. Panelists suggested that the process for evaluating DDIs should be centralized so that the burden of resources needed could be shared and such a list could serve a larger public good. Names of organizations that were suggested as neutral entities for maintaining the list of DDIs, were the American Society of Health-System Pharmacists, the College of American Pathologists, the National Library of Medicine, or the US Pharmacopoeia. An alternate suggestion was to authorize the organization that undertook its development and have periodic discussions with KB vendors and feedback from clinicians to modify the list based on new evidence that becomes available and on physician acceptance of alerts in actual clinical settings. Either approach would involve creating a clearing house of information where KB vendors and the institution responsible for maintaining the standard list could make decisions on its content, such as suggesting severity levels based on previously agreed upon criteria and being able to nominate, elevate, or demote interactions from the list. Further, the critical set of alerts should contain information on further stratification based on patient context variables, which implementers could tune based on the level of specificity of alerts relevant to the clinical setting. Such assessments could be further guided by real-world clinical input through implementation in EHRs, to assess outcomes associated with the implementation of this set of DDIs. Further, the heuristics, editorial guidelines, severity levels, and other definitions adopted by the group, in developing the DDI list, should be transparent and should take into consideration the current systems followed by KB vendors to facilitate easy integration with their own databases without the need to for re-programming.
This study had a number of limitations. We had limited resources and a strict schedule so that we could not perform a comprehensive literature review for each DDI evaluated. Despite this limitation, we identified a small set of interactions with strong consensus, and described a process that could be utilized to gain consensus on assessing additional DDIs. The expert group was limited in size, but included many of the leading experts in this domain and it represented a diverse set of perspectives. Future work should include analyses of critical DDIs from multiple sources to improve representation of interactions in this list, and a determination of impact on alert override rate.