Keywords
clinical decision support - alerting - alert fatigue - safety - governance
Background and Significance
Background and Significance
Electronic health record (EHR) alert fatigue, while widely recognized as a concern nationally, lacks a corresponding action plan for management.[1]
[2]
[3]
[4] This manuscript synthesizes several approaches and recommendations for better alert management, derived from four U.S. health care institutions that presented their experiences and recommendations at the American Medical Informatics Association (AMIA) 2019 Clinical Informatics Conference (CIC) in Atlanta, Georgia, United States. The assembled health care institution leaders represent academic, pediatric, community, and specialized care domains.
There is increasing national attention on the impact of EHRs on clinician wellness, with potential threats including usability limitations, imposition of functions and processes that do not correspond to actual clinical work, onerous documentation and data entry requirements, and a perception of excessive, burdensome alerting experienced by frontline clinicians, among others.[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12] Nearly one in four medication orders generate an alert.[13] While alerts can change clinician behavior and improve care processes, clinicians largely dismiss them, an action known as overriding.[14] Overrides may be clinically appropriate, such as when a clinician deems the likely benefit of administering a medication to far exceed the potential medication risks. In other cases, overrides may represent not carefully considered clinical decisions, but reflexive dismissals by clinicians who have become inured to the large number of EHR alerts. EHR alert override rates are as high as 96%[15] and the override rates of drug-allergy alerts have increased over time.[11]
Several authors have considered the reasons behind these high override rates. More than two-thirds of all drug-allergy alerts presented to clinicians were for non-life-threatening allergic reactions,[11] raising the question of whether these alerts were important enough to have shown to clinicians in the first place. How much of a threat to a patient's health does a medication or other therapy need to pose to warrant interrupting a clinician's workflow with an alert? Only immediately life-threatening reactions? A low chance of temporary reversible harm? Perhaps something in between? In addition, almost one-third of medication alerts shown to primary care physicians (PCPs) in one study were repeats of alerts that had fired for the same patient in the last year.[16] These findings illustrate (1) the persistent and repeated decisions (appropriate or not) by clinicians to ignore alert guidance, suggesting the guidance is not clinically helpful nor likely to alter clinical management, and (2) the inability of current-state EHR systems to adapt alert-firing based on prior end-user actions.
The more alerts one experiences, the more likely one is to ignore them.[17] Clinicians overrode drug allergy alerts appearing two or more times more frequently than drug allergy alerts that appeared only once.[11] Reminder alerts to perform certain tasks were also less likely to be heeded with increasing numbers of alerts presented[16] as were responses to clinical trial recruitment alerts.[18]
Alerts can have significant costs. Notably, there is an opportunity cost to the time it takes clinicians to process alerts. McDaniel studied nearly 26,000 drug-drug interaction (DDI) alerts and found that the median time spent processing those alerts was about 8 seconds.[19] Schreiber and colleagues estimated the hourly cost of a physician at between $108 (U.S. dollars [USD])) and $234 per hour.[12] Using the lower value yields a physician time cost of about $0.03 per second; multiplying by 8 seconds gives an approximate cost of $0.24 per physician per DDI alert. This cost increases to $0.52 per DDI alert using the higher value for physician time. With alert override rates typically above 90%, it is reasonable to believe that much of the time physicians spend engaging with alerts represents lost productivity. Given that the above calculations are derived from DDI alerts, it is not possible to extrapolate the lost productivity cost related to other types of alerts. However, the cumulative cost to health care organizations across many thousands of alerts is still likely to be substantial.
Clinicians perceive alert fatigue. Peterson and Bates define alert fatigue as a “condition in which too many alerts consume time and mental energy to the point that both important warnings and clinically unimportant ones can be ignored.”[20] In a survey of 2,590 PCPs, 86.9% reported the alert burden was excessive and more than two-thirds indicated the number of alerts they received was more than they could manage effectively.[21] PCPs report experiencing a median of 63 alerts per day.[21] The perception of, and not the actual alert burden, was positively associated with burnout among PCPs.[22] Improving EHRs will require far more than measuring alert performance if we are to address alert effectiveness and burden.[23]
What has been done to manage alerts better, to increase their effectiveness, and reduce the burden that alerts pose to clinicians? Targeted deactivation of alerts deemed to be of low quality or low effectiveness is one approach.[24] In another study, 18 DDI pairs were downgraded to be no longer visible to clinicians (e.g., warfarin and enoxaparin).[12] This reduced the DDI alert rate by 10.3%, yet the alert override rate remained 96.7%. Another approach is for organizations to implement alerts according to severity settings that their drug knowledge vendors provide. However, such an approach can lead to variation in alert implementation, as drug knowledge vendors define alert severity, classify drugs, and categorize DDI differently.[25] Other authors have advocated running alerts silently as a starting point, so that the alerts are not visible to clinicians.[13] In doing so, organizations can gather data on alert performance to inform implementation decisions. Still others have promoted the notion of adaptive clinical decision support (CDS) that learns user behavior over time and has the capability to filter alerts to which a clinician has previously responded.[26] While current-state EHRs may allow individual users some discretion over the alerts that they wish to continue receiving, it is not clear how many EHRs, if any, may have this built-in adaptive alert filtering capability at present.
Another approach to effective alert management is to follow expert guidance about alert configuration and settings. Payne and colleagues advocated for the consistent use of seven core elements in DDI-alert presentations.[27] Despite the publication of high- and low-priority DDI alerts that should be interruptive and that should be demoted to noninterruptive status, respectively,[10]
[28] a later study revealed that even within the same vendor product no two EHR instances had the same alert settings.[25] There remains wide variability in alert settings and no prevailing standard of care about how to implement alerts.[25]
Why is managing alerts effectively so difficult? A variety of factors may contribute including regulatory mandates, public reporting initiatives, liability concerns, and other external pressures that may incline institutions to advocate for more rather than fewer alerts to avoid preventable harm.[29] In addition, variations in clinicians' drug knowledge and experience can make alerts that are appropriate for one clinician, perhaps a July intern, inappropriate for another, such as an experienced attending physician. Lastly, there is system inertia where alerts, once created and deployed, may accumulate over time and compete for ongoing attention and resources with new, higher health system priorities.
Several unanswered questions deserve researchers' attention. Is there an optimal override rate and if so does it vary depending on the type of alert and other contextual factors? Is there an absolute or relative reduction in the number of alerts that fire that can reduce the alert override rate? What is the best metric for evaluating the effectiveness and appropriateness of alerts? What guiding principles should determine whether alerts should be interruptive versus noninterruptive, hard stops versus soft stops? Are there CDS alternatives to alerts that could be widely scalable to achieve safety goals? What is the best way to lighten alert burden without jeopardizing patient safety and ideally in a way that demonstrably improves patient safety?[30] This paper explores the approaches found in the literature and that the authors, all of whom are clinicians, have found of benefit to answer these questions whenever possible, although much more research is needed.
Alert Governance
There is no prescriptive form of CDS governance that assuredly works for every institution.[4]
[31] What works for one may fail gloriously for another. An institution's organizational culture generally predicts whether a given governance structure works for it. There are numerous governance structures, worthy of a focused literature review. As an overview, most large, integrated delivery health care systems govern in a top-down approach,[31] often called hierarchical governance. Smaller institutions may use this approach but often use a consensus-driven model, usually termed either spoke and wheel governance or star network, where anyone can say “no,” but no one person can say “yes.”[32]
Regardless of the approach to governance, we recommend that alert governance include regularly scheduled reviews of all alerts that breach the organization's established thresholds, discussed in more detail below. Alerts should not be removed automatically but presented to a committee for review and consideration for removal or revision (if possible). Regular review will gradually prune the existing alerts with the threshold then adjusted as needed for more specific fine-tuning. In cases where alerts are turned off, vigilance for unintended consequences is an important goal, recognizing that identifying any safety event, let alone a safety event clearly attributable to deactivation of a single alert, is known to be extremely challenging.[33]
There is no widely accepted definition of successful alert governance. In the absence of one, it may be reasonable to define governance success as a process for evaluation and decision making that enables an organization to achieve predefined goals for its clinical decision support systems.
Governance Structure and Foundational Processes
On what do information technology (IT) departments and end-users agree? There is little argument about the need for the individuals and groups shown in [Table 1] to participate in alert governance and to be part of alert decision making from the beginning.
Table 1
Key participants in alert governance[a]
|
Subject matter expert groups
|
Role as participant in governance
|
|
Clinicians
Advanced practitioners
Nurses
Pharmacists
Physicians
|
Define clinical goal of alerts
May overlap with informatics specialists, if available
|
|
Informatics
Nursing
Pharmacy
Physician
|
Expertise regarding knowledge management
Arbiters between clinicians, technical staff, and administrators
Broad cross-sectional knowledge of clinical and technical domains
|
|
Information Technology
Analysts, builders
Data scientists, researchers
Education leaders
Human factors engineers
Optimization staff
|
Inform decision makers what is possible from a technical capability standpoint
Subject matter experts for issues as they arise
|
|
Administration
Legal staff
Regulatory
Safety officers
|
Define personnel, budgetary, time resources
|
a Not all institutions have the breadth or depth of all participants. If available, each has a unique role.
Despite the necessity of these governance participants, they are accompanied by sociotechnical domains which may impede effective governance as shown in [Table 2].
Table 2
Sociotechnical domains which may impede effective governance
|
Domains
|
Issues
|
References
|
|
Physicians
Other clinicians
|
Limited availability to participate
Lose interest rapidly
May not receive compensation
|
Authors' experience
|
|
EHR analysts
|
May not know system limitations early in project
May be few knowledgeable analysts available in health IT market for some applications
Failure to recognize alert errors and anomalies
|
Authors' experience[34]
[35]
|
|
Budget
|
Total cost of ownership not always clear
Allotments may not be realistic
|
[36]
[37]
|
|
Administration
|
Impaired or interrupted institutional knowledge due to:
Frequent hospital leadership turnover (avg.: 5.5 years)
Frequent medical executive leadership turnover (often 1 year)
Organizational priorities unlikely focused on system optimization
Prioritization of other projects
Risk intolerance variability
|
[38]
|
|
Regulatory
|
Rapid cycle changes (e.g., meaningful use, ICD-10)
|
Authors' experience
|
Abbreviations: avg., average; EHR, electronic health record; ICD, international classification of diseases; IT, information technology.
Governance structures evolve over time, sometimes as the result of changing organizational needs or leadership. Health care integration, too, can impact governance structures. For example, as community hospitals become members of larger health systems, the culture of governance often changes from local- to system-level control. A previously consensus-driven structure often yields to top-down governance, compelling institutions to reflect on and refine governance structures. Moreover, resources may expand with mergers in such a way that facilitates better governance. More informatics resources, however, may become available to community hospitals after health system integration.
A strong governance process adheres to a defined strategy and sets metrics and specific goals. Decision making accelerates when explicit rules exist regarding prioritization, metrics, and accountability.[39] In addition to making decisions about creation and removal of alerts, each operational governance body will have to weigh the value of alert optimization against other optimization efforts that may compete for the same resources. In the interest of efficiency, each governance committee should have the authority to make decisions about the topics presented to them, rather than serving in a purely advisory role to other decision makers.
When attempting to evaluate existing alerts, there may be staff available with the institutional knowledge of who created alerts, and why and when, who may be able to shed light on prior processes. Alert assessments should be performed in a nonjudgmental manner, recognizing that prior parties created the current alerts with good intentions and felt doing so was important. The original impetus for alert creation must be addressed, especially when removing an alert.
Governance includes setting up a categorization of alerts into standard groups, such as medication alerts or alert recipients which facilitates alert build processes, evaluation, and maintenance.[40]
[41] Prioritization models used by clinical informaticists, focusing on patient safety, revenue, and clinical workflow optimization, have been published.[42]
It is not uncommon for groups who will not receive an alert to request it on another group's behalf. Regardless of the requestor and respecting the intent of the request, a facilitator should clarify the goal and try to determine the original problem that the requestor is trying to resolve. Ideally both the requestor and the alert receiving party (if different) should be involved in a collaborative discussion to seek resolution of the original problem, with or without an alert.
New alert requests require justification. First, prior to creating an alert, there should be a system search for any alert or other CDS tool which already meets the intent of the proposed alert. Second, alert requests should be vetted by an informaticist (or by whatever means an organization determines) to determine whether an alert or reminder is the appropriate tool with which to achieve the desired objective. Once these two prerequisites are met, the request may come before an appropriate governance group. Consider use of an intake form which asks the submitter to delineate information consistent with the five rights of CDS.[43]
If an alert is the best solution to the issue and meets build criteria, we recommend following the guidance in [Table 3] before building the alert. In addition, ideal alerts meet our proposed seven CREATOR rules for alerts, also shown in [Table 3].
Table 3
Author-recommended checklist to justify alert appropriateness
|
Alert justification criteria
|
Questions to determine appropriateness
|
|
Considerations prior to build
|
|
Problem identification
|
What problem will the alert solve?
|
|
Beneficial
|
Is there a clearly defined return on investment (e.g., increased screening referrals)?
Will alert reduce potential adverse events?
|
|
Appropriateness
|
Is the alert consistent with clinician workflow?
|
|
No better alternative
|
Is there an existing clinical decision support that accomplishes the same thing?
Is there a less intrusive mechanism that may succeed?
|
|
Not so complex as to inhibit system performance
|
Does testing reveal system malfunction or slowing?
|
|
Metrics defined
|
How will alert success be measured?
|
|
Scheduled review of alert
|
When are the first and subsequent dates of review?
|
|
Consistent with organizational strategy and principles
|
Is the alert compatible with institutional policy, financial goals, and clinician workflows?
|
|
Considerations for alert build
|
|
CREATOR: 7 new rules for ideal alerts:
|
|
Consistent with organizational strategy and principles
|
Addresses high priority goals, adheres to established alert guidelines
|
|
Relevant and timely
|
Appropriate for clinical workflow impacted
|
|
Evaluable
|
Predefined metrics
|
|
Actionable
|
Allow delete or modify of triggering orders from within alert
|
|
Transparent
|
Rationale of alert is clear, highlighting patient-specific data which triggered the alert
|
|
Overridable
|
Clinical workflows may not be predictable by alert designers; clinician may be presented with a scenario which exceeds the alert designers' foresight
|
|
Referenced
|
Citing literature as appropriate, supporting intent of alert
|
After alerts have been implemented, health care organizations must recognize that a clinician who overrides an alert is making an individual patient level decision, while at the same time creating an important signal for the organization. Organizations should appreciate the value of these messages and seek to learn from them.[44] While we do not know of specific override rate goals, each organization should determine what alert override rates should prompt further alert revision, changes in alert actions, or further clinical education.
Following the guidance to “start with the end in mind,”[45] ideally governance design should welcome feedback about alerts from end-users. There should be a process to understand how alerts function in the production environment (see sections “Alert Metrics” and “Thinking of Alerts like Diagnostic Tests”), whether the result is as expected (see section “Designing Alerts”), and whether users find the alerts helpful, discussed further in section Case Examples of Alert Maintenance and Reduction “Geisinger Health System.”
Description of a Governance Structure at Geisinger Health System
At Geisinger Health System, there is a system Chief Medical Informatics Officer (CMIO) and three associate CMIOs who report to the system CMIO. Separate from the CMIO team, there are unique structures for nursing and pharmacy informatics. These groups collaborate regularly at conjoint informatics huddles, optimization meetings, and collaborative informational and educational meetings. Numerous ad hoc committees are created as different issues arise, such as order set optimization for enhanced postoperative recovery.
Geisinger has revamped its process for managing alert requests. Requestors must first obtain managerial approval. A team of nursing and physician informaticists and analyst staff then triages requests for new or altered alerts for assignment to one of seven major governance committees: alerts and reminders, documentation, orders and order sets, interface, optimization, usability, and education. Importantly, each of these groups is authorized to make decisions regarding rejection, acceptance, or altering of the requests. For more complex issues involving more than one constituency, or when there is disagreement, the request rises to the appropriate optimization group.
While this structure appears to function well in this large, integrated, dispersed, complex medical system, those creating governance in other settings must be cognizant of local culture, existing reporting structures, and mindful of prior governance that succeeded or failed. At Geisinger, this governance has been more successful than prior governance versions by decreasing the number of pending requests, accelerating the time to request completion, and increasing the ability to create system-wide consistency in alerting.
Governance is challenging yet essential for operational efficacy and efficiency. Informatics teams help, but are not always available, especially in smaller community hospitals. Participation by noninformatics staff is difficult, time-consuming, and often unsustained. At all institutions, governance structures evolve over time.[4]
Alert Metrics
Quantitative alert assessment is difficult in that there is no agreed upon measure to assess alert effectiveness and burden. Consider metrics such as the total number of alerts or an alerts per orders ratio, number of interruptive versus noninterruptive alerts, alert override rate, or time required to act on an alert.[12]
[19] Whether these metrics or some combination are the appropriate indicators of alert effectiveness and burden is unclear. Ideally, outcome metrics for each alert would establish the alert's clinical impact but measuring clinical outcomes directly attributable to alerts is challenging.[46] In the absence of outcome measures, process measures may still be beneficial. [Table 4] is a compendium of alert metrics which the authors have found in the alert literature.
Table 4
Alert metrics with definitions, advantages, and disadvantages
|
Parameter
|
Definition
|
Advantages
|
Disadvantages
|
|
Override rate
|
% of alerts dismissed[a]
|
Easy to calculate
|
No clear baseline or desirable target goal
|
|
Acceptance rate
|
% of alerts where user selected suggested action
|
Easy to calculate
|
No clear baseline or desirable target goal
|
|
Volume of alerts
|
Total number of alerts fired
|
Easy to calculate
|
Crude metric
|
|
Alerts/patient
|
Ratio of total alerts divided by number of patients on whom alerts fired
|
Easy to calculate
|
Crude metric that depends on illness severity, types of medications and orders
|
|
Alerts/clinician
|
% of alerts fired for an individual clinician
|
Easy to calculate; may offer ability to compare clinicians with similar types of patients
|
Comparing clinicians with different patient populations may not be meaningful
|
|
Alerts/patient/day
|
Ratio of total alerts divided by number of patients on whom alerts fired per day
|
An improved assessment of daily workload
|
Comparing different patient populations may not be meaningful
|
|
Alerts/clinician/day
|
% of alerts fired for an individual clinician
|
An improved assessment of daily workload
|
Comparing clinicians with different patient populations may not be meaningful
|
|
Alerts/orders entered
|
Proportion of total orders entered on which an alert fired
|
Another analysis of workload burden
|
Comparing different individual's order burden may not be meaningful
|
|
Alerts/order session
|
Average number of alerts fired per order session
|
May offer an alternative view of workload burden
|
Comparing different individual's order burden may not be meaningful
|
|
Alerts/specific order item
|
Alerts that fire on a given order
|
Helpful for analysis of specific orders where clinicians see frequent alerts
|
No clear baseline or desirable target goal
|
|
Harm occurred when clinician overrode true positive alert
|
Number of alerts correlated with adverse events reports
|
Perhaps the most pertinent in terms of PPV
|
Very difficult data to obtain
|
|
Dwell time[19]
|
Time that an alert is on screen
|
Easy to obtain metadata
|
May not reflect total cognitive load of alert
|
|
Think time[12]
|
Total time between firing of alert and resolving it
|
Easy to obtain metadata, only analyzes time of dealing with alert
|
May be an overestimate if user unfamiliar with screens, or underestimate for those who dismiss alerts rapidly
|
|
Effectiveness[47]
|
Proportion of patients on whom an alert fired where the clinician chose the alert's intended action
|
Measures success of alert on a per patient basis at achieving intended action
|
May be difficult to quantify if alert not actionable or several acceptable choices are possible;
May require chart review
|
|
Efficiency[47]
|
Proportion of alerts for which the clinician chose the intended action
|
Measures burden of alerts required to fire before intended action occurs
|
May be difficult to quantify if alert not actionable or several acceptable choices are possible;
May require chart review
|
|
Number needed to alert or prompt
|
Number of times alert needs to fire to elicit the intended action
|
A variation on effectiveness and efficiency, but measures total alert firing rate, rather than per patient
|
As a number, and not a ratio, may suffer from differing denominators
|
|
Outcome
|
Many possible definitions
|
Optimal metric for effectiveness
|
Proving cause (the alert) and effect (outcome) is challenging; concurrent and evolving changes create bias and confounding
|
Abbreviation: PPV, positive predictive value.
a Definitions of override vary.
One intriguing metric distinguishes alert effectiveness and efficiency. The former is the number of patients for whom the alert's intended action was taken divided by the total number of patients on whom the alert fired (what proportion of patients on whom the alert fired did it prompt the desired action?), while the latter is the number of alerts on which the desired action was taken divided by the total alerts fired (how many times did the alert fire on all patients to achieve its purpose?).[47] Values for effectiveness and efficiency could be very similar but could diverge when an alert fires multiple times for one clinician for a single patient or when the alert fires to multiple clinicians for a single patient.
Another complication with alert metrics is that there is not a uniform definition for what counts as an alert override. There is considerable commentary in the CDS literature that high override rates are unacceptable, yet there is no agreement about a proper rate and it is unclear if the override rate is even a valid measure of alert effectiveness. The authors' experience in the case of DDI alerts[25] showed that there were several options that were equivalent to an override action. Clinicians could click on a button called “override” to reject the alert guidance, with or without indicating a reason for doing so. Alternatively, clicking on the “X” dismisses the alert as does clicking “continue,” but in neither case is there an indication of the clinician's intention other than to skip the alert. For data analysis purposes, it is worth noting that some available override reasons may not even be relevant to the alert displayed.[48] In addition, dismissal of an alert does not always mean that the user ignored the advice; rather, the user may sometimes enter an appropriate order later. These examples illustrate the complexity of interpreting alert metrics.
Thinking of Alerts like Diagnostic Tests
Thinking of Alerts like Diagnostic Tests
Whenever possible, informaticists and others designing alerts and CDS tools need to think of them as they would a diagnostic test with true and false positives and negatives. Decision support relies on positive and negative criteria and whether the patient truly meets the condition in question (e.g., sepsis). These data are necessary to calculate the familiar test characteristics of sensitivity, specificity, and positive/negative predictive value (PPV/NPV; [Table 5]). Importantly, these test characteristics apply to the alert performance itself (whether it fired appropriately or not), rather than to the clinician's response to the alert (such as to accept, override, or dismiss).
Table 5
Calculation of an alert performance measure
|
Patient clinical condition
|
Performance measures
|
|
Condition present
|
Condition absent
|
|
Alert behavior
|
Condition criteria triggered
(alert fired)
|
(A) True positive: condition identified
|
(B) False positive: condition incorrectly identified
|
Positive predictive value: A/(A + B)[b]
[c]
|
|
Condition criteria not triggered
(no alert)
|
C) False negative: condition missed
|
(D) True negative: condition truly absent and alert correctly did not fire
|
Negative predictive value: D/(C + D)[b]
|
|
Performance measures
|
Sensitivity: A/(A + C)[a]
|
Specificity: D/(B + D)[a]
|
|
a Not affected by condition prevalence.
b Affected by condition prevalence.
c Easiest to calculate since both numerator and denominator are based on easily retrieved alert firing data. All other performance measures have either true or false negatives, both of which require methods like manual chart review or retrospective query of validated cohort criteria to calculate.
The consensus is that current alert configurations are overly sensitive, with rampant false positives and subsequent pervasive alert fatigue.[16] Where the field has struggled, and perhaps introduced the greatest potential harm, is the failure to incorporate PPV into the design of CDS and to consider the level of control (LOC). Level of control is the degree to which the alert is attempting to alter clinical decision making. For example, interruptive alerts which require entry of an override reason are much more controlling than those that are dismissible with one click. While a higher PPV is always a goal, this can be challenging, especially in the case of conditions with low prevalence or where the severity of outcomes, such as sepsis or cardiac arrest in the pediatric population, warrants a higher alerting LOC.
The key principle is that the PPV should align with the LOC. Without this alignment, which is typically low PPV and high LOC, there are two prominent risks: first, an often incorrect alert with onerous requirements contributes to alert fatigue, and second, a clinician may heed a false positive alert with high LOC (i.e., strongly recommending something) and take the wrong action for the patient, again potentially leading to harm. Over time, informaticists have concluded that alerts will always be more successful when the design and format, specifically the level of control, align with its test characteristics, perhaps most importantly the PPV.[24]
Designing Alerts
There are several types of interruptive alerts. A complete hard stop prevents the user from proceeding (e.g., trying to prescribe isotretinoin for a pregnant woman). Partial hard stops require that one cannot proceed without supplying required elements, while soft stops require the user to pause, even if data entry is not required. Soft stops are less controlling but may still contribute to end-user perceptions that all alerts can be a nuisance.
Noninterruptive or advisory alerts do not interfere with the user's workflow but may not be seen and clinicians may ignore them more readily. Language of an alert can also play a role in LOC, depending on the strength of the verbiage regarding the recommendation in the alert.
There is growing interest in novel designs of alert appearance which may improve user-computer interaction when applied to alerts.[49]
[50]
[51] Although this is a promising area of research, a full discussion of these topics is beyond the scope of this paper.
Alert Testing
There are limitations in the ability to test alerts, most notably that scenarios in testing environments typically represent a small fraction of the potential clinical variations that can trigger an alert in an environment with actual patient data. Moreover, scenarios for testing almost always are limited to anticipated true positive and negative behavior. This risks underestimation of false positive and negative alerts. Although feedback, after the alert is live, can lead to adjustments that improve an alert's performance, this is only after the potentially poorly performing alert has been live in the system with the associated risks that it could cause unintended patient harm and negative impressions on clinicians.[17] Finally, the task of correcting a problematic alert postdeployment can be time sensitive and is less likely to enable a thoughtful work environment for informaticists relative to the more planned and deliberative conditions predeployment.
Is there a better way to test alerts? The evidence is growing that alert testing can and should be performed with real, dynamic patient data, and fully functional interfaces.[52] Actual patient data rather than scripted testing scenarios will help to refine and improve the alert criteria and design. Once alert criteria are optimized, final test characteristics, specifically the PPV, should update alert formatting elements, such as LOC and language. There are two possible approaches that alert designers can take to achieve more rigorous level of testing and alert performance evaluation: retrospective analysis and background deployment.
Retrospective Analysis
Retrospective alert analysis is a relatively high-resource requiring method. It involves application of potential alert criteria using retrospective data. This approach offers several clear advantages. First, it can apply the logic of an alert to a very large cohort. Second, one can derive all four test characteristics (sensitivity, specificity, PPV, and NPV) for the alert, if there is a validated cohort group of patients who truly meet the criteria of interest ([Table 5]). In retrospective analysis, determining a specific cohort of patients defining a “true condition” may be a distinct analytics project, since the cohort definition data may not necessarily be the same as the criteria in an alert. For example, the criteria used for sepsis screening of a general population (vital signs, certain laboratory tests and assessment documentation) may differ from those describing a true sepsis cohort using data that may not be present at the time of screening (end-organ damage and positive culture data), but is present in the retrospective analysis, which will impact the PPV and NPV.
An additional programming challenge is that the degree of data manipulation necessary in retrospective analysis is greater than the manipulation and tools available in commercial EHRs. As such, informatics teams must guard against building a retrospective alert that is impossible to build and deploy for use with actual patients.
Background Deployment
“Background deployment” or “silent mode” means activating an alert in the live environment without making it visible to clinicians while recording all potential firings. This “lower tech” approach offers many of the same advantages of retrospective analysis while avoiding some of the limitations. It allows testing of the alert in a setting with similar limitations to those encountered by deployed, clinician-visible alerts, for example, erroneous data entries and lack of final cultures or final billing codes.
The primary limitation of the background approach, especially when compared with retrospective analysis, is the time it may take to gather enough data to perform an adequate performance analysis. While a retrospective query will have a large number of patient visits immediately available, a background alert must be left to run for some period of time for initial analysis and any subsequent criteria refinement steps. For high-volume alerts, this may be of little practical consequence, whereas for rare conditions, it could be a significant impediment to analysis. Fortunately for the purposes of minimizing alert fatigue and false positives, typically even a limited period of background analysis yields significant insights that can significantly reduce the potential alert burden.[53]
Either of these approaches to “going live before go-live,” in which teams can include in the alert planning phase accurate performance data, is strongly recommended, in particular for alerts where complex logic is involved. Both can then provide real-time prospective data which when combined with other variables, such as clinical severity of the targeted condition, can enable the configuration of an alert that is more likely to achieve its goals and avoid unintended consequences.
Case Examples of Alert Maintenance and Reduction
Case Examples of Alert Maintenance and Reduction
Geisinger Health System
Geisinger Health System has over 1,500 alerts and reminders in its Epic EHR (Verona, Wisconsin, United States) which it installed in 1996 (ambulatory) and 2006 (inpatient). More recently Geisinger contracted with a third party CDS software company (Stanson Health, Sherman Oaks, California, United States) which supplies its own CDS, as well as analyzes currently installed alerts, including those in silent mode. Stanson supplies and Geisinger reviews alert statistics regarding firing rates, override percentages, acceptance rates, alert comments, and other vital alert data.
Armed with this data, Geisinger turned off alerts that users always override or ignore, or which violate one or more of the five rights.[43]
[54] Comparing monthly alerting rates between January 2018 and January 2019 Geisinger reduced the absolute number of active alerts fired to nurses from 1,674,429 to 763,132 (54%) and for physicians, from 630,690 to 511,705 (19%). Alerts that were turned off included those directed to incorrect users or that were poorly timed (e.g., reminding nurses to obtain a patient's smoking history before the nurse had a chance to complete the nursing intake interview) or that included guidance that was inconsistent with current workflow.
Penn Medicine
Penn Medicine is a large academic and community-based institution that has undertaken efforts to optimize alerts. Penn Medicine also uses the Epic EHR and calls this optimization initiative EHR wellness, the practice of continuously analyzing the performance and efficacy of clinical decision support and other tools to assure that they are functioning appropriately and supporting clinician workflow as intended. The goal of this iterative process is to eliminate noisy and burdensome alerts that cause cognitive load on ordering providers, nurses, and pharmacists while optimizing the important and necessary alerts, all while continuing to support patient care. The EHR wellness program addresses interruptive and noninterruptive care-guidance alerts, medication alerts, and order sets. Only 14% of physicians find that they have the time they need to provide the highest standard of care,[55] which served as a driving force for the EHR wellness campaign. The campaign aims to proactively guide providers and make it easier to do the right thing at the right time in the EHR.
The Penn Medicine EHR wellness team includes operational and clinical leaders, technical analysts, and informaticists from the CMIO team. The team uses a third-party platform (Phrase Health Inc., Philadelphia, Pennsylvania, United States)[56] that provides detailed performance data on EHR alerts, which complements the EHR vendor-supplied reporting tools to pinpoint the largest areas of opportunity. [Table 6] lists the goals of the team.
Table 6
Guiding principles of electronic health record wellness
|
1. Correct design inconsistencies and tailor alerts to meet the needs of the target population
|
|
2. Engage directly with impacted clinicians to redesign workflows (user-centered design/optimization)
|
|
3. Make all alerts actionable: assure the ability to jump directly to the intended action
|
|
4. Set standards for inclusion/exclusion logic across all care settings so that alerts do not impact unintended areas or users
|
|
5. Review trigger actions, align acknowledgment reasons, streamline verbiage
|
|
6. Standardize analyst capture of metadata when alerts are changed, to assure a reliable record of alert adjustments and the reasons for them, as well as the routine review of the alerts during change control
|
The initial focus of the optimization efforts targeted 17 of the most “burdensome” alerts that accounted for nearly 1.7 million alerts/month across the health system. At Penn Medicine, an alert is considered potentially burdensome if it fires >100,000 time per month, if the alert has an elevated average number of interruptive alert firings per day for the population that is exposed to it relative to other alerts,[56] or if the alert has inconsistent build/design according to institutional standards. Once troublesome alerts were identified, a detailed review of individual alert settings was performed to assess for elements, such as acknowledgment option consistency compared with other alerts, ease of jumping to the intended action from the alert itself, and alert triggers. A fundamental question guiding the process was, “is this alert even necessary?” Three months of analysis and alert optimization resulted in complete removal of three of the most burdensome alerts and editing of the remaining 14 alerts. These changes resulted in the reduction of interruptive alerts by 67,863 alerts/month (45%), and overall alerts by 251,505 alerts/month (15%).
In August 2017, Penn Medicine initiated a secondary effort to evaluate EHR medication alert performance specifically. Baseline alert data revealed 675,613 alerts per month that users overrode 94.6% of the time. The goal was to safely reduce unnecessary medication alerts by 3 to 5% and provide more effective guidance to ordering providers, pharmacists, and nurses. Pharmacy residents, under the supervision of the Director of Pharmacy, conducted a literature review, analyzed the evidence available to support medication alerts, such as drug–drug and dose-range alerts, and made recommendations for which of the 20 alerts that fired most often should be continued, edited, or retired. These recommendations gained approval from key medical, nursing, and pharmacy leaders. The results of these combined efforts exceeded original expectations. Overall medication alerts decreased by 23%. The number of alerts per 100 orders dropped by nearly 34% as shown in [Table 7].
Table 7
Results of medication alert reduction
|
Medication alerts
(per mo)
|
Number of alerts
(per 100 orders placed)
|
Number of overridden alerts
(per mo)
|
Override rate[b]
|
|
July 2017
|
675,613
|
55.4
|
581,958
|
94.6%
|
|
August 2018
|
521,005[a]
|
36.8
|
445,088
|
92.9%
|
|
Difference
|
−154,608
|
−18.6
|
−136,870
|
−1.7%
|
|
Overall change
|
23% reduction in all medication alerts
|
34% fewer alerts per 100 orders
|
24% reduction in overridden alerts
|
∼2% reduction in overall override rate
|
a An additional approximately 215,000 orders placed (per month) attributed to an additional hospital going live on the EHR.
b Override rate only includes unfiltered alerts.
Of note, these efforts only minimally impacted the override rate for medication alerts, a conundrum previously noted.[12]
Discussion
Managing alerts within an EHR is a complex undertaking, with notable considerations being organizational history, expectations, and governance related to alerts. Complicating management is the lack of a widely accepted metric to judge the effectiveness and burden of alerts. New, scalable methods of evaluating alerts and responses to them are necessary[57] but this requires further research. High-alert volume can lead to alert fatigue which contributes to increased mental workload, potential patient harm via workarounds, and mistakes in ordering and treatment.[17]
[57]
[58]
There are no nationally-developed, endorsed standards for which alerts are appropriate and which are not, which ones should be interruptive and which ones should be passive. Some proposals are gaining acceptance, such as the list of DDIs that should be interruptive.[28] Standardization of alert nomenclature would enhance understanding and promote better research. This may also assist vendors to coordinate their alert design to enhance crossvendor comparisons. The authors believe that new standards, such as Fast Healthcare Interoperability Resources (FHIR), will enable enhanced CDS in general and alerting in particular by sharing successful processes across institutions. Machine learning will also offer new approaches to improve alerting through analysis of huge datasets.
Notwithstanding these limitations organizations have taken proactive steps toward optimizing alerts with some early success. Organizations can begin an alert optimization program by evaluating alerts with high firing or override rates, or those assessed to be burdensome on clinicians. Doing so will likely uncover alerts that are relatively less valuable and that may be optimized for better effectiveness or alternatively, deactivated. More work is needed to understand at what level of alert reduction clinicians respond more appropriately to the guidance of remaining alerts. When considering deployment of new alerts, analysis of alert performance prior to go-live can improve PPV and design before any clinician ever experiences the alert.
Conclusions and Recommendations
Conclusions and Recommendations
Alert management programs must strive to meet common goals of improving patient care while at the same time decreasing the alert burden on clinicians. In doing so, organizations have an opportunity to promote the wellness of patients, clinicians, and EHRs themselves. There are multiple components to ensure a successful alert management program:
-
Governance is complex but essential infrastructure for effective alert management.
-
Organizations should conduct ongoing analysis and review of alerts.
-
Absent an agreed-upon optimal metric for analyzing alert performance, each organization must select metrics appropriate for itself.
-
For guidance regarding implementation of an alert management program, look to organizations that have been successful and have reported their experiences.
-
New design paradigms, data and alert visualization displays, and emerging technologies also offer promise for improved alerting.
Clinical Relevance Statement
Clinical Relevance Statement
Electronic health record alerting provides tools for clinical decision support and can help clinicians to provide improved care, while also preventing medical errors. Yet there is widespread agreement that over-alerting leads to alert fatigue, with the subsequent risks of potential patient harm and clinician burnout. This paper presents the analysis and recommendations for mitigation of this problem from informatics leaders from four major health care organizations which may provide useful guidance for small and large as well as community and academic institutions.
Multiple Choice Questions
Multiple Choice Questions
-
Recognizing that an alert is a form of one-way communication (system to user), what is an objective way to measure its effectiveness?
-
Committee review and discussion to modify or delete alerts.
-
Reports showing the rate and changes of alert firing over time.
-
Total percentage of alerts which elicit the intended action.
-
User narrative feedback to modify, delete, or add alerts.
Correct Answer: Option c is the most objective and direct measure. The other answers have a higher potential for bias or false attribution of effect.
-
Institutions with a long history of clinical decision support (CDS) and those newer to creating governance for CDS, including alerts, share the same struggles regarding the scope and complexity of the task. The most critical factor to assure success in alert governance is:
-
Establish best infrastructure for alert installation.
-
Manage and lead change transformation processes.
-
Purchase of the newest technology from an outside vendor.
-
Reward staff for meeting or exceeding goals.
Correct Answer: The best option is b because use of CDS depends on supporting the workflow requirements of users. Although solid infrastructure is necessary, it is not sufficient. The latest and greatest technology is only as good as the ability of personnel to use it. Rewarding staff has merit but is also not sufficient.
-
Interruptive alerts should be used for critical decision-making processes. Which statistical measure is the most helpful in determining whether an alert should be interruptive?
Correct Answer: The best option is b. Several factors can determine the format of a CDS intervention, but of the statistical measures positive predictive value (true positives/all positives) is most useful in determining whether CDS should be interruptive and prescriptive.
