We developed a novel clinical-decision support tool to prevent adverse drug events associated with the administration of warfarin. We found that stealth alerts—alerts triggered by a clinician’s prescription but delivered to the anticoagulation management service, rather than to the prescribing clinician—increased the proportion of patients who underwent INR monitoring within 5 days following the prescription of a medication that could potentially interact with warfarin. The effect of this intervention was robust, with evidence of effectiveness in increasing monitoring rates as early as 3 days following the alert and as long as 7 days following the alert, and a result that was consistent in time-to-event analysis. Although this study was not designed with sufficient power to detect a reduction in adverse drug events, more vigilant monitoring of INR following the co-administration of potentially interacting medications, as demonstrated in this study, would likely prevent both hemorrhagic and thrombotic complications of warfarin therapy.
Clinical-decision support aids have been shown to reduce co-prescription of warfarin with potentially harmful interacting medications.1
When these co-prescriptions are required, targeted alerts to prescribers can increase the margin of safety.4
Alerts and reminders may improve therapeutic monitoring of medications. Many of these medication safety alerts have been built into EHR’s to enhance patient safety.13,14
However, few prior studies of computerized clinical-decision support have focused on improving warfarin monitoring.15,16
To our knowledge, no prior studies have evaluated alerts targeted to accelerate monitoring in the specific clinical context of the co-prescribing of warfarin and potentially interacting drugs.
Four novel dimensions of this intervention deserve mention. First, we were concerned that busy primary care clinicians would have “‘alert fatigue,”3,17
and might not respond to the new co-prescription safety alerts, having already just seen the standard drug-drug interaction alert in each case. Others have proposed more parsimonious alerting (e.g., fewer triggers in a given system18
) or tailoring alerts to the specific characteristics of individual patients.4,19
Our strategy was to direct these alerts to the AMS managers, rather than to the ordering clinicians. Kesselheim and colleagues recently suggested a similar approach of directing alerts to individuals who have responsibility for implementing them.2
The stealth alerts in this study fostered sharing of responsibility between the primary care team and the disease management team, which can focus on the safety issues pertinent to its area of expertise.20
Integrating specialized care for patients with chronic conditions is a hallmark of the patient-centered medical home.21
Second, while co‐prescribing alerts might cause the clinician to modify the prescription, they infrequently provide guidance regarding laboratory monitoring or follow up.15,16,22
The stealth alert guided the AMS manager to arrange appropriate follow‐up with the patient, which was operationalized as an accelerated INR test if determined to be clinically indicated.
Third, this intervention could be easily replicated in other health care systems using the Epic® EHR; currently more than 250,000 physicians in the US use Epic®, and by 2013, more than 127 million patients’ records are expected to reside on an Epic® EHR.23
Moreover, the “stealth” approach is not specific to Epic®, and could be implemented in any EHR meeting the standards of the Certification Commission for Health Information Technology.24
Fourth, this alerting system functions irrespective of the pharmacy where the patient fills the prescription, because the alert is triggered by the clinician’s medication order, not by the dispensing of the medication. Although this feature does create some false positive alerts (i.e., medication ordered but not filled by the patient), the communication by the AMS with the patient can easily resolve these alerts.
The magnitude of the intervention’s effect in this study was considerably smaller than that seen in other studies of alerts to improve monitoring. For example, Feldstein and colleagues found that reminders in the EHR sent to the prescribing clinicians more than doubled the rate of monitoring.22
In that study, the patients were not enrolled in a formal program for anticoagulation monitoring; in such programs, as in the present study, patients would be expected to have more vigilant routine monitoring, making it challenging for any intervention to demonstrate a large effect.
The modest size of the intervention effect—a 5 % increase, from 34 % monitored before intervention to 39 % monitored after intervention—should not diminish the potential value of this approach. Many of the co-prescriptions that triggered the alert, such as a 3-day course of a short-acting antibiotic, would have been considered low-risk by the AMS and, therefore, would not have prompted more urgent monitoring. Furthermore, we recognize that in our EHR, many of the alerts occurred in the context of what appeared to be a new prescription for a potentially interacting agent, but instead was a re-writing of an ongoing steady regimen of a medication that would not require accelerated monitoring. Given these considerations, one would not necessarily expect 100 % monitoring within 5 days of co-prescribing. In this context, an absolute 5 % increase may appear more meaningful.
In addition, other factors may have converged to limit the effectiveness of the intervention. Although an alert may have guided the AMS staff to instruct the patient to complete urgent laboratory monitoring, in some cases patients may not have complied with this recommendation. Unfortunately, our data do not afford us the opportunity to examine the individual conversations between AMS and patients to characterize the nature of these interactions.
Several other study limitations should be considered. First, because we did not have a concurrent comparison group, there is a possibility that changes temporally associated with the intervention were attributable to secular factors other than the intervention itself, such as a general increase in ordering INR tests for all patients, regardless of timing. To address this issue, we used an interrupted time-series analysis to demonstrate a stable monthly outcome rate in the pre-intervention period.
Second, as mentioned above, although the stealth alert was designed to occur only in the context of a new prescription for a medication potentially interacting with warfarin, we learned that the alert was also occurring in some cases of a medication refill, e.g., patients receiving a standing dose of phenytoin or simvastatin. This occurred when the prescribing clinician created a new order in the EHR, rather than selecting to “reorder” a standing medication. These “false positives” could have diluted the effect of the intervention, as patients on established medications may not need the same monitoring intensity as patients on new medications. Additionally, the excess alerts could have had the effect of desensitizing the AMS staff, thereby reducing their effectiveness.
Third, conversely, some clinicians may have ordered a new medication using the “re-order” or “refill” function of the EHR. For example, in a patient treated in the remote past with trimethoprim-sulfamethoxazole now presenting with a new infection, the clinician might simply refill the previously prescribed medication from the medication history list, rather than writing a completely new prescription. In this clinical situation, an alert would be appropriate, but none would occur, since the system did not recognize this prescription as “new.” Using a system that would permit certain drugs (such as antibiotics) to initiate alerts at the time of refills, while excluding alerts on drugs more likely to be refills (such as statins), would prevent this problem.
Fourth, on occasion, clinicians prescribed potentially interacting medications that were not filled or taken by the patients. Although the prescription would generate an alert to the AMS manager, the AMS manager could make the determination, after communicating with the patient, that accelerated INR testing was unnecessary. In such cases, lack of a follow-up INR would not imply failure of efficacy of the alert. We suspect that the rate of unfilled prescriptions was small and relatively constant throughout the entire study period, thereby minimizing its potential to diminish the measured effect of the alert.
Finally, the study did not measure outcomes such as intracranial bleeds or strokes, or the time in therapeutic range, which would have been a reasonable surrogate measure for efficacy of the alert. However, we note that time in therapeutic range has consistently hovered at over 80 % for our AMS, with no differences discerned before or after the implementation of the alert.
The results of this study support the conclusion that stealth alerts result in higher rates of anticoagulation monitoring following the initial prescription of medications that potentially interact with warfarin in patients enrolled in an AMS. Increased INR monitoring in these situations likely results in improved patient safety, since the co-prescription of these medications may cause INR values to drift higher or lower than the anticoagulation goals and increase the risk of complications. Earlier monitoring in these situations fosters more frequent warfarin adjustments, and therefore likely maintenance of anticoagulation goal ranges, in addition to prevention of potential adverse effects of over-anticoagulation and under-anticoagulation. Future studies should examine the effect of alerts to accelerate laboratory monitoring on rates of patient outcomes, such as adverse drug events. Because stealth alerts have the potential to reduce the burden of excessive clinical-decision support, this approach should be explored for reducing alert fatigue and improving the quality of care for other chronic disease management systems.