We developed an alerting system integrated with computerized provider order entry to continuously monitor and inform health care providers about medication safety in the setting of acute kidney injury. The intervention significantly improved provider response to rising creatinine levels, increasing the frequency at which nephrotoxic and renally cleared medications were modified or discontinued. The greatest effect occurred for medications with the highest potential for toxicity. Response rates exhibited a modest increase during the post-intervention period, possibly demonstrating a learning effect. The findings of our evaluation are important in defining how clinical decision support can reduce drug selection and dosing errors that are a significant proportion of overall ordering and monitoring errors27,28
. Improved management of medications in patients with AKI may reduce drug toxicity and represents improved compliance with the standard of care for hospitalized patients.
The alert system featured two levels of workflow intrusiveness to maximize provider attention to important prescribing issues while avoiding an over-abundance of interruptive alerting for medication dosing problems of lower urgency or clinical significance. The passive alerts for medications requiring review, which persisted on the main screen until the alert expired or an expected drug action was taken, did not change the prescribing behavior of providers. Similar to previous studies using clinical decision support, the interruptive alerts were more effective at producing change in prescribing behavior than passive alerts42,43
. We could not determine whether the low response rates to passive alerts were due to provider unawareness of the alert or deliberate judgment that no action was required.
The system allowed deferral of the alert so that users not primarily responsible for medication decisions could pass responsibility for responding to the alert to the next user. Providers frequently used this option, sometimes to apparent excess, which may signal that the medication decision needs to be escalated to a more knowledgeable member of the care team, such as the attending physician or rounding pharmacist. Targeting the alert only to the responsible team member (i.e. all members of the primary team and covering teams) would be a desirable enhancement to the current approach, where all users receive the alert.
The high rate of deferred and overridden alerts is similar to other evaluations of clinical decision support. Chertow, et al.14
and Galanter, et al.17
also found high rates of alert overrides (42% and 48% respectively) in systems which targeted medication safety in chronic or acute kidney disease. In our study, some of the overrides appeared to have been appropriate due to low initial dosing such as a prophylactic dose of acyclovir in transplant patients. Patients may have been experiencing transient changes in serum creatinine, and providers may have anticipated that the changes would be reversed with fluid administration. When deferring or overriding the alerts, providers may also be demonstrating alert fatigue or uncertainty about what therapeutic changes to make in the face of AKI. Our analysis of provider interactions showed that although initial deferral rates were high, more than half of alerted drugs are eventually modified or discontinued.
The decision support capabilities developed for this intervention have not previously been combined in a single system. A number of systems successfully provided guidance for renally dosed drugs when providers initially prescribed medications14,17,20,44
. Other systems have implemented surveillance to alert providers about kidney function changes but have distinct differences to our system15,29,45
. One developed by Rind, et al.29
generated alerts for 55 medications that appeared to providers as e-mail messages outside of the workflow. The study found a 21.6 hour reduction in physician response time, compared to our finding of 19.6 hours for a comparable set of high toxicity drugs. Systems developed by Evans, et al.15
and Kilbridge, et al.45
also include a surveillance approach, though the systems assessed for changes only once daily, and the alerts appeared to pharmacists rather than ordering providers. Alerts displayed to pharmacists may complement those displayed to physicians or other prescribers, as pharmacists may filter alerts for priority and clinical significance. However, they may not always have updated knowledge of the patient condition, or do not have sufficient staffing to monitor all patients at all times. By continuously monitoring for updated lab results and alerting ordering providers directly in the computerized provider order entry system, we gave prescribers the opportunity to make dosing changes more quickly.
Our study has a number of limitations. The measured outcomes may have been affected by the changing case mix of patients admitted during the two sequential study periods or parallel efforts within the institution to improve prescribing quality. However, the regression analysis suggests that the change was abrupt and associated with activating the intervention. The attempt to exclude dialysis patients from the intervention using a user-defined flag and a filter for low baseline calculated GFR was imperfect and may have led to both false positive and false negative alerting. A more reliable coded method to identify dialysis patients is needed.
Because recommendations for patients with rapidly changing creatinine are difficult to provide without significant error, the intervention did not give specific dosing advice, and we considered both medication discontinuations and dose reductions as responses. We are taking steps to further evaluate the intervention by retrospectively reviewing cases and adjudicating the appropriateness of alert responses. Approaches to providing recommendations in the future may include providing approximate dosing recommendations for mild, moderate, or severe AKI, or facilitating the expert dosing advice of a pharmacist or nephrologist.
Finally, it is not known whether the intervention prevented adverse drug event rates. The interactions between kidney function and medications are numerous and difficult to measure in a comprehensive fashion for 122 medications. Additionally, there is no validated method to distinguish nephrotoxic drug effects from native kidney disease in patients selected for a changing creatinine.
In conclusion, we developed, implemented, and evaluated an intervention to continuously monitor for and alert providers within the computerized provider order entry system about AKI in the presence of target nephrotoxic and renally cleared drugs. The intervention significantly increased the rate and timeliness of provider modification or discontinuation of target drugs.