Electronic Medical Record Alert Associated with Reduced Opioid and Benzodiazepine Co-prescribing in High-Risk Veteran Patients

Abstract

Background.

Over the past decade, overdoses involving opioids and benzodiazepines have risen at alarming rates, making reductions in co-prescribing of these medications a priority, particularly among patients who may be susceptible to adverse events due to high-risk conditions.

Objectives.

This quality improvement project evaluated the effectiveness of a medication alert designed to reduce opioid and benzodiazepine co-prescribing among Veterans with known high-risk conditions (substance use, sleep apnea, suicide-risk, age ≥65) at one VA healthcare system.

Methods:

Prescribers were exposed to the point-of-prescribing alert for 12 months. For each high-risk cohort we used interrupted time series design to examine population trends in co-prescribing 12 months after alert launch adjusting for co-prescribing 12 months prior to launch, demographics and clinical covariates. Trends at the alert site were compared to those of a similar VA healthcare system without the alert. Secondary analyses examined population trends in opioid and benzodiazepine prescribing separately.

Results:

Over 12 months, the alert activated for 1,332 patients. Proportions of patients with concurrent prescriptions decreased significantly post-alert launch among substance use (AOR=0.97, 95% CI=0.96-0.99, 12-month decrease=25.0%), sleep apnea (AOR=0.97, 95% CI=0.95-0.98, 12-month decrease=38.5%), and suicide-risk (AOR=0.94, 95% CI=0.91-0.98, 12-month decrease=61.5%) cohorts at the alert site. Decreases in co-prescribing were significantly different from the comparison site among suicide risk (AOR=0.92, 95% CI=0.86-0.97) and sleep apnea (AOR=0.98, 95% CI=0.96-1.00) cohorts. Significant decreases in benzodiazepine prescribing trends were observed at the alert site only.

Conclusions:

Medication alerts hold promise as a means of reducing opioid and benzodiazepine co-prescribing among certain high risk groups.

Introduction

Opioids and benzodiazepines are the most common prescription classes involved in pharmaceutical overdoses, with substantial increases in drug overdose fatalities involving these medications over the last decade.^1-3 Although clinical practice guidelines (CPG) discourage their co-prescribing due to risks,⁴ dual use of opioids and benzodiazepines is common.⁵ The abuse potential, side effects of sedation, respiratory depression and impaired coordination, and risks of intentional⁶ and unintentional^7,8 overdose and injury⁹ associated with concurrent use of these medications may be higher among particular individuals, including those with substance use disorders (SUDs),¹⁰ suicide risk,^11,12 and sleep apnea,¹³ as well as the elderly (age ≥65).¹⁴

Recent estimates suggest that 27% of Veterans enrolled in VA care who are prescribed opioids also are prescribed benzodiazepines.¹⁵ The risk of overdose among Veterans who are co-prescribed opioids and benzodiazepines is 3- to 4-times greater compared to those prescribed opioids alone.¹⁵ In response the 2014 VA Opioid Safety Initiative (OSI) prioritized reductions in opioid and benzodiazepine co-prescribing and other non-recommended opioid prescribing practices.¹⁶ Given that factors driving co-prescribing of opioids and benzodiazepines are complex (e.g. provider beliefs that medication benefits outweigh risks^17-19 and lack of medication/behavioral alternatives,^19,20 provider time,¹⁹ and system-level tapering strategies¹⁹), tools are needed to assist in reduction efforts.

Clinical decision support (CDS) systems^21,22 offer one approach to decrease non-recommended prescribing practices. Errors at the point of prescribing,²³ common and costly sources of preventable adverse events,²⁴ are caused by inadequate knowledge about drugs and patient-level contraindications, drug therapy errors (e.g. wrong/duplicative drug prescribed),^23,25 and system factors, such as heavy workload.^21,22,25 Advanced medication alerts are one type of CDS system that perform complex processes such as reviewing the electronic medical record (EMR) for patient-specific risk factors²¹ that contraindicate medication use. Such systems have demonstrated effectiveness in reducing non-recommended prescribing.^26-29

This quality improvement project evaluated the implementation of an advanced medication alert designed to identify Veterans with known high-risk conditions (SUD, suicide risk, sleep apnea, age ≥65) who were co-prescribed opioids and benzodiazepines. The alert was implemented in 2014 at one multi-site VA healthcare system and evaluated over a 1-year period. Specifically, we examined: 1) alert activity at the patient- and provider-level; 2) changes in opioid and benzodiazepine co-prescribing among patients activating the alert; and 3) population-level trends in opioid and benzodiazepine co-prescribing at the implementation site and a comparison site in the 12 months preceding and following alert launch among Veterans with targeted risk conditions.

Methods

Setting

VA Puget Sound Health Care System (VAPSHCS) includes two medical centers and seven community-based outpatient clinics (CBOCs). Investigators partnered with mental health (MH), primary care (PC), pharmacy, and specialty-pain service leadership at VAPSHCS, with input from VA Central Office, to develop a medication alert aimed at improving the quality and safety of clinical care. The project was considered a quality improvement evaluation and did not require VAPSHCS Institutional Review Board approval.

To assess if co-prescribing trends following alert launch were due to the alert itself or secular trends, a comparison VA healthcare system was selected based on its similarity to VAPSHCS with respect to facility complexity, size and organization (two urban medical centers with a system of CBOCs). Further, no local projects aimed at reducing opioid and benzodiazepine co-prescribing were carried out at the comparison site during the implementation period. Prescription and other data generated at the comparison site were utilized, but no alert-related activities were completed there.

Procedures

Figure 1 provides a project overview. Initial steps involved engagement of pharmacy, specialty pain, PC and MH services. Pharmacy buy-in was essential due to its leadership in reducing unwanted prescribing practices and role in granting approval for new EMR-based medication alerts. Specialty-pain provided expertise and guidance. Given that opioids most frequently are prescribed in PC clinics and benzodiazepines in MH clinics,⁹ leadership support from both services was critical. Following engagement of key services, approval was obtained to access resources for alert development.

Figure 1.

Project Overview

Advanced Medication Alert.

The alert was designed in consultation with facility leaders, informed by the CDS literature,^30-32 and underwent pilot testing by PC and MH prescribers. Three criteria were required to activate the alert: 1) prescriber ordered an outpatient benzodiazepine/opioid medication, 2) patient had an active VA or documented non-VA prescription for the other medication class, and 3) patient had a risk condition documented in the EMR in the past 12 months. Risk conditions (detailed in Supplemental Table 1) activating the alert were: SUD diagnosis or severe alcohol misuse per the Alcohol Use Disorders Identification Test-Consumption (AUDIT-C)³³; suicide risk, defined as a patient record flag,³⁴ suicide attempt, or inpatient mental health admission;³⁵ sleep apnea; and age ≥65. The alert activated at the point of prescribing within the EMR,³⁶ synthesizing patients’ current benzodiazepine and/or opioid prescriptions and risk factors that activated the alert (see example in Supplemental Figure 1). Relevant EMR details were provided (e.g. substance use or sleep apnea diagnosis date), enabling prescribers to review information. The alert was not activated if no targeted risk factors were present. Providers could ignore or override the alert without justification.

In the month before alert launch, project leadership (EH) attended staff meetings in PC and MH clinics to introduce the alert and address concerns. Prescribers (114 PC and 67 MH) were informed about project goals and components, including an anonymous survey¹⁹ and qualitative interviews with prescribers and leadership (data pending), details about alert activation (conditions required for and frequency of alert activation), and launch date.

Provider Reports.

In addition to point-of-prescription alerts, PC and MH prescribers received reports every 4 months for 12 months identifying patients on their panels who met alert criteria. The report included: specific risk condition(s) present, recent benzodiazepine and opioid prescription details (fill date, medication name, dose, quantity), and an indicator of long-term use (≥90 days’ supply in prior 4 months). Reports complemented the alert by synthesizing key information about all patients on providers’ panels, allowing for targeted review to aid decision making with respect to tapering/discontinuing.

Data Source

This study used the VISN 20 Data Warehouse, a data repository for VA facilities in the Pacific Northwest that stores patient-level data from the VA EMR. Demographics, diagnoses, AUDIT-C results, suicide risk indicators, and utilization data were used to identify risk cohorts described above at both VAPSHCS and the comparison site. Prescription data from VA outpatient pharmacy, fee basis pharmacy, and non-VA medication files were used to identify all opioid and benzodiazepine prescriptions for cohort patients. In addition, alert-related data elements, including patient and prescriber identifiers, date/time of alert activation, and high-risk conditions that caused alert activation, were pulled for VAPSHCS patients and linked to pharmacy data.

Outcomes

The primary outcomes were change in the proportion of patients who were co-prescribed opioids and benzodiazepines (medications listed in Supplemental Table 2) at VAPSHCS in the 12 months before and after alert launch for each of the four risk conditions. These measures included all patients meeting risk condition criteria as described above in the year preceding and/or following alert launch (those age ≥65 were required to have an outpatient visit during the 24-month period to ensure that they were receiving VA care). Patients were excluded if they died or had a cancer diagnosis in the 24-month period. A binary indicator (yes/no) was used to identify patients who were co-prescribed opioids and benzodiazepines by month. Changes in co-prescribing trends at VAPSHCS were compared to those at the comparison site without the alert.

Secondary outcomes included: 1) change in proportion of patients with each risk condition prescribed opioids in the 12 months before and after alert launch, 2) change in proportion of patients with each risk condition prescribed benzodiazepines in the 12 months before and after alert launch, 3) among patients activating the alert, changes in proportion prescribed benzodiazepines, opioids and both medication classes in the 6 months after patients’ initial alert activation.

Covariates. Covariates for primary analyses were selected based on associations with benzodiazepine and opioid use, factors potentially impacting co-prescribing in the 12 months before and after alert launch, and availability in the EMR. Demographic covariates included age, gender, race (White, Black, other, unknown), and ethnicity (Hispanic/Latino, non-Hispanic/Latino, unknown). Clinical covariates included pain, mental health and medical comorbidity. A binary indicator of pain was determined by at least one outpatient or inpatient Internal Classification of Diseases, Ninth Revision Clinical Modification (ICD-9-CM) diagnosis indicating chronic pain.³⁷ A count (0, 1, ≥2) of ICD-9-CM disorders (anxiety, depressive, bipolar, psychotic and eating disorders) was calculated to measure mental health comorbidity. A modified version of the Charlson Comorbidity Index (CCI),³⁸ calculated from ICD-9-CM codes that identify 17 health conditions, measured medical comorbidity.³⁹ CCI scores were assigned to three groups: 0, 1, ≥2. Clinical characteristics were based on EMR data from one year prior to alert launch (July 15, 2014) unless otherwise specified. An indicator of opioid treatment program (i.e. methadone/buprenorphine treatment) attendance during the 24-month period, and a time-varying indicator (0, 1) of VA visit attendance for each of the 24 months, were derived from outpatient visits. Lastly, an indicator (0, 1) reflecting when the VA OSI was implemented in May 2014 was included in case it influenced prescribing practices.

Data analysis

Descriptive statistics were used to characterize patient- and provider-level alert activity. Monthly benzodiazepine, opioid and concurrent medication fills in the 6 months following patients’ initial alert activation were examined using counts and percentages. Patients receiving ≥90 days of both opioids and benzodiazepines (long-term co-prescriptions) were examined separately from those receiving <90 days of one or both medications (short-term co-prescriptions). Although the alert activated in all outpatient settings, only PC and MH prescribers were involved in implementation efforts; analyses described above are limited to patients seen by these prescribers.

Primary analyses examining population-level change in co-prescribing trends in the 12 months before and after alert launch were evaluated using segmented logistic regression with interrupted time series design.⁴⁰ Models estimated change over time in the proportion of patients in each risk group who were co-prescribed opioids and benzodiazepines after alert launch, adjusting for co-prescribing trends 12 months prior to launch. Models included terms for the outcome (co-prescribed opioids and benzodiazepines, 0, 1) for each of the 24 months, month since the beginning of the pre-launch period (1 to 24), month since alert launch (1 to 12), and a pre-/post-alert launch indicator (0, 1). Segmented logistic regressions used generalized estimating equations (GEE) with an auto-regressive (AR1) correlation structure and were adjusted for covariates described above. Analyses comparing changes in co-prescribing trends at VAPSHCS to the comparison site used similar adjusted models with interaction terms for site X month since beginning of pre-launch period, site X month since alert launch, and site X indicator of pre-/post-alert launch. Analyses were conducted using Stata MP, version 13.⁴¹

Results

Alert Activity

As seen in Table 1, the alert activated a median of 3 times per patient over 12 months, with the most common reason for activation being age, followed by sleep apnea and SUD and small numbers (<5%) for suicide risk. Nearly one-quarter of patients had two or more risk conditions. In the 6 months preceding their initial alert activation, two-thirds of patients filled opioid prescriptions for ≥90 days, with slightly fewer (60%) filling benzodiazepine prescriptions of ≥90 days. Fewer than 20% of patients were newly prescribed opioids and benzodiazepines at the time of their initial alert activation.

Table 1.

Overview of 12-Month Alert Activity

	n (%) or Mdn (IQR)
Total Patients with Activated Alert	1,332
Number of Times Alert Activated per patient, Mdn(IQR)	3 (1-7)
Risk Factor Activating Alert, n(%)*
≥65	854 (64.1)
Sleep Apnea	397 (29.8)
Substance Use Disorder	356 (26.7)
Suicide Risk	61 (4.6)
Total Risk Factors, n(%)
1	1,026 (77.0)
2	277 (20.8)
3 or 4	29 (2.2)
Opioid Use in the 6 Months Prior to Initial Alert
Days of Medication Use
0	194 (14.6)
>0 and <90	256 (19.2)
≥ 90	882 (66.2)
Morphine Equivalent Daily Dose, Mdn(IQR)**	32.8 (17.2-68.1)
Benzodiazepine Use in the 6 Months Prior to Initial Alert
Days of Medication Use
0	246 (18.5)
>0 and <90	287 (21.6)
≥ 90	799 (60.0)
Diazepam Equivalent Daily Dose, Mdn(IQR)**	15.3 (9.8-27.9)
Total Providers with Activated Alert	153
Number of Times Alert Activated per Provider, Mdn(IQR)	22 (7-54)
Unique Patients per Provider with Alert Activated, Mdn(IQR)	10 (5-20)
Total PC Providers with Activated Alert	100
Number of Times Alert Activated per Provider, Mdn(IQR)	32 (13.5-80.5)
Unique Patients per Provider with Alert Activated, Mdn(IQR)	12 (6-24)
Total MH Providers with Activated Alert	53
Number of Times Alert Activated per Provider, Mdn(IQR)	13 (5-26)
Unique Patients per Provider with alert Activated, Mdn(IQR)	8 (4-11)

Mdn: Median; IQR: Interquartile range

^*Percentages add to >100% because more than one risk factor may be present.

^**Excludes patients who did not receive medication prior to initial alert.

The alert activated a median of 22 occasions per provider, with PC providers treating more patients with activated alerts and viewing the alert on more occasions relative to MH providers.

Prescribing Patterns among Patients with Activated Alert

Patterns of concurrent, opioid and benzodiazepine fills in the six months following patients’ initial alert are presented in Figure 2. Among patients who filled long-term (≥90 days) prescriptions for both opioids and benzodiazepines in the prior 6 months (n=615), concurrent fills fell from 97% to 69% in the 6 months following the alert. Patients with shorter-term (<90 days) prescriptions for one or both medications in the prior 6 months decreased concurrent fills from 64% to 22% over 6 months. Note that while we would anticipate 100% of patients would have concurrent fills when they first triggered the alert, some patients had active prescriptions (medication ordered by provider) but had not filled their medication that month. Further, benzodiazepine and opioid fills may not have overlapped, resulting in a concurrent fill percentage lower than that of either opioids or benzodiazepines alone.

Figure 2. Percent of Patients with Medication Fills in the Six Months Following Patients’ Initial Alerts (N=1332).

^a Individualized to each patient activating the alert

^b ≥90 days in 6 months prior to intial alert

^c <90 days in 6 months prior to intial alert

Population-Level Prescribing Patterns

Population-level trends in co-prescribing 12 months before and after alert launch for each risk cohort at VAPSHCS and the comparison site are presented in Table 2 and Figure 3. After adjusting for covariates and co-prescribing trends in the prior 12 months, significant monthly decreases in co-prescribing were seen post-alert launch at VAPSHCS among SUD (p=0.010), sleep apnea (p<0.001), and suicide risk (p=0.006) cohorts. This translated to 12-month decreases in co-prescribing of 25.0% (95% CI: 5.9%–43.7%) post-launch, relative to 18.5% (6.2%–30.6%) pre-launch, in the SUD cohort, a decrease of 38.5% (21.7%–55.1%), relative to an increase of 1.9% (−13.8%–9.9%), in the sleep apnea cohort, and decreases of 61.5% (18.2%–103.2%), relative to 6.5% (−25.3%–37.4%), in the suicide risk cohort. Decreases were not detected among those aged ≥65. At the comparison site, no significant decreases in co-prescribing were detected in any risk group post-alert launch. Larger monthly decreases were observed in suicide risk (AOR=0.93, 95% CI: 0.87–0.99, p=0.022) and sleep apnea (AOR=0.98, 95% CI: 0.96–0.99, p=0.012) cohorts at VAPSHCS relative to the comparison site.

Table 2.

Monthly Change in Likelihood of Prescription Fill in the 12 Months Following Alert Launch by Risk Cohort.

	Alert Site			Comparison Site
	AORa	95% CI	p-value	AORa	95% CI	p-value
Substance Use Disorder
Concurrent Medications	0.979	0.964 - 0.995	*	0.994	0.977 - 1.011
Opioids	0.984	0.978 - 0.991	***	0.993	0.986 - 1.000	*
Benzodiazepines	0.976	0.966 - 0.986	***	0.998	0.987 - 1.009
Age ≥65
Concurrent Medications	0.990	0.977 - 1.003		0.988	0.974 - 1.003
Opioids	0.990	0.986 - 0.995	***	0.992	0.987 - 0.996	***
Benzodiazepines	0.991	0.984 - 0.999	*	0.995	0.987 - 1.004
Sleep Apnea
Concurrent Medications	0.968	0.954 - 0.982	***	0.993	0.979 - 1.006
Opioids	0.984	0.978 - 0.989	***	0.985	0.980 - 0.990	***
Benzodiazepines	0.977	0.968 - 0.986	***	0.999	0.989 - 1.008
Suicide Risk
Concurrent Medications	0.949	0.914 - 0.985	**	1.023	0.970 - 1.078
Opioids	0.964	0.945 - 0.984	***	0.994	0.971 - 1.018
Benzodiazepines	0.971	0.948 - 0.994	*	0.994	0.963 - 1.027

AOR: Adjusted odds ratio. 95% CI: 95% confidence interval.

Note: Sample sizes were as follows: Substance Use Disorder – Alert Site: N=10,657, Comparison Site: N=10,754; Age ≥65 – Alert Site: N=35,326, Comparison Site: N=32,725; Sleep Apnea – Alert Site: N=12,690, Comparison Site: N=13,628; Suicide Risk – Alert Site: N=1,304, Comparison Site: N=1,141.

^*p<0.05

^**p<0.01

^***p<0.001

^aAdjusted for prescription fills in the 12 months prior to alert launch, age, gender, race, ethnicity, pain, mental health and medical comorbidity, opioid treatment program participation, VA visit attendance, and OSI implementation.

Figure 3. Percent of Patients Prescribed Concurrent Medications over 24 Months by Risk Group.

Greater decreases in co-prescribing following alert launch were observed at the alert site relative to the comparison site in the suicide risk (AOR=0.93, 95% CI: 0.87, 0.99, p=0.022) and sleep apnea (AOR=0.98, 95% CI: 0.96, 0.99, p=0.012) cohorts after adjusting for covariates and co-prescribing trends in the year prior to launch.

Separate analyses for opioids and benzodiazepines found significant monthly decreases in the likelihood of opioid (all p-values<0.001) and benzodiazepine (all p-values<0.05) fills for all risk cohorts at VAPSHCS post-alert launch (Table 2). At the comparison site, significant monthly decreases were seen in opioid prescribing among the SUD, age ≥65, and sleep apnea cohorts (all p-values<0.05), with no differences seen in benzodiazepine prescribing in any cohorts. No significant differences between VAPSHCS and the comparison site were seen in opioid prescribing trends in any risk group. However, larger monthly decreases in benzodiazepine prescribing were detected among SUD (AOR=0.98, 95% CI: 0.97–0.99, p=0.008) and sleep apnea (AOR=0.98, 0.97–0.99, p=0.001) cohorts at VAPSHCS relative to the comparison site.

Patient characteristics for each risk cohort at VAPSHCS and the comparison site are presented in Supplemental Table 3.

Discussion

While advanced medication alerts have demonstrated effectiveness in reducing non-recommended prescribing practices, to our knowledge, this is the first examination of an alert specifically targeting opioid and benzodiazepine co-prescribing. Given current trends in overdose deaths involving opioids and benzodiazepines,⁵ tools to assist prescribers in reducing co-prescribing are needed. To prevent “alert fatigue,” this quality improvement project attempted to reduce co-prescribing by focusing on patient groups at greatest risk of adverse events. Further, the project targeted two distinct prescriber groups, those in PC and those in MH. The alert appeared to have a wide reach, activating for 153 providers and 1,332 patients within PC and MH clinics at one VA healthcare system. Changes in co-prescribing trends both amongst those patients who triggered the alert, particularly those with long-term use, and at the population level indicated that the alert holds promise in reducing co-prescribing of opioids and benzodiazepines among certain subgroups of high-risk patients.

The most robust decreases in co-prescribing were observed in sleep apnea and suicide risk cohorts, with significant changes in trends from pre- to post-launch at VAPSHCS and significantly greater decreases than the comparison site, suggesting that changes might be due to the alert rather than secular trends. It is unclear whether changes seen in some cohorts but not others were due to provider attitudes about targeted risk factors, provider knowledge with respect to patient characteristics, or other factors. In a survey of PC and MH prescribers who participated in this project,¹⁹ over 90% of MH prescribers agreed/strongly agreed with CPG recommendations of caution in co-prescribing across all risk groups. Fewer PC prescribers (76%) agreed/strongly agreed with recommendations regarding patients ≥65 and with sleep apnea, with more agreeing with recommendations with respect to SUD and suicide risk (83% and 89%, respectively).¹⁹ Given the generally high endorsement of caution across all groups, it may be that the alert conveyed information regarding sleep apnea and suicide risk that is not readily known and/or accessible in the EMR to all prescribers, in contrast a factor such as patient age, which is readily apparent. This alert was developed to display specific risk factors activating the alert, including the source data (e.g. visit date) associated with those factors, in a succinct format.^30-32 While presence of a suicide flag is conveyed to all providers when accessing a patient’s EMR per VA policy,³⁴ psychiatric hospitalizations and suicide-related diagnostic codes in the prior year are not, and providing prescribers with this additional information may have influenced prescribing decisions. Likewise, sleep apnea diagnoses may not be as readily known/accessible to prescribers as information on SUDs, which likely require more frequent treatment contacts and impact multiple domains of patient functioning. Co-prescribing in those ≥65 was notably lower in comparison to the other cohorts, suggesting that prescribers may already have taken steps to reduce co-prescribing in this group.

While there were notable decreases in co-prescribing among long-term users of these medications, 69% of these patients continued to receive concurrent prescriptions 6 months after their initial alert. In some cases this may have resulted from a slow tapering schedule appropriate to patient needs. However, despite prescribers’ general agreement with the goal of reducing co-prescribing, multiple factors may influence clinical decision making and complicate achievement of this goal, including beliefs that tapering/discontinuing would cause patients to suffer, ¹⁹ doubt about benefits of discontinuing long-term treatment among patients who are stable,^17,18 and needs for additional time to negotiate tapering schedules,^17,18,42 alternative medication- and behavioral-based treatments,¹⁹ and clinic-wide policies discouraging co-prescribing.¹⁹ While medication alerts can raise prescriber awareness about particular risk factors, they represent one piece in larger efforts to address prescribers’ concerns and resource needs.

Interestingly, the comparison site saw decreasing trends in opioid prescribing following alert launch in 3 of 4 risk cohorts but no changes in co-prescribing or benzodiazepine prescribing, whereas at VAPSHCS changes were seen in co-prescribing, as well as benzodiazepine and opioid prescribing alone. Current VA initiatives to decrease opioid prescribing¹⁶ likely have concentrated efforts in PC settings, where most opioid prescribing occurs. This project, which focused both on PC and MH, may have encouraged changes in benzodiazepine prescribing, resulting in a greater drop in co-prescribing for at least some cohorts. Over one-third of prescribers (34% PC, 45% MH)¹⁹ targeted by this project endorsed the belief that it is too difficult to coordinate tapering/discontinuation when medications are prescribed by different prescribers. Given that benzodiazepines are primarily prescribed in MH,⁹ and prescribers are often wary of addressing medications prescribed by others, it is critical to bring MH prescribers into efforts to reduce co-prescribing.

Limitations

This quality improvement project has several limitations. Because data on cancelled prescription orders were not retained in the data repository,⁴³ we could not determine whether the medication alert resulted in cancelled benzodiazepine or opioid prescriptions. Our data do not allow us to determine if decreased co-prescribing resulted in fewer adverse events. Future studies are needed to assess whether changes in co-prescribing result in changes in adverse events, including death. This alert was implemented at one VA healthcare system, limiting the generalizability of results to other VA or non-VA facilities. Population-level analyses compared the implementation site to a single VA healthcare system; we cannot comment on how co-prescribing at the implementation site compared to other similar healthcare systems or to the VA nationally. Given the number of initiatives within VA to reduce risky opioid prescribing practices, it is possible that changes in co-prescribing trends seen here were due to secular trends rather than the alert itself, although analyses involving the comparison site mitigate this concern somewhat. Although we included non-VA medication files in our analyses, we were not able to capture all prescriptions received in the community. Patients may have sought prescriptions from community prescribers following alert launch to compensate for decreases in VA prescribing. Because individual patients typically received opioid and benzodiazepine prescriptions from multiple providers, we were not able to adjust analyses for ordering provider. Additional study is needed to assess the impact of the alert on individual providers. Lastly, our analyses do not address whether other aspects of the larger project, such as provider reports, were associated with changes in prescribing trends.

Conclusion

Advanced medications alerts hold promise as a means of reducing co-prescribing of opioids and benzodiazepines among certain high-risk groups, most notably those with suicide risk indicators and sleep apnea. Our findings suggest that the ability of the alert to reach both MH and PC prescribers and raise their awareness of patient-level contraindications may have contributed to greater decreases in co-prescribing at the implementation site relative to the comparison site. While medication alerts do hold promise, most long-term users with an activated alert continued to be co-prescribed these medications six months later. Future work should examine whether combining medication alerts with other interventions (e.g. academic detailing, audit and feedback, greater availability of behavioral supports for tapering and/or pain management) that support providers and patients further decreases the number of patients prescribed this risky medication combination.