Keywords
opioids - musculoskeletal pain - chronic pain - clinical decision support - implementation - alert
Background and Significance
Background and Significance
Musculoskeletal diseases affect approximately half of U.S. adults and up to three-quarters of older adults.[1] Globally, 1.3 billion people are living with a musculoskeletal disorder (e.g., rheumatoid arthritis, osteoarthritis, low back pain, neck pain, gout), contributing to 138 million years lived with disability.[1] Simultaneously, opioid use disorder is the seventh leading cause of disability-adjusted life-years in the United States, and overdose deaths remain high.[2]
[3] Therefore, clinicians are serving a large population of patients suffering from chronic musculoskeletal pain during an epidemic of opioid use disorder.
Due to the increased risk of overdose and opioid use disorder, prescribing opioids requires clear evidence of benefit to the patient.[4]
[5]
[6] However, the literature does not support significant improvement in pain or function for osteoarthritis patients on chronic opioid therapy as compared with nonopioid analgesics.[7]
[8]
[9] While opioids may have short-term efficacy for chronic back pain, their long-term effectiveness is unknown and complicated by drug tolerance and hyperalgesia.[10]
[11] Adverse effects include dependence on opioids after arthroplasty, inadequate analgesia following arthroplasty, poor surgical outcomes, hypogonadism, constipation, and decreased pain tolerance.[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19] Overall, existing American Academy of Orthopaedic Surgeons clinical practice guidelines (CPGs) do not include opioids due to limited efficacy and risk of adverse effects.[8]
To provide guidance in balancing chronic pain management and the risks of opioids, the Centers for Disease Control and Prevention (CDC) released a CPG in 2016, and then provided an updated guideline in 2022.[20]
[21] This release was associated with significant changes in opioid prescribing nationwide (e.g., lower prescribing rate, fewer high-dose prescriptions).[22]
[23] An interrupted time series assessing impact of CDC guideline among surgical patients found total morphine milligram equivalents (MME) prescribed postoperatively decreased, particularly following hip and knee replacement.[24] The 2016 CPG was associated with a decline in opioid dispensing across most specialties; greatest among family medicine clinicians and surgeons.[25] It was also associated with reduction in high-risk opioid prescriptions for chronic pain patients (i.e., coprescribing opioid and benzodiazepines)[26] and a decrease in opioid dose and days' duration at discharge following surgery.[27] However, some literature found no change in opioid outcomes following release of the guideline and targeted quality-improvement interventions.[28]
CDS interventions might improve uptake of the CPG into clinical practice. In one example, default prescriptions in accordance with guidelines in electronic health record (EHR) increased guideline-concordant prescriptions even more than the release of the CDC guideline.[29] A scoping review of CDSS for prescribing opioids for chronic noncancer pain in primary care identified 14 studies assessing CDSS tools. While prescription drug monitoring programs (PDMPs) were the most common CDS tool and no studies assessed patient outcomes, all studies found that CDSS led to improved prescribing practices.[30] A larger systematic review of interventions targeting opioid prescribing identified 142 articles.[31] PDMPs were the only type of clinical decision support intervention assessed. PDMPs were associated with a lower prescribing rate; however, it is unknown whether that decrease is clinically appropriate or represents guideline-concordant care. Overall, it seems CDS interventions are often effective in improving behaviors, like prescribing naloxone, using urine drug screenings, and prescribing for a short duration, but not necessarily for other prescribing outcomes such as MME.[29]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
Objectives
This study addressed the following research questions: (1) “Was implementation of a CDS intervention associated with a decrease in the percent of chronic musculoskeletal pain patients receiving opioid prescriptions and/or opioid dose?,” and (2) “Which prescriber and facility characteristics are associated with adherence with safe opioid-prescribing practices in response to a CDS intervention?”
Materials and Methods
Study Design
After approval from the Institutional Review Board, we conducted an interrupted time series analysis to assess trends from 2016 to 2020 and the change in percentage of patients receiving an opioid and the average opioid dose associated with implementation of a CDS toolkit to operationalize the CDC CPG in October of 2017. We also conducted a retrospective cohort study to assess the association between prescriber and facility characteristics and safe opioid-prescribing practices.
CDS Intervention
This health care system launched a clinical decision support tool called Prescription Reporting with Immediate Medication Utilization Mapping (PRIMUM) in 2015 in the Cerner platform (Kansas City, Missouri, United States).[39] In October 2017, additional CDS tools were added to the PRIMUM platform to address controlled substance prescribing and operationalize the CDC CPG ([Supplementary Table S1] [available online only]). Specific components of the CDS tool assessed in this study are described below. These components were included into the Cerner Opioid Toolkit, which is available to organizations using Cerner.
Original PRIMUM alert: this alert notifies the prescriber in real-time of the presence of potential risk factors for abuse, misuse, and diversion of prescription opioids. The PRIMUM alert includes eight patient risk factors. Relevant to this study, it provides an alert when a prescriber is initiating an opioid prescription for a patient with a current benzodiazepine prescription or initiating a benzodiazepine prescription for a patient with a current opioid prescription. Prescribers have the option to cancel the prescription, proceed with the current prescription, or make changes based on the alert.
Extended Release alert: this alert triggers when a prescriber initiates an extended-release opioid for an opioid-naïve patient (defined as no opioid prescribed in past 30 days). Prescribers can either continue or cancel the prescription.
Pain Agreement alert: this alert suggests completion of a standardized pain agreement for patients who have exceeded 90 days of continuous opioid therapy (90 days of the past 110 days) with no current (within past year) pain agreement in their record. Prescribers can either click to launch the pain agreement or continue without starting a pain agreement.
Naloxone alert: this alert suggests a naloxone prescription for patients at high risk of overdose (patients receiving opioids >50 MME, receiving concurrent benzodiazepine and opioid prescriptions, or with a history of opioid or benzodiazepine overdose documented via diagnosis code at any point in their medical record) and no current naloxone prescription.
Controlled Substance Review Component: a page in the EHR was created to display all controlled substance information, including prescriptions and on-site administrations, as well as all risk factors in the PRIMUM alert “on demand.” This means a prescriber can proactively access this page before initiating a prescription. In addition, this page displays the patient's opioid dose in MME. If the patient has an MME greater than or equal to 90, the text displays in red, indicating high risk.
Study Population
The target population was patients presenting to a large health care system from 2016 to 2020 with a diagnosis of a chronic musculoskeletal condition ([Supplementary Table S2] [online only]). We included arthropathies (inflammatory and degenerative) and chronic conditions of the spine (deforming dorsopathies, spondylopathies, and other dorsopathies) because they are common, chronic, painful musculoskeletal conditions in which opioid therapy is not recommended. We removed records missing facility name. To exclude postoperative prescriptions, we limited to encounters in the ambulatory setting. Furthermore, once a patient had an inpatient visit with an orthopaedic surgeon, we removed all subsequent ambulatory visits during the study period for that patient. Finally, we excluded urgent care and oncology encounters. To assess opioid dosage, we removed records missing information necessary to calculate MME ([Fig. 1]).
Fig. 1 Encounter and patient selection.
Data Collection
All data were sourced from the EHR and the PRIMUM database generated by the CDS intervention described above, which included prescriber information, prescription information, and prescriber response to each alert.
Variables
The main outcome is safe opioid-prescribing practices, operationalized as a composite score of the frequency at which a prescriber does the following behaviors in response to the CDS intervention: (1) cancels an opioid prescription when alerted the patient is already prescribed a benzodiazepine; (2) initiates a pain agreement when alerted a patient has reached >90 days of continuous opioid therapy; (3) prescribes naloxone when alerted the patient is at high risk for overdose; (4) cancels a prescription for an extended-release opioid when alerted that patient is opioid-naïve; and (5) prescribes opioids ≤90 MME. We calculated a prescriber-level weighted percentage of such encounters to obtain a composite score ranging from 0 to 100 using the formula below. Weighting was necessary to prevent a small number of encounters in any one category from leading to an over- or under-representation of the prescriber's overall opioid safety behavior.
where E
coprescribe is the number of encounters in which the prescriber received the coprescription alert and p
cancelled is the proportion of those encounters in which the prescriber cancelled the prescription and did not prescribe an opioid; E
90 day is the number of encounters in which the prescriber received the 90-day alert and p
agreement is the proportion of those encounters in which the prescriber initiated a pain agreement; E
high risk is the number of encounters in which the prescriber received the high risk of overdose alert and p
naloxone is the proportion of those encounters in which the prescriber prescribed naloxone; E
ER is the number of encounters in which the prescriber received the extended release alert and p
cancelled is the proportion of those encounters in which the prescriber cancelled the extended release opioid (they could have prescribed an immediate release opioid); N
rx is the number of opioid prescriptions written by the prescriber and p
<90MME is the proportion of those prescriptions that are less than or equal to 90 MME.
The independent variable is implementation of the CDS toolkit in October of 2017. Data prior to October 2017 were “preimplementation” and data on or after October 2017 were “postimplementation.”
The number of patients prescribed an opioid, the number of opioid-prescribing encounters, specialty, and type of prescriber were included as prescriber-level covariates. In addition, we summed up patient-level variables to the prescriber level to indicate case mix regarding diagnosis, gender, age, and race. Finally, we included the facility type (large multisite practice vs. single clinics) and whether the facility specialized in musculoskeletal conditions (i.e., orthopaedic surgery, neurosurgery, or rheumatology) as facility-level covariates.
Data Analysis
Research Question 1: we plotted the rate of opioid prescribing and the average MME by month from 2016 to 2020. We used an interrupted time series linear regression analysis to determine if the percentage receiving an opioid prescription or the average MME changed significantly after implementation of the CDS interventions as compared with baseline. Because chronological data are subject to autocorrelation, we used Newey–West autocorrelation-adjusted standard errors with linear regression with a lag order of 1.
Research Question 2: we included all “postimplementation” encounters for the population described above. We removed prescribers who prescribed an opioid to <10 patients during the study. We used descriptive statistics to characterize prescribers and facilities. We compared median safe opioid-prescribing score by prescriber and facility characteristics using Kruskal–Wallis H-tests and Spearman's rho. We utilized a two-level hierarchical linear regression model to predict safe opioid prescribing, accounting for fixed and random effects at the prescriber and facility levels. We assessed model fit and selected the most appropriate model using Akaike information criterion and Bayesian information criterion (BIC) values.
Results
After applying the inclusion/exclusion criteria, we included 1,289,697 encounters in the time series assessing whether an opioid is prescribed, 154,299 encounters in the time series assessing MME, and 606 prescribers (79,741 encounters) in the hierarchical linear regression analysis ([Fig. 1]). [Fig. 2] displays the percentage of encounters resulting in an opioid prescription and average MME of opioid prescriptions over time, with the implementation of the CDS intervention depicted by a vertical line. For percent of patients receiving an opioid, the preintervention trend was statistically significant, with rates of opioid prescriptions decreasing by 0.2% per month (β1, trend preintervention) ([Table 1]). After controlling for this trend, the level change following implementation of the intervention decreased by 1.6% (β2, level change postintervention, p < 0.001). The postintervention trend differed significantly from the preintervention period (β3, change in trend postintervention, p = 0.044). The postintervention slope was also significant (p < 0.0001), with rates continuing to decrease by approximately 0.1% per month (β1 + β3). Rates of opioid prescribing increased modestly in April 2020. For average MME, the only significant finding was the postintervention slope, with average MME decreasing by 0.16 MME per month (b1 + b3, p < 0.0001). Change in MME was not associated with the intervention.
Fig. 2 Percentage of encounters resulting in opioid prescription and average morphine milligram equivalents over time. Note: the vertical line indicates when the clinical decision support intervention was implemented.
Table 1
Interpretation of regression coefficients, interrupted time series, encounters resulting in opioid prescription, and average MME
|
Percent receiving an opioid
|
Average MME
|
|
|
Parameter
|
Estimate
|
Standard error
|
p-Value
|
Estimate
|
Standard error
|
p-Value
|
Interpretation
|
|
β0
|
0.175
|
0.003
|
<0.0001
|
59.779
|
1.069
|
<0.0001
|
Intercept
|
|
β1
|
−0.002
|
0.000
|
<0.0001
|
−0.074
|
0.071
|
0.298
|
Trend preintervention
|
|
β2
|
−0.016
|
0.004
|
0.0002
|
0.479
|
1.046
|
0.649
|
Change in level postintervention
|
|
β3
|
0.001
|
0.000
|
<0.0001
|
−0.087
|
0.080
|
0.279
|
Change in trend postintervention
|
|
β1 + β3
|
−0.001
|
–
|
<0.0001
|
−0.161
|
–
|
<0.0001
|
Trend postintervention
|
Abbreviation: MME, morphine milligram equivalent.
Descriptive characteristics for the encounters included in the retrospective cohort study (n = 79,741 encounters) are provided in [Supplementary Table S3] (available online only); however, this analysis was conducted at the prescriber level (n = 606 prescribers). Overall, the mean safe opioid-prescribing score was 74.4%, and the median was 77.1%. The prescriber population ([Table 2]) was mostly physicians (73.1%) who practiced as family or internal medicine clinicians (86.1%). They prescribed opioids mostly to patients who were ages 18 to 64 (59.3%), white (83.3%), and female (60.6%). The most common diagnosis was “other dorsopathies” (74.1%). The median number of patients clinicians prescribed to was 42. About half of these prescribers practice at large, multisite practices (52.5%); very few (10.6%) practice at a clinic specializing in musculoskeletal conditions.
Table 2
Bivariate analyses comparing prescriber and facility characteristics and safe opioid prescribing score
|
Sample,
n (%)
Median (Q1, Q3)
(n = 606)
|
Safe opioid prescribing score
Median (Q1, Q3);
Spearman correlation coefficient
|
p-Value
|
|
Prescriber characteristics
|
|
Prescriber type
|
|
|
<0.0001
|
|
Physician
|
443 (73.1%)
|
80.8 (73.3, 87.5)
|
|
|
Advanced practice provider
|
163 (26.9%)
|
75.7 (64.2, 83.3)
|
|
|
Specialty
|
|
|
0.1240
|
|
Family practice/internal medicine
|
522 (86.1%)
|
76.4 (66.0, 84.6)
|
|
|
MSK
|
65 (10.7%)
|
78.6 (73.5, 85.5)
|
|
|
Other
|
19 (3.1%)
|
77.8 (71.9, 92.3)
|
|
|
Case mix—age
|
|
|
|
|
Percent patients, <18
|
0 (0, 0)
|
−0.035
|
0.3929
|
|
Percent patients, 18–64
|
59.3 (48.4, 70.5)
|
0.245
|
<0.0001
|
|
Percent patients, ≥65
|
40.7 (29.5, 51.6)
|
−0.244
|
<0.0001
|
|
Case mix—race
|
|
|
|
|
Percent patients, white
|
83.3 (71.8, 91.4)
|
−0.269
|
<0.0001
|
|
Percent patients, black
|
14.8 (7.1, 24.1)
|
0.255
|
<0.0001
|
|
Percent patients, other
|
0 (0, 2.3)
|
0.021
|
0.6099
|
|
Case mix—gender
|
|
|
|
|
Percent patients, female
|
60.6 (51.8, 71.0)
|
0.115
|
0.0045
|
|
Case mix—diagnosis
|
|
|
|
|
Percent encounters, dorsopathy
|
0 (0, 1.5)
|
−0.133
|
0.0011
|
|
Percent encounters, inflammatory polyarthropathy
|
6.8 (3.6, 11.1)
|
−0.058
|
0.1539
|
|
Percent encounters, osteoarthritis
|
18.5 (9.4, 30.7)
|
−0.045
|
0.2650
|
|
Percent encounters, other dorsopathy
|
74.1 (58.8, 83.3)
|
0.037
|
0.3580
|
|
Percent encounters, spondylopathy
|
5.7 (2.0, 11.1)
|
−0.113
|
0.0054
|
|
Number of patients
|
42.0 (20.0, 88.0)
|
−0.408
|
<0.0001
|
|
Facility characteristics
|
|
|
|
|
Large, multisite practice
|
318 (52.5%)
|
78.6 (69.3, 86.7)
|
<0.0001
|
|
Yes
|
|
|
|
|
No
|
288 (47.5%)
|
75.0 (63.9, 82.6)
|
|
|
MSK specialty clinic
|
|
|
0.1201
|
|
Yes
|
64 (10.6%)
|
78.6 (73.6, 85.5)
|
|
|
No
|
542 (89.4%)
|
76.9 (66.1, 84.6)
|
|
Abbreviation: MSK, musculoskeletal.
Note: Bold values are statistically significant at the p < 0.2 level and included in multivariable model.
We included the variables that were significant at the p <0.2 level in a multivariable linear regression model ([Table 3]) and in building a mixed-effects model ([Table 4]). While the BIC was slightly higher for Model 4 as compared with Model 3, we selected Model 4 to align our empirical data with our conceptual model of important covariates at both the prescriber and facility levels.
Table 3
Multivariable linear regression, safe opioid prescribing score
|
Estimate
|
p-Value
|
|
Intercept
|
86.1
|
<0.001
|
|
Prescriber characteristics
|
|
Prescriber type
|
|
|
|
Physician
|
–
|
–
|
|
Advanced practice provider
|
4.70
|
0.0002
|
|
Specialty
|
|
|
|
Family practice/internal medicine
|
–
|
–
|
|
MSK
|
2.59
|
0.5355
|
|
Other
|
0.13
|
0.9649
|
|
Case mix—age
|
|
|
|
Percent patients, 18–64
|
0.09
|
0.0060
|
|
Case mix—race
|
|
|
|
Percent patients, white
|
-0.16
|
<0.0001
|
|
Case mix, gender
|
|
|
|
Percent patients, female
|
0.05
|
0.2482
|
|
Case mix—diagnosis
|
|
|
|
Percent encounters, dorsopathy
|
−0.22
|
0.3035
|
|
Percent encounters, inflammatory polyarthropathy
|
−0.04
|
0.3654
|
|
Percent encounters, spondylopathy
|
0.05
|
0.4660
|
|
Number of patients
|
−0.08
|
<0.0001
|
|
Facility characteristics
|
|
Large, multisite practice
|
|
|
|
Yes
|
2.77
|
0.0113
|
|
No
|
–
|
–
|
|
MSK specialty clinic
|
|
|
|
Yes
|
1.87
|
0.6465
|
|
No
|
–
|
–
|
Abbreviation: MSK, musculoskeletal.
Table 4
Estimates for two-level linear mixed-effects models of safe opioid prescribing score (N = 606)
|
Model 1
|
Model 2
|
Model 3
|
Model 4[a]
|
|
Model description
|
No predictors, just random effects for intercept
|
Model 1 + prescriber level fixed effects
|
Model 2 + random slopes for prescriber level predictors[b]
|
Model 3 + facility level fixed effects
|
|
Fixed effects
|
|
Intercept
estimate (standard error)
|
73.26 (1.12)
|
83.52 (4.84)
|
84.57 (4.71)
|
83.97 (4.72)
|
|
Prescriber type
|
|
Physician
|
|
–
|
–
|
–
|
|
Advanced practice provider
|
|
2.96 (1.21)
|
2.82 (1.18)
|
2.92 (1.19)
|
|
Specialty
|
|
Family Practice/internal medicine
|
|
–
|
–
|
–
|
|
MSK
|
|
2.08 (2.93)
|
0.77 (2.81)
|
2.39 (3.80)
|
|
Other
|
|
−1.70 (2.97)
|
−1.64 (2.89)
|
−1.20 (2.90)
|
|
Case mix—age
|
|
Percent patients, 18–64
|
|
0.08 (0.03)
|
0.08 (0.03)
|
0.08 (0.03)
|
|
Case mix—race
|
|
Percent patients, white
|
|
−0.17 (0.03)
|
−0.16 (0.03)
|
−0.17 (0.03)
|
|
Case mix—gender
|
|
Percent patients, female
|
|
0.05 (0.04)
|
0.05 (0.04)
|
0.04 (0.04)
|
|
Case mix—diagnosis
|
|
Percent encounters, dorsopathy
|
|
−0.13 (0.20)
|
−0.15 (0.19)
|
−0.13 (0.19)
|
|
Percent encounters, inflammatory polyarthropathy
|
|
0.01 (0.06)
|
0.03 (0.05)
|
0.03 (0.05)
|
|
Percent encounters, spondylopathy
|
|
0.08 (0.06)
|
0.05 (0.06)
|
0.06 (0.06)
|
|
Number of patients
|
|
−0.10 (0.01)
|
−0.11 (0.01)
|
−0.11 (0.01)
|
|
Large, multisite practice
|
|
Yes
|
|
|
|
3.75 (1.87)
|
|
No
|
|
|
|
–
|
|
MSK specialty clinic
|
|
Yes
|
|
|
|
−4.65 (4.51)
|
|
No
|
|
|
|
–
|
|
Error variance
|
|
Level-2 (attending) intercept
|
54.60 (16.34)
|
39.58 (11.36)
|
27.87 (10.53)
|
25.62 (9.74)
|
|
Random effects
|
|
Number of patientssite
|
|
|
0.002 (0.001)
|
0.002 (0.001)
|
|
Model fit
|
|
AIC
|
4,938.5
|
4,759.4
|
4,747.8
|
4,747.1
|
|
BIC
|
4,945.6
|
4,790.2
|
4,781.0
|
4,785.0
|
Abbreviations: AIC, Akaike information criterion; BIC, Bayesian information criterion; MSK: musculoskeletal.
Note: Bold indicates p < 0.05; ICCSite = 0.2354; values based on SAS PROC MIXED. Estimation model = maximum likelihood. Parameter estimates are shown with standard errors in parentheses.
a Best-fitting model.
b Inclusion of random slopes for each of the prescriber-level covariates produced too complex of a model; we fit models including each prescriber-level covariate as a random slope individually. Only percent of white patients and number of patients were statistically significant. When including both in the model as random effects, the percent of white patients was no longer statistically significant, and the model fit did not improve. Therefore, Model 3 only includes the number of patients as a random slope.
Based on the intraclass correlation coefficient calculation, 23.5% of the variance in safe opioid-prescribing score is accounted for by practice. Thus, approximately 76.5% remains to be accounted for by prescriber factors. Advanced practice providers (APPs, i.e., nurse practitioners, physician assistants, certified nurse midwives) were associated with a higher safe opioid-prescribing score of 2.92 percentage points. As percentage of patients aged 18 to 64 increases, the safe opioid-prescribing score also increases. Conversely, as the percent of white patients increases and as the number of patients prescribed opioids increases, the safe opioid-prescribing score decreases. At the facility level, large, multisite practices were associated with a higher safe opioid-prescribing score of 3.75 percentage points.
Discussion
Overall, the implementation of a CDS platform to operationalize the CDC CPG had a statistically significant influence on the percentage of encounters resulting in an opioid prescription. However, the influence was small (1.6%). The small effect is not surprising, as the intervention was aimed toward prescribing safely rather than preventing prescriptions. The significant trends before and after the intervention indicate a gradual reduction over time. This reduction is likely due to other factors, such as increased awareness of the harmful effects of opioids, increased oversight of opioid prescriptions by the health care system, changes in legislation, and/or changes in patient preferences. The trend is encouraging, showing that while 18% of patients received an opioid in January 2016, only 8% received opioids in December 2020. In comparison, literature shows approximately 27% of patients with osteoarthritis receive opioids—a rate higher than any time point in the present study.[11] Interestingly, opioid prescribing increased modestly at the beginning of the COVID-19 pandemic (April 2020). Given that this was at the beginning of COVID-19 lockdowns and clinic closures prior to the successful shift to virtual visits, the group of patients choosing to go to the clinic during this time likely represents a group of patients in more severe pain. Alternatively, prescribers may have prescribed opioids just in case patients needed them to avoid return visits to clinic in the context of the pandemic. However, rates quickly returned to normal and followed the decreasing trend over the rest of the study period. The fact that the prescribing rate was already low, leaving little room for drastic improvement, demonstrates the need for focus on other measures such as safe prescribing practices.
The average MME was approximately 50 to 60 throughout the study period, and the intervention was not associated with a change in the average MME of prescriptions. This finding is unsurprising considering the average MME was already below the CDC CPG recommendation (<90 MME). To determine the impact on high-dose prescriptions, future studies may assess the percent of opioids >90 MME over time.
Previous studies in the emergency department found significant effects of CDS on whether patients received opioids and MME. However, those studies did not control for existing trends or had small sample sizes.[40]
[41] Another study found no reduction in MME after implementation of a CDS intervention, but did find an increase in use of urine drug screens and naloxone prescriptions.[38] Synthesizing the literature is challenging because each CDS intervention is unique.[30] Therefore, research which rigorously evaluates and disseminates effective interventions should be prioritized by researchers, health care administrators, electronic medical record vendors, and funders.
After we developed this CDS intervention, the EHR vendor then provided to all their clients, providing the unique opportunity to replicate analyses across institutions. Pierce et al conducted an interrupted time series following implementation of these same clinical decision support tools included in our study, with some local customizations, including the 90-day opioid therapy, concurrent benzodiazepine, extended release, and naloxone alerts.[36] They found no change in number of opioid prescriptions or opioid dose associated with the launch of the intervention, which is similar to our findings. They assessed each guideline recommendation individually and found no change in extended-release opioids among opioid-naïve patients or in concurrent benzodiazepine prescriptions. However, they did find an increase in naloxone prescribing associated with the CDS intervention. We cannot compare these findings since we compiled all the behaviors into a single composite outcome.
The median safe opioid-prescribing score for prescribers was 77.1%. Practice site accounted for 23.5% of variation. Another study using a similar mixed-effects model found a substantial amount of variation in opioid-prescribing practices explained by practice.[42] Our hierarchical model identified two prescriber-level factors associated with safer prescribing: the percent of patients ages 18 to 64 and the prescriber being an APP. Previous literature has found APPs prescribe higher MME and frequency as compared with physicians.[43]
[44] However, those studies were not conducted in the context of a CDS intervention, and APPs may respond more to CDS tools. It is also possible that some practices delegate monitoring of chronic opioid therapy patients to APPs, and therefore, they are more experienced in treating these patients and following standardized protocols. However, we do not have information regarding how patients are assigned by practice across our large health care system.
We found two prescriber-level factors associated with less safe prescribing: the percent of white patient and the total number of patients to whom that clinician prescribed an opioid. Our findings align with historical patterns where white patients are more likely to receive opioids[45]
[46]
[47]
[48]
[49]
[50] and black patients are more likely to receive guideline-concordant care (i.e., urine drug testing).[51] Clinicians who prescribe to more patients may have a patient population that is more likely to require opioids, or they may be treating patients referred to them already established on opioids. Alternatively, since opioids are not recommended for these conditions, clinicians who prescribe to many patients may be less aware of current guidelines. Future research could explore these hypotheses. The effect of the number of patients may also be due to our formula for the safe opioid-prescribing score that, despite its weighting, might distort the performance of lower and higher-volume prescribers. Further research is needed to validate the measure and understand the relative importance of each behavior included in the composite score.
Interestingly, the random slope of number of patients by site also was statistically significant. This finding means the influence of number of patients on safe prescribing differed by practice site. Sites may have internal policies and practices which cause these differences. Some practices may designate a few clinicians to treat patients requiring opioids, while others distribute these patients among clinicians. Only one site-level covariate was statistically significant: practices that were large, defined as having multiple physical locations, were associated with higher safe opioid-prescribing scores. Large practices may be more likely to have formal policies and procedures to reduce variation in practice among clinicians. Additionally, these practices may have more resources to engage in surveillance and quality-improvement initiatives.
One limitation of this analysis is using diagnosis codes to identify patient encounters. These codes encompass chronic conditions that are likely to be included on any encounter the patient has with his or her primary care clinician, so we are unable to know for certain whether the opioid was specifically prescribed for the musculoskeletal condition. Some of these diagnoses may present with acute pain as well, so this methodology might capture some patients with acute pain. We also chose to remove ambulatory visits after an inpatient orthopaedic surgery for the remainder of the study period in an attempt to only capture chronic pain being managed nonoperatively. It is possible this exclusion biases the sample in the later years of the study toward less severe pain that did not require surgery. However, pain levels after surgery should improve, and we did not want to conflate postoperative recovery with pain prior to surgery. Additionally, the case mix variables are based only upon the patients receiving an opioid rather than the entire patient panel. Data for all patients are only available by attending clinician. However, prescriber is not always the same as the attending (e.g., APPs/residents). Because the outcome was directly related to clinician response to prescription alerts, we decided it was more important to describe the prescriber's behavior, rather than attribute these behaviors to the attending. In addition, the model could likely be improved by the addition of variables which were not available for this analysis (e.g., clinician demographics, years of experience, and whether the practice is associated with an academic center) or by including interaction. Finally, we were unable to test the impact of the intervention on safe opioid-prescribing behaviors because implementation of the CDS platform itself facilitated the measurement of these behaviors.
The interrupted time series design, correcting for autocorrelation, is a rigorous methodology for studying large-scale interventions. The trend for both of these outcomes over time was declining, indicating a shift in clinical practice likely associated with many simultaneous factors. Without controlling for the trend over time, we could have mistakenly attributed the change from the preintervention to postintervention period to the CDS intervention. Many of the statistically significant associations in the bivariate and multivariable regression models were attenuated or no longer significant in the multilevel model, highlighting the need for hierarchical modeling, particularly in the field of health care. This study used a relatively large sample of clinicians, and these results are likely generalizable to a broad range of primary care clinicians. Finally, we conceptualized and applied an objective measure of safe opioid prescribing, based upon the CDC CPG. An objective measure like this one would be useful both for research and clinical practice to monitor clinician performance and identify opportunities for improvement and intervention. Clinical dashboards have been developed creating comparative visualizations of opioid prescribing adjusting for case mix.[52] A similar tool could be developed based on the composite score developed for this study that would be widely applicable across subspecialties. These data and dashboards would be useful for both prescribers and teams focused on quality-improvement initiatives.
These results demonstrate that clinicians are not prescribing opioids for chronic musculoskeletal conditions frequently, and that, when they do prescribe opioids, they are generally adhering to guidelines. Therefore, future research might explore the trend in outlier behavior over time (e.g., prescriptions over 90 MME). This study also could be replicated with a broader range of chronic pain conditions outside of the musculoskeletal system. Finally, these data can be used to identify prescribers with very high and very low safe opioid-prescribing scores for purposive sampling in a subsequent qualitative study to identify targets for intervention. Recent qualitative research has identified clinicians' perceived utility of clinical decision support interventions and informs the design of CDS tools by identifying barriers and facilitators to adoption.[53]
[54] It is critical to pursue mixed-methods research to optimize the design and utility of CDS interventions while monitoring effectiveness to ensure CDS tools are useful and minimally burdensome on clinicians.
Conclusion
A CDS intervention was associated with a small improvement in percent of patients receiving an opioid, but not on average dose. Overall, the results of this study demonstrate clinicians are adhering to guidelines regarding opioid prescribing for patients with chronic musculoskeletal conditions. The implementation of a CDS tool presented the opportunity to objectively measure safe opioid-prescribing behavior. This measure could be used to assess the impact of future interventions in this health care system. Health care systems could also use this measure to tailor interventions for certain prescribers or practices where safe opioid-prescribing rates are low.
Clinical Relevance Statement
Clinical Relevance Statement
Musculoskeletal clinicians largely adhere to existing clinical practice guidelines regarding safe opioid prescribing. These results can be used to identify populations or groups of prescribers where additional training or intervention is needed to optimize adherence (e.g., small practices, MD prescribers). This outcome measure can be used to assess the impact of future interventions.
Multiple Choice Questions
Multiple Choice Questions
-
After implementation of a clinical decision support intervention to operationalize clinical practice guidelines, this health care system experienced:
-
A twofold reduction in the rate of opioid prescribing
-
Ability to comprehensively measure safe opioid prescribing
-
Reduction in average opioid dose
-
Increase in prescribers checking the prescription drug monitoring program prior to prescribing opioids
Correct Answer: The correct answer is option b. While this intervention did not lead to a clinically meaningful change in individual opioid-prescribing outcomes, it did create the ability to objectively assess overall compliance with clinical practice guidelines.
-
Which of the following was associated with lower safe opioid prescribing scores?
Correct Answer: The correct answer is option a. See [Tables 1]
[2]
[3]. APPs have statistically significant higher safe opioid-prescribing scores.
