Effect of Integrating Access to a Prescription Drug Monitoring Program Within the Electronic Health Record on the Frequency of Queries by Primary Care Clinicians

Key Points Question Does direct access to the prescription drug monitoring program (PDMP) from within the electronic health record (EHR) increase the frequency of PDMP queries by primary care clinicians? Finding This cluster randomized clinical trial involving 309 clinicians in 43 primary care clinics found that providing direct access to a PDMP tool from the EHR increased PDMP queries by 60% compared with clinicians in control clinics who did not have EHR-integrated access. Meaning The findings of this trial demonstrate that direct access from the EHR to a PDMP can increase the provision of guideline-concordant care.


54
Chronic pain is highly prevalent, affecting an estimated 11% of the U.S. adult population. 1 The CDC's chronic pain guidelines call for prescribers to check their state-administered 76 prescription drug monitoring program (PDMP) to identify potentially problematic opioid use 77 patterns. Some studies examine the effects of mandating clinicians to check the PDMP. 15 Intervention strategies can also make it less difficult for clinicians to access or utilize information 106 for appropriate pain treatment decisions. Unnecessary points of friction can undermine the 107 likelihood of pursuing or completing a task, even one that is construed favorably and people 108 indicate they will complete. 24 For instance, clinicians may affirm the value of information about 109 a patient's prescribing history, yet not use that information because it is difficult to access. Thus, 110 strategies that remove impediments to this information may facilitate clinicians delivering care 111 that is guideline-supported and concordant with their intentions.  The PRINCE study investigates two interventions that were informed by behavioral economics 123 principles. These interventions were chosen based on input from PCPs, pain treatment, and 124 opioid use stakeholders in the Fairview system.

126
The first intervention (the "choice architecture" intervention) is two alerts within Epic that the 127 study team designed based on choice architecture principles from behavioral economics. One The delivery system will implement the interventions in the same way that EHR enhancements 168 and changes are implemented normally. "Tip sheets" about the interventions and how to use 169 them, will be shared with all PCPs in the appropriate intervention arm clinics by their clinic 170 leadership (included in the Supplemental Materials).

171
Overview of study design and randomization 172 A clinic-randomized factorial design will be used to test the effects of the two interventions on 173 pain treatment and opioid prescribing decisions by PCPs. The design contains up to two stages of 174 randomization ( Figure 5). In the first stage, 43 primary care clinics will be randomized 1:1:1:1 to 175 receive the "choice architecture" intervention alone, the PDMP integration intervention, both 176 interventions, or care as usual. All PCPs (physicians, physicians assistants, and nurse 177 practitioners) working in the study clinics were exposed to the intervention(s) that their clinic is 178 assigned to. To ensure similar characteristics of the clinics across randomized groups, covariate 179 constrained randomization procedures will be used. 25,26 Specifically, randomization is The intervention period will last for 12 months. An interim outcomes analysis is planned to 187 inform a possible second stage of randomization after six months, when the study team will 188 consider re-randomizing the clinics receiving any active intervention to reduced intensity (i.e.,  The primary aim of the study is to test whether two separate interventions affected PCP decisions 209 around pain treatment and opioid prescription. We hypothesize that each of the interventions will 210 increase the likelihood that pain treatment and opioid prescribing decisions will be consistent 211 with clinical guidelines. Specifically, we hypothesize that each of the interventions will reduce 212 initiation of opioid-prescriptions without a concurrent or previous non-opioid recommended 213 treatment among patients who are opioid naïve. Similarly, we hypothesize that each intervention 214 will lead to increased use of appropriate opioid tapering among patients currently prescribed a 215 "high risk" opioid. We also will examine several secondary outcomes representing more-specific 216 features of opioid prescribing and pain treatment. interventions firing less-frequently affects outcomes differently than when the intervention fires 224 whenever triggered.

226
As an exploratory aim, we will assess potential moderators of the intervention as well as assess 227 whether there is an interaction between the two interventions being studied. We hypothesize that 228 the two interventions will interact to have a stronger effect on the study outcomes. PCPs and patients for being exposed to the study interventions and for the study team to Because the choice architecture arm of the study uses separate alerts for opioid-naïve and 245 current, "high-risk" opioid-using patients, there are separate primary outcomes for visits with 246 each of those populations. The data for those outcomes are at the encounter-level and for all 247 cohorts includes patient visits during the 12 months before and after the start of the study 248 interventions for patients 18 and older who did not have a current cancer diagnosis.

249
For the opioid-naïve population, the analysis cohort includes patient visits meeting the above  For the current "high risk" opioid population, the analysis cohort included primary care visits for 265 patients with a current opioid prescription with an MME of 50 or higher, or currently prescribed 266 an opioid and a benzodiazepine. The primary outcome for this population is a variable with three 267 15 mutually-exclusive categories. Category 1 (appropriate taper): Whether a patient visit had an 268 order that would reduce MME by no greater than 20%, relative to the current prescription, and 269 there is documented evidence that the reduction is consistent with CDC guidelines. Category 2 270 ("inappropriate" taper): Whether a patient visit had an order that would reduce MME without 271 documented evidence that the reduction was consistent with CDC guidelines, or, decreased 272 MME by greater amounts than recommended (>20% relative reduction in MME). Category 3 (no 273 taper): Whether patient visit had no reduction in MME. Data for this outcome are derived from 274 EHR data. For encounters in which the MME was reduced by no more than 20%, members of 275 the study team blinded to randomization will audit the chart to assess whether the taper was 276 consistent with CDC guidelines (the chart audit tool is included in the supplemental materials).

277
Two secondary outcomes specific to this population include 1) Whether there is a partial 278 reduction in the MME or prescription length of refill orders, versus a total opioid 279 discontinuation. 2) Whether there is an increase in the MME/day. PCP's behalf, and checks that were done using the integrated PDMP tool within the EHR. Data 286 will be available for the 12 months prior to the study and the 12-month study period. To test our primary hypotheses, we will fit mixed effects logistic regression models using data 300 from the 12-month pre-intervention period, and the post-intervention period prior to any second-301 stage randomization (the model for the current "high risk" opioid group will be multinomial 302 logistic regression due to the 3-category outcome). All models will include fixed effects for 303 whether the encounter was at an intervention arm clinic, an indicator for whether the , and patient-level (age, sex, race/ethnicity, insurance status, and in-person or virtual 312 visit). Separate models will be fit to assess the PDMP integration and choice architecture 313 intervention and for the opioid-naïve and current, high-risk opioid users. In each model, the 314 intervention group pools across two arms of the factorial design (e.g., the choice architecture 315 intervention pools across arms 2 and 4) and the control condition pools across the remaining two 316 groups (e.g., arms 1 and 3 to continue the example). A Bonferroni correction will adjust 317 inference for multiple comparison across two different populations (opioid naïve and current, 318 high-risk opioid users) but we will not adjust for multiple tests across the two different 319 interventions.

320
As an exploratory analysis, we will assess whether PCP characteristics (clinician type and tenure 321 in medicine) and patient characteristics (age, sex, and race/ethnicity) moderate the interventions' 322 effect. Tests for treatment effect heterogeneity will be implemented by adding the interaction 323 between the potential moderator and the intervention indicator to the models for the primary 324 outcome (adjusting for the same factors described above). The effect of the intervention within 325 subgroups formed from the potential moderators will be estimated by fitting separate models 326 within each subgroup.

327
Subgroup analyses will not be adjusted for multiple comparisons; they are supportive to the 328 primary outcome analysis. Subgroup analyses will be interpreted with caution due to limited 329 power and uncontrolled type I error.

330
If the study proceeds to the second stage of randomization, we will fit mixed effects logistic 331 regression models using data from 12 months pre-and post-intervention. All models will include 332 fixed effects for intervention arm (control, continued intervention, 50% intervention, or 333 intervention turned-off), indicators for whether the measurement was after the initial 334 18 interventions began or after the second stage randomization, and their interaction. The interaction 335 is the primary measure of the intervention effect. The models will include the same random 336 effects and covariate adjustment as above.

337
Secondary Outcomes Analysis: PDMP and Web Survey data 338 For other EHR-derived secondary outcomes, we will use the same general modeling framework 339 as for the primary outcome but will fit mixed effects logistic or linear models depending on 340 whether the outcome was categorical or continuous.

341
To assess intervention effects on the frequency of checking the PDMP we will fit mixed effects 342 Poisson regression models using data from the 12 months pre-and post-intervention. All models 343 will include fixed effects for whether the PCP was at an intervention arm clinic, indicator for 344 whether the month was during the intervention period, and their interaction. The models will 345 adjust for clinician type, length of tenure in medicine, and health system indicators, and include 346 random effects for clinic and PCP (nested in clinic) to account for within-clinic and within-347 provider correlation. As an exploratory analysis, we will assess whether PCP characteristics 348 (clinician type and length of tenure) moderate the effect of the interventions using a similar 349 process as the primary outcome.

350
A similar approach will be used for the web survey data, except we will fit mixed effect linear 351 regression models using data from the pre-and post-intervention surveys. All models will 352 include fixed effects for whether the PCP was at an intervention arm clinic, an indicator for 353 whether the survey was the 12-month follow-up, and their interaction. The University of Minnesota IRB and the study sponsor agreed that a Data Safety and 356 Monitoring Board was unnecessary. However, a Data Monitoring Committee (DMC) will be 357 formed and asked to recommend graduation to the second stage of randomization if there is clear 358 and substantial evidence of treatment efficacy. As a guideline, the Lan-DeMets spending 359 function analog of the Pocock boundaries will be used to monitor the primary outcome 360 comparison for the primary aims. 27 The study team, with the exception of the unblinded 361 statistician, will be blinded to interim outcomes. We plan to conduct a single interim analysis 362 occurring midway through the study.

363
The interim analysis of the effect of the interventions on key secondary outcomes will be 364 provided to the DMC. Should the stopping boundary be crossed with clear evidence of an 365 intervention's effect, we will ask the DMC to recommend proceeding to the secondary 366 randomization of intervention arm. Without clear evidence of an intervention's effect, we will 367 continue the study for an additional 6 months without re-randomizing any clinics (to ensure 368 sufficient power to detect an effect of the intervention). The decision to move to the second stage 369 of randomization will be made separately for both intervention arms. If the PDMP integration 370 intervention shows substantial efficacy at the interim analysis, we will recommend that the 371 PDMP integration be made available to all clinics. A decision on what constitutes substantial 372 efficacy will be made in consultation with the DMC. As the study interventions have small 373 potential adverse effects, the DMC would not be asked to recommend early termination for 374 futility. The PRINCE study will use a clinic-randomized factorial design with potential for secondary 377 randomization to test the effects of two separate behavioral economics-informed interventions on 378 20 a range of outcomes related to guideline-concordant practices of opioid prescribing and pain 379 treatment. One limitation of the study is that the primary outcome for the current-opioid using 380 population relies on audits of medical charts to classify opioid tapers as appropriate or 381 inappropriate. This classification may miscode some tapers as inappropriate to the extent that the 382 clinical rationale was not entered into the patient's chart.

383
The study contains several novel aspects. One intervention arm includes two novel interventions 384 to the choice architecture around opioid prescribing, and another intervention arm represents an 385 already-diffusing EHR tool to facilitate PDMP use that has not yet been rigorously evaluated. An 386 adaptive design will test whether the interventions can be "titrated" without compromising 387 efficacy. We considered a fixed design where all active intervention clinics would be re-