In this quasi-experimental retrospective pre–post analysis, we examined the effects on prescribing patterns at our institution before and after modifications of opioid prescription settings. Thomas Jefferson University Hospitals, as an enterprise, has 908 acute care beds and over 2700 physicians and practitioners caring for more than 1.4 million people throughout the inpatient, outpatient, and ED settings [28]. This study was performed at the Center City Division, our urban academic institution comprising a large tertiary-care hospital, a community hospital, and multiple ambulatory clinics. Our study encompasses all outpatient prescriptions for opioids written between September 1, 2017, and August 31, 2019, in our EHR (Epic Systems Corporation) and the patients who received them. We did not include prescriptions for buprenorphine or methadone as these medications are used in medical assisted treatment (MAT) for opioid use disorder [29]. We limited our analysis to only oral capsule and tablet formulations (excluding oral liquid, buccal films, patches, etc.).

Our interest was in the effect of our intervention on opioids prescribed for acute pain. To this end, we excluded patients with chronic pain from our analysis via a modified approach to a validated algorithm [30]. Tian et al [30] achieved an accuracy of 95% for the identification of patients with chronic diseases via (1) a single International Classification of Diseases, 9th Revision (ICD-9) code [31] “highly likely” to represent chronic pain OR (2) 2 or more ICD-9 codes “likely” to represent chronic pain separated by at least 30 days OR (3) receipt of at least 90 days of opioids OR (4) an ICD-9 code “likely” to represent chronic pain AND 2 or more pain scores greater than or equal to 4.

We mapped ICD-9 codes utilized in the algorithm to the 10th Revision (ICD-10) [32] using the Centers for Medicare and Medicaid Services General Equivalence Mappings [33] via the R package “touch” [34]. ICD code mappings were also manually reviewed. There were 9476 prescriptions in our cohort that did not include an ICD-10 and remained in our analysis as if they did not meet criteria for exclusion. Our data set included unreliable pain score documentation, so we eliminated this step in the algorithm. Pain scores held the lowest positive predictive value for identifying patients with chronic pain in the initial study [30], and we presumed that excluding these patients would only bias our results against our objective, as patients with chronic pain are presumed to have long-term prescriptions. Finally, we excluded all prescriptions written in the ED given the results of our previous intervention.

Using our analytics software Qlik Sense (QlikTech International), we identified all prescriptions for non-MAT opioid medications written during the study period and extracted a number of variables for each individual prescription and the associated patient.

At the prescription level these variables were medication, number of tablets, dosage unit (ie, mg), route of administration (oral vs rectal), and frequency of administration. Using our previously described method, we also included the calculated duration of therapy [35]. We utilized this value to represent the duration of the prescription as our prior work demonstrated that what is documented in the prescription is at times unreliable [35]. We also extracted precalculated MME/day for each prescription. At the patient level we extracted demographic data including age and sex.

In order to mitigate bias, all information was obtained through the same data query, and all participants were selected in the same way (eg, a script for an opioid medication). Participants were retrospectively recruited from a continuous 24 months.

Participants were divided into 2 cohorts (12 months before and after the intervention). The intervention was introduced on August 24, 2018, so the month of August 2018 was included in the preintervention cohort to avoid influencing postintervention results.

We calculated and compared frequencies of missing values for metrics of interest before and after the intervention. Demographic information was compared to assess similarity between the 2 cohorts.

The total number of prescriptions and unique patients were evaluated before and after the intervention. To ensure our intervention did not inadvertently cause an increase in the frequency of opioid prescribing, we compared the median number of prescriptions per patient per month as well as the odds of a patient receiving more than 1 prescription or refills.

The proportion of prescriptions was calculated for each class of opioid medication. Opioid class was determined by the active ingredient in the compound. We chose not to separate by schedule class as these can change [36,37]. We compared the proportions of opioid classes before and after the intervention and then measured if our intervention changed the quantity, duration, or proportion of prescriptions greater than 90 MME/day for all prescriptions, as well as for each opioid class.

For statistical analysis, Wilcoxon rank-sum test was utilized for nonparametric data, and t test for parametric data. Chi-square and Fisher exact test were performed for comparison of categorical values. Confidence intervals were included for all appropriate analysis. Statistics were performed in R statistical software (R Core Team).

There were 128,933 non-MAT opioid prescriptions written during our study period. Of these, 103,338/128,933 (80.15%) were written for oral tablet or capsule formulations. We excluded 23,961/103,338 (23.19%) prescriptions that were written for patients with chronic pain. Of the remaining 79,377, we excluded 1131/79,377 prescriptions (1.42%) that were written by the ED. The median duration for these ED prescriptions was 2.7 days (IQR 2.5-3.3) with a median dispensed quantity of 9 tablets (IQR 9-31).

The remaining 78,246 prescriptions were for 30,975 unique patients. There were 38,976/78,246 (49.81%) prescriptions written for 16,464/30,975 (53.15%) patients in the preintervention period and 39,270/78,246 (50.19%) prescriptions written for 17,399/30,975 (56.17%) patients in the postintervention period. Figure 1 demonstrates the inclusion and exclusion criteria leading to our study cohorts.

Of the 78,246 prescriptions, 7167 (9.16%; 95% CI 8.96-9.36) were missing a calculated duration. The proportion of missing calculated durations decreased from 9.91% (7758/78,246; (95% CI 9.62-1.02) to 8.41% (6578/78,246; 95% CI 8.14-8.70; P<.001) before and after the intervention. In addition, 7156/78,246 (9.15%; 95% CI 8.94-9.35) prescriptions were missing documented quantity dispensed with a reduction from 9.90% (7750/78,246; 95% CI 9.60-10.20) to 8.40% (6570/78,246; 95% CI 8.13-8.68; P<.001) before versus after the intervention. Finally, 1139/78,246 (1.46%; 95% CI 1.37-1.54) prescriptions were missing the MME/day field with a significant increase in this proportion before versus after the intervention (915/78,246 [1.17%; 95% CI 1.07-1.28] to 1358/78,246 [1.74%; 95% CI 1.61-1.87]; P<.001). Missing values were excluded from their respective analysis.

Overall, 17,344/30,975 (55.99%; 95% CI 55.44%-56.55%) of all unique patients were female. There was no significant difference in patient sex before and after the intervention (17,300/30,975 [55.85%; 95% CI 55.10%-56.61%] vs 17,237/30,975 [56.65%; 95% CI 55.91%-57.39%] female; P=.14). The median age was 59, which did not change.

There was no significant change in the median number of prescriptions written per month, before and after the intervention (3254.5 [IQR 3176.25-3331.25] vs 3338 [3157.0-3391.5]; P=.59). The median monthly prescriptions per person remained 1 (IQR 1-2). The odds of getting more than 1 opioid prescription after the intervention did not significantly increase (1.003, 95% CI 0.967-1.04) and the odds of being prescribed a refill decreased (0.801, 95% CI 0.747-0.858).

The majority (48,261/78,246, 61.68%; 95% CI 61.34%-62.02%) of prescriptions during the study period were for oxycodone. Table 1 demonstrates the proportion of individual opioids.

There were small but significant reductions in the proportion of prescriptions for morphine—6.30% (2456/38,976; 95% CI 6.06%-6.55%) to 5.95% (2335/39,270; 95% CI 5.72%-6.19%); P=.04)—and oxymorphone—0.37% (143/38,976; 95% CI 0.31%-0.43%) to 0.24% (96/39,270; 95% CI 0.20%-0.30%; P=.002). There was no change in the proportion of the other opioids.

There was a significant reduction in the overall median quantity of opioid tablets dispensed before versus after the intervention (54 [IQR 40-120] vs 42 [IQR 18-90]; P<.001). There was also a reduction in median duration of treatment before and after the intervention (10.5 days [IQR 5.0-30] vs 7.5 days [IQR 3.0-30]; P<.001). Finally, there was a significant reduction in the proportion of prescriptions greater than 90 MME/day (27.46% [10,704/38,976; 95% CI 27.02%-27.91%] vs 22.86% [8979/39,270; 95% CI 22.45%-23.28%]; P<.001), despite no change in the median of 45 MME/day per prescription before and after the intervention. These results are displayed in Figure 2.

There was a significant reduction in all metrics for oxycodone. Tramadol and codeine demonstrated a reduction in tablets dispensed, while codeine also had a reduction in the duration of treatment (P<.001 for all three metrics). Hydromorphone had a significant reduction (P<.001) in MME greater than 90/day despite no change in the other metrics. Table 2 demonstrates the effect of our intervention overall, and on each type of opioid medication before and after the intervention.

Given the majority of prescriptions were for oxycodone, Figure 3 demonstrates monthly numbers of tablets dispensed, median duration of treatment, and proportion of prescriptions greater than 90 MME/day for oxycodone.

In this quasi-experimental pre–post study our intervention resulted in a significant reduction (P<.001) in the duration of treatment, quantity of tablets dispensed, and proportion of prescriptions greater than 90 MME/day for all opioid prescriptions written for patients with acute pain without an increase in the rate of prescribing or MME/day per prescription, while maintaining clinician autonomy. Our postintervention median duration of treatment for all opioids was 7.5 days, which is slightly longer than the CDC’s recommendations of 7 days [16], but an improvement over our preintervention duration of 10 days. In addition, we improved compliance with CDC recommendations that opioid dosing should not exceed 90 MME/day [16].

Oxycodone accounted for more than 60% (48,261/78,246, 61.68%) of our institutions prescriptions and was the only opioid to demonstrate reductions in all metrics, while tramadol, codeine, and hydromorphone had reductions in at least one metric. Our results show an improved median duration of treatment for oxycodone of 6 days, which is within CDC’s recommendations for acute pain.

Modifications to prescription presets is a relatively simple and effective way to combat the opioid epidemic. Our results are similar to other studies of single departments. Delgado et al [23] studied 2 EDs prescribing oxycodone. They examined the difference in prescribing patterns before and after the implementation of a new EHR, where they previously had no prescribing presets and the new EHR was preset to 10 tablets per prescription demonstrating a decrease in the median number of tablets from 11.3 and 12.6 to 10 and 10.9, respectively, in the 2 departments [23].

Chiu et al [22] looked at opioid prescribing in the outpatient surgical setting. Their study included 3 hospitals examining changes in preset opioid prescription quantities from 30 to 12 tablets. The intervention reduced the median number of tablets prescribed from 30 to 20, with a decrease of 5.22 tablets and 34.41 MME per prescription, and no difference in refill rates [22].

Despite their expectation of decreased tablets dispensed, Zwank et al [26] showed an increase from 15.31 to 15.77 in the mean number of tablets dispensed for hydrocodone and oxycodone in the ED after they removed a 15-tablet preset and required manual entry of a dispense quantity [26]. This may suggest that presets are vital to reduce prescribing quantities; however, this is contradicted by Santistevan et al [24] who demonstrated that removal of a default of 20 tablets for hydrocodone and oxycodone reduced the median number of tablets prescribed from 20 to 15, although their preintervention presets were higher than Zwank et al’s.

Montoy et al [25] implemented a block-randomization study at 2 EDs. They examined 6 possible opioid tablet quantity presets: “status quo” 12 and 20, as well as “null,” 5, 10, and 15. They demonstrated that each tablet increase in preset yielded an increase of 0.19 tablets prescribed, and lower default quantities were associated with lower number of pills in 8 of 15 pairwise comparisons [25]. These results support our approach that modifications to prescription presets are an effective way to enact change in opioid prescribing.

Despite a number of publications demonstrating reductions in opioid prescribing through presets [22,23,25], to our knowledge this research is the first to demonstrate reduction in overall opioid prescribing and prescribing of multiple individual opioids for an entire hospital system. Our hospitals and ambulatory clinics are located in one of the most lethal counties in one of the most lethal states associated with the opioid epidemic [38-40]. The CDC estimates that the risk of chronic use of opioids is 13.5% after 8 days of treatment [15] and the World Health Organization estimates the annual rate for opioid-dependent individuals overdosing at 45% and death at 0.65% [41]. We reduced our median duration of treatment from 10.5 to 7.5 (6 days for oxycodone), thus potentially bringing more than half of our institutions’ opioid prescriptions below this high-risk threshold.

Examining our hospital’s nearly 3300 prescriptions per month and postintervention median reduction of 12 tablets per prescription, we can estimate that our hospital system has reduced nearly 39,600 opioid tablets from being prescribed per month, or 475,200 tablets per year. Extrapolate this to the over 168 million prescriptions written in the United States in 2018 [42], and our simple intervention could potentially reduce the number of tablets being prescribed each year by over 2 billion. Our intervention excluded prescriptions for chronic pain, thus these approximations are likely extreme, but even a small percentage of these estimates is a net large and clinically important reduction in opioid tablets. Significant reductions in the number of opioid tablets prescribed, durations of treatment, and doses greater than 90 MME/day can play an important role in curbing the epidemic, improving quality of care, and ultimately saving lives.

Our findings have a number of limitations. First, although a validated algorithm, we did not complete Tian et al’s [30] algorithmic approach to identifying patients with chronic pain with EHR data. We believe this is unlikely to significantly bias our results as this was the least sensitive element in the original study and exclusion of this step, if anything, biases our results away from our objective (ie, including more patients with chronic pain with presumably higher doses and longer prescription durations in our data set). In addition, this algorithm was designed for use with ICD-9 codes. We mapped ICD-9 to ICD-10 codes and performed a manual review, but errors may still have persisted. Second, with regard to our results, we did demonstrate that morphine and oxymorphone had slight reductions in proportion prescribed after the intervention; however, this was unlikely to have biased our results. We also should acknowledge that 9.16% (7167/78,246) of our study cohort were missing calculated durations and quantity of tablets dispensed and the percentage of these missing values decreased after the intervention, while 1.46% (1139/78,246) of prescriptions were missing MME/day and this increased after the intervention. We believe the majority of missing calculated durations are likely due to the missing dispense quantities, which are required for the computation of this metric. Missing MME/day is likely due to missing concentrations, doses, or conversion factors of some formulations. Some of these missing values may be due to prescriptions being refilled and bypassing some of the new requirements we implemented. We excluded missing data from our analysis, but it is possible this exclusion influenced our results.

For our MME/day threshold we chose to use the CDC guidelines for opioid prescribing as the CDC is a federal agency that upholds “the health of the people of the United States.” [43] It should be noted that individual states may have opioid prescribing guidelines that deviate from those presented by the CDC [44].

It should also be noted that our research was a quasi-experimental pre–post study, and we cannot infer causality with our intervention and the results observed. It is possible that other influential factors could have played a role in the reduction of opioid prescribing at out institution during this 2-year period, as the opioid epidemic has come to the forefront of modern medicine. However, we believe our results are encouraging, and other studies have demonstrated similar outcomes to suggest our intervention likely played at least a role in the reduction of prescribing patterns at our institution. Finally, our region has one of the highest opioid overdose and death rates in the nation, making our results of high value to curbing the epidemic; however, our results may not be generalizable to the rest of the country.

In this quasi-experimental retrospective pre–post analysis, we demonstrated that modifications of the opioid prescribing presets in our hospital system’s EHR can improve prescribing practice patterns and reduce the number of pills dispensed, duration of treatment, and proportion of prescriptions greater than 90 MME/day. We have identified a simple and effective way to reduce opioid prescribing, and to our knowledge, we are the first to perform this kind of an intervention throughout an entire hospital system, and not just at the department level. Reduction in opioid prescribing may aid in curbing the opioid epidemic, thus improving quality of care and potentially saving lives.