The MedSafer Study—Electronic Decision Support for Deprescribing in Hospitalized Older Adults

Key Points Question Does providing clinical decision support during an acute care hospitalization improve deprescribing of potentially inappropriate medications and 30-day postdischarge adverse drug events (ADEs) in older adults? Findings This cluster randomized multicenter trial of 5698 hospitalized participants found that providing electronically generated deprescribing reports did not have a significant impact on ADEs within 30 days despite increased deprescribing at discharge. Meaning The findings of this randomized clinical trial indicate that clinical decision support during hospitalization improves deprescribing but has little impact on medication harms in the short term.

D eprescribing is the medically supervised tapering or cessation of medications that are no longer needed or beneficial, including medications that are potentially inappropriate (PIMs), time delimited, ineffective, or that do not align with goals of care. 1 After an acute care hospitalization, adverse drug events (ADEs) are common and up to 60% are considered preventable. 2 Deprescribing PIMs at hospital discharge to prevent ADEs could be advantageous; however, several barriers exist. 3,4 First, hospital practitioners focused on acute care requirements may be reluctant to alter long-term outpatient therapies managed by other prescribers, particularly if they do not follow patients postdischarge. 5 Second, not all centers include routine pharmacist or geriatrician subspecialties to support interprofessional care. Third, some prescribers fear that stopping or reducing a drug can lead to adverse effects. 6 Fourth, despite deprescribing opportunities, therapeutic inertia is common 7,8 ; true even for low-risk drugs with no withdrawal concerns 9 and more so for drugs with higher risks of harm. 10 Even when adverse events occur, opportunities to deprescribe are not always recognized or acted on. 11,12 To our knowledge, prior studies of interventions to reduce PIMs while in hospital [13][14][15] have tended to be underpowered and nonrandomized or did not capture postdischarge ADEs.
We previously developed and tested an electronic decision support tool, 16 MedSafer, in a pilot controlled before and after study of deprescribing in the acute care setting. Med-Safer integrates data available in the electronic health record (EHR) 17 to generate reports with evidence-based deprescribing opportunities, stratified according to prespecified expert consensus of perceived drug risks. In our pilot study, providing MedSafer reports to practitioners increased the absolute rate of PIM deprescribing by 8.3%. 16 The primary objective of the present study was to perform a large stepped-wedge cluster randomized trial to evaluate the impact of MedSafer deprescribing reports on 30-day postdischarge ADEs. The secondary objectives were to assess the impact of MedSafer reports on deprescribing and on safety outcomes related to adverse drug withdrawal events (ADWEs).

Methods
We evaluated the use of MedSafer using a cluster randomized trial at 11 participating acute care hospitals in Canada. All hospital sites obtained study approvals from their local research ethics board and administration. Potentially eligible patients were approached for their consent to the 30-day postdischarge telephone interview; a family member or proxy granted consent for patients lacking capacity, in accordance with the ethics regulations of Canada. 18 The trial protocol and statistical analysis plan are available in Supplement 1. The study followed Consolidated Standards of Reporting Trials (CONSORT) reporting guidelines extension for cluster randomized trials. 19 Design, Study Population, Setting, and Randomization The study was conducted with patients admitted to a participating site from August 22, 2017, to January 13, 2020. Each participating site had a control phase followed by an intervention phase. The timing of the intervention varied by cluster, and the order of entry into the intervention was determined centrally by a randomized sequence generated by statistical software. Study sites were kept blinded until approximately 4 weeks prior to their allocation to the intervention to allow for preparation. Clusters were scheduled to move from 1 period to the next every approximately 500 recruited patients. One cluster (Western Canada) had 2 periods when they were able to recruit only 250 participants of the planned 500 because patients had been transferred to nonstudy units.
Patients aged 65 years and older who were regularly taking 5 or more usual medications prior to admission were eligible. Patients receiving palliative care with an expected prognosis of more than 3 months were included. Readmitted patients were eligible only if they had not been previously enrolled in the study. Additional study design details, including the composition of the clusters, are available in the eMethods, eFigure 1, and eTables 1 and 2 of Supplement 2.

Description of the Intervention
During the control period, patients received usual care (having a best-possible medication history performed), with medication reconciliation at hospital admission and discharge. 20 Any deprescribing that took place was based on local and individual practice by the unit physicians and pharmacists. A trained research assistant extracted medical conditions and specific laboratory values from the comprehensive admission notes and entered these data, along with the prehospital medications list, into MedSafer. Any changes to home medications at discharge were captured from exit prescriptions and/or medication reconciliation documents.
The main intervention was the provision of individualized deprescribing reports based on evidence-based guidelines for safer prescribing in older adults, 21-23 with tapering instructions when indicated. The report prioritized opportunities for deprescribing based on prespecified expert consensus as either: (1) high-risk PIMs (harms outweigh benefits for most); (2) intermediate-risk PIMs (harms may approximate benefits, clinical judgment required); and (3) PIMs of little added value (medication shown to be ineffective or adds to pill burden). Opportunities for deprescribing were generated by crossreferencing medical conditions, laboratory values, and home

Key Points
Question Does providing clinical decision support during an acute care hospitalization improve deprescribing of potentially inappropriate medications and 30-day postdischarge adverse drug events (ADEs) in older adults?
Findings This cluster randomized multicenter trial of 5698 hospitalized participants found that providing electronically generated deprescribing reports did not have a significant impact on ADEs within 30 days despite increased deprescribing at discharge.

Meaning
The findings of this randomized clinical trial indicate that clinical decision support during hospitalization improves deprescribing but has little impact on medication harms in the short term. medication lists with evidence-based guidelines for safer prescribing in older adults, described previously. 16,24 Deprescribing reports were provided to the treating team within 3 business days of patient admission and were designed to engage the physician and the hospital pharmacy team. At discharge, the report was sent by facsimile to the patient's community pharmacy and the self-identified usual treating physician(s). Patients or their caregivers received an educational pamphlet on deprescribing, as well as relevant patient-oriented deprescribing pamphlets from the Canadian Deprescribing Network for select classes of medications, such as sedative-hypnotics. 2

Outcomes
The primary outcome was the proportion of patients experiencing an ADE within the first 30 days after hospital discharge in the intervention vs the control phases, including ADWEs, which are a subset of ADEs (eFigure 2 in Supplement 2). Patients or proxies consented to a structured telephone interview approximately 30 days after hospital discharge, conducted by an experienced interviewer and blinded to the intervention status. 28 The interview consisted of a modified Australian Adverse Reaction and Drug Event report 29 and questions about medication changes, new or worsening symptoms, and planned or unplanned visits to medical professionals. Details of the interview were previously described 16 and are available in the eMethods of Supplement 2. If a patient was readmitted to the hospital, a detailed review of the EHR was performed and the reason(s) for readmission was identified.
The adjudication of ADEs was performed by trained study investigators blinded to patient identification, hospital site, province, and the intervention status. In brief, reviewers were first asked to determine if an adverse event had occurred (eg, a fall, emergency department visit, hospitalization, unplanned health care visit) and then, using a Leape and Bates Likert scale, 30 to rate the likelihood of it having been an ADE. An ADE was defined as an event with a rating of 5 (probably caused by medication) or 6 (definitely caused by medication). 30,31 Adjudication was conducted independently in duplicate, and disagreements were resolved by an independent third reviewer. Details of training and examples are available in the eMethods of Supplement 2 and also have been described elsewhere. 31 The Gwet agreement coefficient (Gwet AC) 32 was used to estimate interobserver agreement between the first 2 reviewers for adverse events and ADEs, respectively. Safety outcomes included ADWEs, whereby the reviewer was asked if the ADE was owing to temporarily withholding, tapering, or stopping (deliberately or by omission) of a medication. A sensitivity analysis was performed that included any ADE that was rated as a 4 (possibly related to medication) or higher on the Leape and Bates Likert scale.
Secondary outcomes included the proportion of patients with 1 or more PIMs deprescribed at hospital discharge compared with usual care. A medication was considered deprescribed if it was discontinued or if a taper to discontinue was prescribed. For medications whose dose was reduced, a blinded review was used to determine if that reduction was a deliberate attempt at deprescribing. We also evaluated 30-day postdischarge adverse effects, falls, emergency department visits and/or hospitalizations, deaths, quality of life (using the EQ-5D-5L questionnaire with a Canadian value set 33 ), and sleep disturbance (using the PROMIS SD 4a). Harms were considered ADWEs, as were the impacts on sleep and quality of life. Prespecified subgroup analyses examined the effect by sex, palliative designation, frailty, and residence in a long-term care facility.

Sample Size
Based on the literature, the proportion of medical patients who could experience an ADE after discharge was estimated to be 15.0%. 34,35 We used an α of .05, 80% power, and an expected cluster correlation coefficient of 0.03. 36 We prespecified an absolute decrease of approximately 4.0% corresponding to a number needed to treat of 25. Using the Stata stepped-wedge function 37 and following the Hussey and Hughes approach, 38 we planned for a 6-cluster trial with each cluster recruiting 200 patients per cluster-period for a total of 8400 patients, resulting in 80% power to detect a change in ADE from 15.0% to 11.2%.
Prior to the recruitment of 200 participants in any cluster, recruitment rates suggested it would be impossible to complete the trial as originally designed. Without examining outcomes data, we realigned the 6 clusters into 3 clusters according to recruitment rates and geography (Ontario, Quebec, and Western Canada), permitting the study to be completed within the funding duration and budget. A 3-cluster trial with 500 patients per cluster per period allowed for 80% power to detect a 4.8% absolute difference with 6000 patients.

Statistical Analysis
Baseline characteristics were expressed as numbers and percentages for categorical variables and median (IQR) for continuous variables. Differences between control and intervention baseline characteristics were compared by χ 2 or ranksum as appropriate.
For the primary outcome, ADEs, a mixed-effects logistic regression model was used controlling for intervention status, time period, and the number of baseline PIMs as fixed effects, and cluster as a random effect. Adjusted risk differences (aRD) were estimated from the model parameters differences. 39 This analysis was conducted with data from patients who consented and participated in the postdischarge follow-up interview.
For secondary outcomes, an identical analysis was conducted for any postdischarge adverse event. The intervention's effectiveness in terms of stopping 1 or more PIM was also evaluated using the same method, restricted to those participants taking 1 or more PIM at admission and discharged alive. Further details of the analysis are described in the eMethods of Supplement 2.

Sensitivity Analyses
To address the low number of clusters in the study, we performed 5 post hoc sensitivity analyses: (1) an analysis adjusting for patient factors that differed between groups by P < .01; (2) treating clusters as a fixed effect 40 ; an analysis with (3) a random and (4) fixed effect for hospital (n = 11, in lieu of cluster), and (5) using permutation, a nonparametric method of evaluating for the weighted within period effect size. 41 Figure). Deprescribing opportunities were identified in 4923 patients (86.4%).

Baseline Data
The control and intervention groups were similar regarding most major medical comorbidities; however, the intervention group was slightly older and more likely to have a diagnosis of a major neurocognitive disorder. The median (IQR) number of home medications was 10 (8-14); median (IQR) number of PIMs was 2 (1-3); and the median (IQR) hospital patient 1-year mortality risk (HOMR) score was 39 (36)(37)(38)(39)(40)(41) or approximately a 30% risk of death within 1 year. 17 The most frequently identified PIMs were proton pump inhibitors, sedative hypnotics, gabapentinoids, and diabetic agents with low (<7.5%) glycated hemoglobin A 1c levels.

Harms
There were 49 ADWEs representing 19.7% of 249 ADEs (aRD, −0.1%; 95% CI, −1.2% to 1.0%). The most common ADWE was fluid overload secondary to dose adjustment of diuretics, with 1 event owed to a MedSafer recommendation-hydrochlorothiazide deprescription triggered by severe hyponatremia. Sleep and quality of life remained stable before and after hospitalization (eTables 5 to 10 in Supplement 2).

Sensitivity Analyses
All 5 sensitivity analyses provided similar estimates of effect size and confidence intervals. None of the sensitivity analyses altered the conclusions of the trial. These results are presented in eTable 11 of Supplement 2.

Discussion
Despite clinically and statistically significant increases in deprescribing, communication of medication changes to community physicians and pharmacists, and sustained deprescribing 30 days postdischarge, this large randomized clinical trial was unable to demonstrate any significant impact of deprescribing decision support on short-term ADEs. Several explanations may address the discrepancy between the impact on deprescribing and the complete lack of effect on ADEs. The overall incidence of ADEs (defined as a Leape and Bates Likert score of 5 or 6) was only 5%, which is lower than the approximately 10% to 15% incidence observed in the seminal studies 2,34,35 used to inform our power calculations. However, those studies were conducted more than 15 years ago. The complexities of patient and medication regimens have increased substantially and identification of probable and/or definite ADEs has become more challenging. 31 In addition, wide-spread hospital pharmacist involvement in medication reconciliation has created opportunities to mitigate more worrisome prescribing practices, such as errors of omission, which may lead to clearly identifiable ADEs. While our intervention identified numerous deprescribing opportunities, many were for low-risk nonbeneficial polypharmacy (eg, nonstatin cholesterol-lowering medications or stool softeners 43 ). Deprescribing these medications is less likely to impact 30-day ADEs, but still has patient and societal value, eg, avoiding excess cost, waste, and pill burden. 9,43 When powering future studies of ADEs, interventions may need to focus specifically on high-risk medications, and the time frame for observing the outcome likely needs to be extended. 44 To our knowledge, few large randomized clinical trials of deprescribing in acute care have been completed. The 1537 patients studied by SENATOR 15 (Software Engine for the Assessment and Optimization of Drug and Nondrug Therapy in Older Persons Trial) found poor uptake (<20%) of computerized pharmacologic recommendations in their hospital setting. An accompanying qualitative study 5 attributed this to multiple concurrent work commitments and a reluctance to take sole responsibility for an older person's pharmacotherapy. Most other deprescribing studies have been small, potentially leading to questions of whether the absent impact on ADEs was related to study size. From our study findings, it appears that the timing of outcome ascertainment and the approach to quantifying medication harms are equally important, and both may need to be adapted for future studies. 31

Limitations and Strengths
This study had limitations; first among them was our low number of clusters (n = 3). When we reduced the number of clusters in the study, we introduced a risk that randomization might not achieve balance between groups across entry times, and indeed  this was the case. For instance, the intervention group was slightly older, and participants had been prescribed more baseline PIMs.
To attempt to address any bias or risk to validity owing to these imbalances, we performed several sensitivity analyses. Although these supported the conclusion of the study, the results still need to be interpreted with this caveat in mind. Second, the rate of ADEs was significantly lower than expected, leaving the study underpowered for the primary outcome despite being the largest deprescribing trial performed to date to our knowledge. A recent study evaluated a multifaceted pharmacist-led intervention on medication safety posthospital discharge. 45 The primary outcome included potential ADEs, in addition to preventable and ameliorable ADEs; however, it failed to demonstrate an impact on ADEs even with this broader definition. While our sensitivity analysis that included possible ADEs appeared to show a stronger effect of the intervention, it still did not demonstrate a reduction in ADEs.
Third, the generation of deprescribing reports required manual data input, which could limit scalability. The software is now updated to process EHR data (eg, codes from the International Statistical Classification of Diseases and Related   Health Problems, Tenth Revision, and Drug Identification Numbers) directly via an application programming interface. This capability has important implications for implementation by reducing the burden of data input. Fourth, reports were provided for home medications at hospital admission and not regenerated at discharge. Thus, the intervention did not formally address new in-hospital PIM starts, 44 which may be particularly relevant to preventing early postdischarge ADEs.
Importantly, the intervention did not lead to an increase in ADWEs. The findings demonstrated a clinically and statistically significant impact on deprescribing of PIMs and generated extensive safety data in support of acute care deprescribing. Other strengths of the study included an intervention that addressed complex medication regimens and is generalizable to academic and community hospitals. We involved inhospital pharmacists in the design of the intervention, which may have increased the proportion of deprescribing opportunities that were accepted. We also engaged patients and their caregivers with deprescribing brochures, and communicated medication changes to the treating community team, which promoted shared decision-making and reinforced longerterm persistence of changes made during the hospitalization.

Conclusions
This randomized clinical trial found that providing deprescribing decision support to the acute care medical teams did not impact 30-day ADEs; however, this intervention effectively stopped PIMs, with no evidence of increased harm. Shortterm ADEs may not be the ideal outcome to measure to capture the benefits of deprescribing during an acute care hospitalization.