Digital Medication Management in Polypharmacy

Abstract

Background

Inappropriate drug prescriptions for patients with polypharmacy can have avoidable adverse consequences. We studied the effects of a clinical decision-support system (CDSS) for medication management on hospitalizations and mortality.

Methods

This stepped-wedge, cluster-randomized, controlled trial involved an open cohort of adult patients with polypharmacy in primary care practices (=clusters) in Westphalia–Lippe, Germany. During the period of the intervention, their medication lists were checked annually using the CDSS. The CDSS warns against inappropriate prescriptions on the basis of patient-related health insurance data. The combined primary endpoint consisted of overall mortality and hospitalization for any reason. The secondary endpoints were mortality, hospitalizations, and high-risk prescription. We analyzed the quarterly health insurance data of the intention-to-treat population with a mixed logistic model taking account of clustering and repeated measurements. Sensitivity analyses addressed effects of the COVID-19 pandemic and other effects.

Results

688 primary care practices were randomized, and data were obtained on 42 700 patients over 391 994 quarter years. No significant reduction was found in either the primary endpoint (odds ratio [OR] 1.00; 95% confidence interval [0.95; 1.04]; p = 0.8716) or the secondary endpoints (hospitalizations: OR 1.00 [0.95; 1.05]; mortality: OR 1.04 [0.92; 1.17]; high-risk prescription: OR 0.98 [0.92; 1.04]).

Conclusion

The planned analyses did not reveal any significant effect of the intervention. Pandemic-adjusted analyses yielded evidence that the mortality of adult patients with polypharmacy might potentially be lowered by the CDSS. Controlled trials with appropriate follow-up are needed to prove that a CDSS has significant effects on mortality in patients with polypharmacy.

Medication errors are responsible for around 30 to 50 percent of all medical errors (1–3). Approximately seven percent of hospital admissions are for adverse drug reactions (ADRs), of which two percent are fatal and 30 to 70 percent avoidable (4–6). Polypharmacy, usually defined as the simultaneous use of five or more medicines (7), affects more than one third of all adults (8) and is associated with adverse health outcomes, poor medication safety, and poor utilization of the healthcare system, which is reflected in higher rates of hospital (re)admissions, emergency admissions, and care home placements (1).

Clinical decision support systems (CDSSs) can improve prescribing quality and patient safety by preventing medication errors, ADRs, and the prescription of potentially inappropriate medications (PIMs) (9–11). They can also support primary care physicians in carrying out risk-benefit assessments of drug therapy (12, 13).

Although CDSSs improve process parameters and increase medication safety (14), patient-relevant outcomes are inconsistent: So far, no randomized study has been able to demonstrate effects of CDSSs on mortality or hospital admissions, often due to a lack of statistical reliability (14).

The aim of the AdAM study (Application of an Electronic Medication Management Support System) was therefore to examine in a cluster-randomized controlled trial whether the application of a user-initiated CDSS by primary care physicians can reduce hospital admissions and mortality in adults with polypharmacy.

Methods

The AdAM project primarily investigated whether a CDSS-supported medication review in primary care practices in Westphalia-Lippe, Germany, leads to a reduction in all-cause hospital admissions and/or all-cause mortality in adults with polypharmacy (at least five medicines). As secondary outcome measures, an assessment was also conducted as to whether the intervention

• reduces all-cause mortality,

• lowers all-cause hospital admissions, and

• improves prescribing quality and medication safety.

A detailed description of the methodology can be found in the eMethods and in the numerous Tables in the eSupplement. The study was registered with ClinicalTrials.gov (NCT03430336), funded by the Innovation Fund (01NVF16006) of the German Federal Joint Committee, and approved by the Ethics Committee of the North Rhine Medical Association (Nr. 2017184, 24.07.2017).

eMethods.

Study design

In order to minimize contamination effects, a cluster design was chosen for the AdAM study (Application of an Electronic Medication Management Support System) in which primary care practices served as randomization units. Although the study was originally planned as a cluster-randomized controlled trial with a parallel group design (parallel group c-RCT), a stepped-wedge design (SWD) was ultimately adopted due to the insufficient recruitment of practices and patients. An SWD with an open cohort and step length of one quarter (= 3 months) was chosen. The clusters changed from the control group to the intervention group in one of eleven steps. Details can be found in the study protocol (e1). The study was conducted from June 27, 2017, thru March 31, 2021, in primary care practices in the region of Westphalia-Lippe, Germany.

The study was approved by the Ethics Committee of the North Rhine Medical Association (No. 2017184 of July 24, 2017). All changes made after the trial had begun were also sent to, and approved by, this ethics committee (number 6000207769, approved on April 03, 2020) and have been published elsewhere (e1).

Inclusion of practices and patients

The Association of Statutory Health Insurance Physicians Westphalia-Lippe (KVWL) had invited primary care practices to participate. Those practices were included which

• cared for Barmer patients,

• employed at least one primary care physician with a specialist qualification in general medicine or internal medicine or without a specialist qualification,

• had at least ten patients potentially eligible to participate (potential patients),

• were able to provide the primary care physician with VPN access to the KVWL website via a secure connection, and

• had at least one primary care physician who agreed to fulfill the contractual obligations arising from the study.

Adult patients were eligible to participate if they were covered by the Barmer health insurance fund during the study period and had been prescribed at least five drugs (defined by the codes of the Anatomical Therapeutic Chemical [ATC] classification system and documented in their billing data) in at least one quarter of the previous year. Each of the five drugs had to have been prescribed during at least two consecutive quarters.

Before randomization and in each quarter of the intervention period, participating patients were identified from Barmer billing data. The patients had to give their written consent to participate in the intervention.

Randomization and blinding

The Department of Medical Informatics, Biometry and Epidemiology (AMIB) at the Ruhr University Bochum randomized the practices every month in a ratio of 1:1 to the intervention or control arm. For this purpose, the KVWL provided lists of the newly recruited practices to the AMIB, where randomization was carried out by an independent person using computer-generated numbers. The results of randomization were sent to the KVWL, and the assigned group was disclosed to the participating practices (e1). Once a practice had been assigned to either the intervention or the control arm, all eligible patients and the medical staff of this practice were assigned to the same arm. Whereas the practices in the intervention group started the intervention immediately after randomization (Q1/2018–Q2/2019), the practices in the control group continued standard care for five quarters before starting the intervention (Q2/2019–Q3/2020; eFigure).

Given the nature of the intervention, neither the primary care physicians, their patients, nor the AdAM study team were blinded with respect to treatment allocation.

Endpoints

The primary endpoint was a combined dichotomous endpoint comprising all-cause mortality and all-cause hospital admissions.

The principle secondary endpoints were mortality, hospital admissions (dichotomous), and a combined dichotomous endpoint representing the 19 different potentially inappropriate medications (PIMs) (“patients with any risk factor and one or more high-risk prescriptions as defined in SOpim measures 20 to 22”, SOpim-C; see eSupplement Table 2 for details). The PIMs were defined in advance from the literature and operationalized via ATC and ICD codes (International Statistical Classification of Diseases and Related Health Problems) (eSupplement Table 2). Other secondary outcomes included cause-specific hospital admissions, preceded by high-risk prescribing, and potential prescribing omissions (operationalization eSupplement Tables 3 and 4; Results eSupplement Tables 13 to 19).

The results were measured on a quarterly basis at patient level for 14 quarters (October 01, 2017, thru March 31, 2021). The information on all patients eligible to participate was extracted from Barmer performance data in pseudonymized form. The billable interactions (health insurance benefits) between the insured persons and the healthcare delivery system are recorded in the performance data.

The hospital admissions included night-only and day-only admissions as well as partial and full inpatient stays. The date of admission defined the assigned quarter. The date of death was determined for mortality.

All patient characteristics were registered for the baseline quarter, which was defined as the first quarter in which the patient fulfilled the criteria for the analysis. Consequently, the data for the last year relates to the four quarters prior to the baseline quarter. To summarize, it can be said that a higher nursing care level (0–5) corresponds to a greater need for care and determines the amount of care services that a patient is entitled to. All ICD codes (inpatient or outpatient) listed in van den Bussche et al. (e2) (eSupplement Table 5) were used to measure the number of chronic conditions. The average number of repeat prescriptions was calculated per patient and was defined as the number of prescribed ATC codes in two consecutive quarters of the last year. The medication-based Chronic Disease Score is a prognostic index which takes into account age and sex as well as a list of evidence-based medications for the treatment of older chronically ill outpatients and was used as a surrogate for morbidity estimation (e3). Higher scores are taken as an indication of a higher mortality risk. The medication-related part of the medication-based Chronic Disease Score is presented separately (only the part calculated using the ATC classification), as it was included in the analysis model in addition to age and sex. Two different screening tools were used to describe PIMs (Priscus [e4] and EU-PIM [e5]). Additionally, the prescription rates of drugs with anticholinergic properties and sedatives were calculated using the Anticholinergic Drug Scale (ADS) (e6), the Anticholinergic Drug Burden (ADB) (e7), and a modified Drug Burden Index (DBI mod.) (e8), whereby the dosages had to be ignored as they were not available in the performance data.

Information on the participating practices and the primary care physicians was provided by the KVWL (e.g., socio-economic data, type of practice, and number and specialization of the primary care physicians per practice). They were collected at the time of enrolment (primary care physician profile) and randomization (practice profile). The KVWL data are not systematically linked to Barmer data on either a primary care physician or patient level. The practice size is expressed as the number of registered patients over a period of three months after randomization.

Safety and adverse events were neither monitored nor reported as it was assumed that the study would not produce a deterioration in treatment.

Statistical analysis

When calculating the sample size, it was assumed that the incidence of hospital admissions and mortality in patients with polypharmacy in the control group would be 35.25 percent during a follow-up period of 15 months (5 quarters) (e9, e10). Based on an intra-cluster correlation coefficient (ICC) of one percent and a recruitment rate of 80 percent, a sample size of 21 500 patients per group (17 200 recruited patients and 539 practices per study arm, i.e. approximately 32 recruited patients per practice) was originally calculated as being required to detect an absolute difference in the combined endpoint of 1.8 percent between intervention and control groups (with a type 1 error of five percent and a type 2 error of 15 percent). However, the targeted number of practices was not achieved, and only 688 practices were randomized in the period from June 27, 2017, thru July 03, 2019 (eSupplement Table 1). Based on the assumptions of 26 832 patients eligible to participate (39 per randomized practice), a participation rate of 60 percent in the intervention group, the same number of practices at the transition times (i.e., the transition from control to intervention period), and a constant event rate in the control group, extensive simulations showed that, under otherwise identical assumptions, a power of 80 percent could be achieved by switching to the SWD (eFigure).

The main analysis followed the intention-to-treat principle (ITT principle) and analyzed the control and intervention periods in accordance with randomization. The full analysis cohort included all patients who had fulfilled the following three criteria between October 01, 2017, and March 31, 2021 (see below for details on inclusion criteria, “Inclusion criteria for the analysis population”):

• The patient fulfilled the inclusion criteria described above.

• The primary care physician who wrote the most prescriptions participated in the AdAM study.

• The patient was treated by the AdAM primary care physician at the beginning of the observation or intervention period. All three criteria were reviewed separately for each period. The patients were included in the analysis cohort of the corresponding period as from the quarter in which they met the above-mentioned entry criteria. Primary care practices and their potentially eligible patients were excluded from the main analysis if none of their patients had participated during the intervention periods (so-called inactive practices, modified ITT analysis).

To account for time effects, the analysis was performed by fitting an appropriate mixed logistic model with a fixed time effect for each quarter (e11–e14). Additionally, fixed effects for treatment group and sex, two random effects correlated at cluster level (one for the control periods and one for the intervention periods) (e15), and an uncorrelated random patient effect were included. For the model, the covariates age, medication-based Chronic Disease Score (e3), and care level were included quarterly as fixed effects in the model in order to control for developments such as the progression of disease burden and age within the cohort. The results are presented as odds ratios (OR) with 95 percent confidence intervals (CI).

Based on an expected learning curve, it was assumed that the intervention effect would gradually increase from the time the primary care physician participates in the intervention and gains access to the AdAM software (¼ in the quarter in which the primary care physician participates in the intervention, ½ in the following quarter, full effect after that). Mixed logistic models were calculated for all primary and principle secondary endpoints.

Three sensitivity analyses were carried out (overview in eTable):

• an analysis of the originally planned parallel group comparison (c-RCT without inactive practices), i.e. practices and their patients were monitored for five quarters after randomization

• an analysis which included only quarters before the first COVID-19-related lockdown (until March 31, 2020)

• an analysis which included the data of all randomized practices (including the inactive practices), but also took only the quarters before the COVID-19 pandemic into account. Similar mixed logistic models were used for all sensitivity analyses with the exception of sensitivity analysis 1, in which age, care level, and medication-based Chronic Disease Score were only considered at the beginning of the c-RCT period.

Although Cox models are commonly used to analyze mortality outcomes, this method of analysis was not initially chosen because extensive simulation studies at the time of the design switch showed that the use of Cox models for the stepped-wedge design (two random effects) was inferior to logistic regression analysis in terms of both power and bias (e16). The analyses were therefore primarily carried out using logistic models in order to achieve good comparability of the results, especially identical indicators (odds ratios). The study protocol and the statistical analysis plan included a sensitivity analysis which used only data from the original c-RCT. Therefore, a Cox regression analysis with robust variance estimation and taking cluster effects into account was performed for sensitivity analysis 1 and the secondary endpoint mortality. Fixed effects for sex, age, care level, and the medication-based Chronic Disease Score, as measured at the beginning of the c-RCT period, were also taken into consideration. Contrary to the study protocol, only quarters before the COVID-19 pandemic were included in sensitivity analysis 3, as initially inactive practices were more frequently randomized to the control group and only started the intervention after the COVID-19 pandemic.

Instead of the planned per-protocol analysis, a dose-response analysis was carried out with the design change, as persons who died in the control phase could not be included in the analysis due to the lack of informed consent and would therefore have led to a systematic bias in the per-protocol approach. A descriptive comparison of the patients revealed relevant imbalances between the groups. These were most likely due to mortality during the observation periods and low recruiting rates in the original control clusters as a result of the COVID-19 pandemic. In the dose-response analysis conducted instead, a mixed Poisson model at cluster level (random cluster effect, offset variable for the number of patients, covariates: mean age, mean care level, mean medication-based Chronic Disease Score of the cluster) was used to calculate whether a higher proportion of patients in a practice treated with the Clinical Decision Support System (CDSS) leads to a greater intervention effect (eSupplement Table 23).

Eight subgroups of patients were first defined by age (below 65 years versus 65 years and older), sex, primary attending physician (primary care physician versus specialist or hospital outpatients department), and by incidental (patients were not on the list of potential patients before the practice was randomized) and prevalent cases (patients were already on the list of the primary care physician’s potential patients prior to randomization). Further details on methodology of subgroup analyses and the results may be found in the annex (eSupplement Table 24).

An appointed expert panel advised on the planning, implementation, and evaluation of the study.

All statistical analyses were conducted using R, version 4.1.2. The study was registered at ClinicalTrials.gov under the number NCT03430336.

Inclusion criteria for the analysis population

Patients had to meet the following three criteria between October 1, 2017, and March 31, 2021, (inclusion criteria) in order to be included in the analysis population:

• Inclusion criteria for participation: Participating patients had to have health cover with Barmer, be at least 18 years old, and have prescriptions for at least five medicines (≥ five different ATC codes) in at least one quarter of the previous year. Each of the five ATC codes must have been prescribed in at least two consecutive quarters of the previous year.

• Criteria for inclusion in the list of potential patients of an AdAM primary care physician: Patients were included in the list of potential patients if the primary care physician had prescribed most of the patient’s medication in the two previous quarters and prescribed at least four ATC codes in this six-month period. If several primary care physicians had prescribed the majority of the ATCs during this period, then the primary care physician with the highest number of daily doses was selected. If the number of daily doses was also the same, the primary care physician was selected at random. The lists were made available with a delay of one quarter.

• Criteria for ensuring contact between the AdAM primary care physician and the patient: Since the list of potential patients was made available with a delay of up to two quarters, an additional criterion was used to ensure that only those individuals treated by a particular AdAM primary care physician were included in the population to be analyzed. This meant that the AdAM primary care physician must have treated the patients in the quarter in which they appeared on the available list of patients eligible to participate or in the following quarter. Treatment was defined as the issuing of at least one prescription or the coding of at least one EBM code (doctor’s fee scale of the statutory health insurance system for the remuneration of medical services) by the primary care physician. At the beginning of the intervention period, it was once again verified whether the patient and AdAM primary care physician were in contact and thus participating in the intervention quarters.

Patients were included in the analysis population as from the quarter in which they had met the above inclusion criteria.

Description of the intervention

AdAM is a multifactorial intervention to support primary care physicians in user-initiated annual CDSS-supported medication reviews. For this purpose, primary care physicians accessed the CDSS, which is not connected to the practice management system, via a secure web-based portal (KV-SafeNet) and which provided the user with information at practice level and – after obtaining informed consent – at patient level. The CDSS contains diagnosis and treatment data relating to drug and non-pharmacological treatments based on performance data of the Barmer health insurance fund (Barmer). It supports cross-patient risk management at practice level (for patients receiving medication mentioned in Direct Healthcare Professional Communications [DHPCs]) as well as patient-specific medication reviews. For the medication review, the doctors accessed the information provided in the CDSS on the enrolled patients, updated and supplemented it with information that was not (yet) available in the Barmer performance data (new prescriptions, dosages, body weight, laboratory results). In response to the review, the CDSS displayed alerts (e.g., on drug interactions, contraindications, dosing errors, and duplicate medications). The physician was then able to optimize the medication and provide patients with a print-out in the format of the nationally standardized medication plan with multilingual explanations on drug therapy. Enrolled patients could be accessed in the CDSS as often as required, and a medication review conducted at least once a year was remunerated with 85 euros per patient. Before starting the intervention, the doctors were given the opportunity to participate in face-to-face and online training sessions on polypharmacy and the technical use of the CDSS. A process evaluation planned in advance examined user behavior and the implementation of the intervention (15, 16).

Study design

The study was originally planned as a cluster-randomized controlled trial with a parallel group design (parallel c-RCT). Practices in the intervention group applied the new electronic medication management support system when treating their patients, while control practices continued to provide standard care. Simulation calculations were conducted due to low recruitment rates of primary care physicians and patients – particularly as a result of the flu outbreak and coronavirus pandemic. We subsequently switched to a stepped-wedge design (SWD) in order to achieve a power of 80 percent. This meant that all the practices of the control group changed to the intervention phase after five quarters. Overall, the observation period was extended to include the data before randomization and after the end of the parallel c-RCT (eFigure).

eFigure.

Schematic representation of the AdAM-study in chronological order. Switching the study design to a stepped-wedge design with an open cohort resulted in an extension of the observation period. The parallel group comparison was subsequently extended as follows: The starting point was the practices included in the parallel group comparison, for which additional data for the period before and after the planned parallel group comparison phase (c-RCT period, shaded) were included in the assessment. This resulted in a control phase and a subsequent intervention phase for each practice (compare waitlist control group design).

* The randomization on March 07, 2019, was assigned to the second quarter of 2019.

AdAM, Application of an Electronic Medication Management Support System; c-RCT, cluster randomized controlled trial

Study population

General practices in Westphalia-Lippe, Germany, were informed about the study via regional media, were contacted by the Association of Statutory Health Insurance Physicians of Westphalia-Lippe (KVWL) by postal mail and invited to participate. Barmer also sent out information flyers to its insurees in the region. General practices willing to participate with

• healthcare services for patients insured with Barmer

• physicians with a specialist qualification in general medicine, internal medicine, or without a specialist qualification

• at least ten potentially eligible patients (“potential patients”)

• access to the KVWL website via a secure connection, and

• consent of the physicians to fulfill the contractual obligations arising from the study were admitted to the trial and randomized to the intervention or control group.

Potential patients are adult patients of the participating primary care practices insured with Barmer with at least five different medication prescriptions (defined as number of the ATC codes) for at least two quarters. They were identified during the observation period using Barmer performance data. All potential patients of the control and intervention practices whose GP participated in the study were included in the analysis if a contact had taken place during the observation phase. The patients were included in the analysis cohort of the corresponding period (open cohort) as of the quarter in which they met the entry criteria, regardless of consent (secondary data analysis). Primary care practices and their potential patients were excluded from the main analysis if none of their patients participated during the intervention period (inactive practices), as this is due to various organizational reasons (17).

Primary and secondary endpoints

The primary combined dichotomous endpoint comprised mortality and hospital admission (whichever occurred first). The principle secondary endpoints were mortality, hospital admissions, and a combined dichotomous endpoint which represented the 19 high-risk prescriptions for gastrointestinal bleeding, cardiovascular risks, and falls (eSupplement Table 2). The endpoints were measured quarterly at patient level for 14 quarters (October 01, 2017, thru March 31, 2021). The information regarding patients eligible to participate was obtained in pseudonymized form from Barmer performance data.

Statistical analysis

The analysis was conducted according to the intention-to-treat principle (modified ITT after inactive practices were excluded) using the Barmer performance data (secondary data analysis) and employing a mixed logistic model.

Three sensitivity analyses were performed:

• an analysis of the originally planned parallel group comparison (parallel c-RCT without inactive practices), i.e. the practices and their patients were monitored for five quarters after randomization,

• an analysis which included only the quarters before the first COVID-19-related lockdown (until March 31, 2020) (when routine care was restricted in hospitals and general practices), and

• an analysis which included the data of all randomized practices (including the inactive practices) but also took into account only the quarters before the COVID-19 pandemic.

Additionally, a Cox regression analysis with robust variance estimation taking cluster effects into consideration was conducted for sensitivity analysis 1 in the secondary endpoint of all-cause mortality (was not originally planned).

In a dose-response analysis, a mixed Poisson model was used to calculate at cluster level whether a higher proportion of patients treated with CDSS leads to a greater intervention effect of the practice.

Results

From June 2017 thru July 2019, 1348 practices (identified using main practice number) in Westphalia-Lippe were invited to participate in AdAM; of these, 688 practices were randomized (intervention/control: 343/345). A total of 937 primary care physicians agreed to participate in the study, of which four study physicians had no potential patients. Since access to software was linked to the primary care physician, several physicians could participate in the study per practice premises. Similarly, the medical staff could change premises during the study or work at several (main/subsidiary) premises. This increased the total number of clusters (practices) in the intention-to-treat analyses to 746 practices (Figure 1).

Figure 1.

Flow chart for the project “Application of an Electronic Medication Management Support System” (AdAM)

*¹ 1 Since access to the software was linked to the primary care physician, randomization of the main practice determined the group assignment of the PCPs. A total of 72 PCPs changed practice premises during the course of the study or worked at several practice premises at the same time. As a result, there are more clusters than randomized main practice sites in which study physicians work.

*² This corresponds to the blue box prior to the c-RCT period in the eFigure.

*³ Only clusters and PCPs that had evaluable potential patients in this phase were counted.

*⁴ Includes four practices which were in both groups

c-RCT, cluster-randomized controlled trial; PCPs, primary care physicians

The data of 42 700 patients were included in the intention-to-treat analyses. Of these, 23 582 were monitored both during the control period and during the intervention period (blue and orange boxes, respectively, in the eFigure). Data for 6181 patients were available only for the control period and for 12 937 only for the intervention period. A total of 391 994 patient-quarters were monitored during the period October 01, 2017, thru March 31, 2021 (eFigure).

Of the 746 analyzed practices, 411 (55%) were active practices, the median (IQR) of the enrolled patients was 18 (8 to 33), corresponding to a median (interquartile range – IQR) enrollment rate of 35.8 percent (14.8 to 53.6%) of the potential patients. Enrolled patients were on average somewhat older, less in need of care, and took slightly more prescribed medications (eSupplement Table 25).

Characteristics at baseline

There were no significant differences between the two groups at the baseline examination (Table 1).

Table 1. Characteristics of patients, physicians, and practices at baseline.

Characteristics of patients
	Intervention (n = 21 543)	Control (n = 21 157)
Age, years	71.6 (13.0)	72.0 (13.0)
Sex (female)	13 512 (62.7%)	13 287 (62.8%)
Gainfully employed	4367 (20.3%)	4359 (20.6%)
Care grade 1–5	5666 (26.3%)	5704 (27.0%)
Chronic illnesses during the last year	11.9 (4.63)	11.9 (4.67)
Repeat prescriptions in the last year*	6.31 (2.27)	6.32 (2.29)
Practice characteristics
	Intervention (n = 343)	Control (n = 345)
Practice type: single practice (otherwise group practice or outpatient care centers)	206 (60.1%)	235 (68.1%)
Practice size (according to number of registered patients)
1–870	49 (14.3%)	39 (11.3%)
871–1163	63 (18.4%)	72 (20.9%)
1164–1468	61 (17.8%)	68 (19.7%)
1469–2027	79 (23.0%)	74 (21.4%)
≥ 2028	91 (26.5%)	92 (26.7%)
Characteristics of physicians
	Intervention (n = 478)	Control (n = 459)
Age, years	54.3 (8.77)	54.7 (8.62)
Sex (female)	173 (36.2%)	154 (33.6%)

Categorical data are displayed as frequencies and percentages.

Either the mean (standard deviation) or the median (1st quartile to 3rd quartile) is stated for continuous data.

*ATC codes must have been prescribed in at least two consecutive quarters during the last year. See eSupplement Table 12 in the annex for the long version.

ATC, Anatomic Therapeutic Chemical classification system;

SD, standard deviation

Primary Endpoint

In the main analysis, no significant decrease was observed in den intervention periods for the combined endpoint (odds ratio [OR] 1.00; 95% confidence interval: [0.95; 1.04]; p = 0.872; violet circle in Figure 2). Sensitivity analysis 1 in the parallel c-RCT period, where each cluster represented either a control practice or an intervention practice, produced similar results (OR 0.99 [0.93; 1.06]; p = 0.852; blue square in Figure 2).

Figure 2.

Main and sensitivity analysis of the primary endpoint and the principle secondary endpoints. The results for the different populations to be analyzed (colors) and the different adaptation variants (symbols) are presented.

An odds ratio (OR) less than one favors the intervention.

Adjustments (Adj.): quarterly adj. – quarterly adj. of age, care level and medication-based Chronic Disease Score;

Adj. at the beginning of the parallel c-RCT period – Adj. on admission to the c-RCT period. Owing to the pandemic, analyses 2 and 3 were specified as further sensitivity analyses.

Parallel c-RCT, cluster-randomized controlled study with a parallel group design

A significant decline in hospital admissions and, with it, also in the combined primary endpoint was noted at the start of the COVID-19 pandemic. Before the start of the pandemic, a primary endpoint event occurred in 16.6 percent (18 599 of 111 811) of the intervention quarters, compared with 17.3 percent (21 993 of 126 886) in the control periods. This had an impact on the analyses with the exclusion of the COVID quarters: The effect estimates for the OR were larger in favor of the intervention in both sensitivity analysis 2 (OR 0.97 [0.92; 1.02]; p = 0.237; green circle in Figure 2) and sensitivity analysis 3 (OR 0.97 [0.93; 1.01]; p = 0.138; red circle in Figure 2) for the pre-pandemic quarters.

Secondary Endpoints

As regards the principle secondary endpoints, the intervention did not result in any reductions in the main analysis (hospital admissions: OR 1.00 [0.95; 1.05]; mortality: OR 1.04 [0.92; 1.17]; high-risk prescription: OR 0.98 [0.92; 1.04]; violet circles in Figure 2 and Table 2). Sensitivity analysis 1 in the parallel c-RCT period produced similar results (hospital admissions: OR 1.00 [0.94; 1.06]; mortality: OR 0.93 [0.8; 1.07]; high-risk prescriptions: OR 0.98 [0.89; 1.08]; blue squares in Figure 2).

Table 2. Main and sensitivity analyses of the primary endpoint and the principle secondary endpoints.

	Combined primary endpoint		Hospital admissions	Mortality	Combined endpoint of inappropriate medications
Population	OR [95% CI]	p-value	OR [95% CI]	OR [95% CI]	OR [95% CI]
Main analysis	1.00 [0.95; 1.04]	0.872	1.00 [0.95; 1.05]	1.04 [0.92;1.17]	0.98 [0.92;1.04]
Sensitivity analysis 1 (parallel c-RCT)	0.99 [0.93; 1.06]	0.852	1.00 [0.94; 1.06]	0.93 [0.80; 1.07]	0.98 [0.89; 1.08]
Sensitivity analysis 2 (before COVID-19 pandemic)	0.97 [0.90; 1.02]	0.237	0.97 [0.92; 1.02]	0.97 [0.85; 1.10]	0.95 [0.88; 1.01]
Sensitivity analysis 3 (before COVID-19 pandemic and with inactive practices)	0.97 [0.93; 1.01]	0.138	0.97 [0.93; 1.01]	0.98 [0.88; 1.08]	1.00 [0.94; 1.05]

Presented here are the results for the various populations to be analyzed. An odds ratio (OR) less than 1 favors the intervention.

Adjustments (Adj.): Main analysis, sensitivity analysis 2 and 3 – quarterly adj. for age, care level and medication-based Chronic Disease Score; sensitivity analysis 1 – Adj. for the time of the start of the parallel c-RCT period. Observed endpoints: combined endpoint comprising all-cause mortality and all-cause hospital admission – primary endpoint, hospital admission – secondary endpoint, mortality – secondary endpoint, combined endpoint of inappropriate medications – secondary endpoint, patients with any risk factor and one or several high-risk prescriptions according to the definition in SOpim-20 to 22, see eSupplement Table 2.

CI, confidence interval; OR, odds ratio; parallel c-RCT, cluster-randomized controlled study in parallel group design

The post-hoc analysis of the Cox model showed a lower probability of death for the intervention (hazard ratio [HR] 0.89 [0.787; 0.997]; p = 0.0451; Figure 3).

Figure 3.

Kaplan-Meier plot of time since randomization (months) and after the randomized treatment group in the originally planned c-RCT period.

c-RCT, cluster-randomized controlled trial

The dose-response analysis showed that treating more patients per practice with the new model of care was associated with a greater reduction in the event rate (relative risk 0.95 [0.90; 0.99]; p = 0.0182; corresponding to the assumptions regarding the effect size when planning the study).

Discussion

On the whole, the planned analyses did not reveal any significant effect of the intervention on the combined primary endpoint and the principle secondary endpoints. Given the inadequate practice and patient recruitment and the high rate of inactive practices (45%), extensive simulations were carried out, which revealed a strong loss of power of the original study design. By switching the study design to an SWD and the associated extension of the observation period as well as an evaluation strategy using mixed effects logistic regression with repeated measurements (GLMM), it was possible to ensure the required 80 percent power in the simulations. Therefore, the preplanned GLMM was used for the analysis of all endpoints (18).

The COVID-19 pandemic had a significant impact on the study, its endpoints, and analyses, as care for chronically ill patients was restricted during this time, they were hospitalized less frequently (19, 20), and the SWD responds sensitively to time effects (21–23). The parallel c-RCT design, which is insensitive to time effects, was therefore adopted as the planned sensitivity analysis for the primary endpoint and the principle secondary endpoints. An unplanned Cox model analysis was conducted as a further sensitivity analysis, since this provides unbiased estimates of mortality in a parallel group design that is insensitive to time effects. The Cox model analysis revealed that mortality during the intervention period fell by ten percent. The effect estimates of other sensitivity analyses, which included only the pre-pandemic quarters, pointed in the same direction (19–23). Practices with above-average patient inclusion showed a higher intervention effect (dose-response analysis), indicating that the low patient recruitment due to the pandemic prevented a stronger effect on the primary endpoint. Another effect of the pandemic was that many of the original control practices did not conduct any intervention after the switch (inactive practices), which in turn reduced the power particularly in the main analysis.

We consider this to be the first prospective, randomized controlled study to indicate that a CDSS-supported medication review in adult patients with polypharmacy in primary care could prevent deaths. A recently published non-randomized retrospective study showed that a collaborative medication management involving physicians and pharmacists significantly lowered mortality of patients with polypharmacy (24). Furthermore, a meta-analysis showed a relative mortality reduction of 26 percent (corresponding to an absolute reduction of 1.4%) in an elderly population with polypharmacy who were subjected to a comprehensive medication review (25). A systematic review covering a greater number of studies, however, did not reveal any reduction in mortality (26). In line with previous studies on polypharmacy, our intervention had no impact on hospital admissions (26–29). Unlike our study, Dreischulte et al. showed a decrease in hospital admissions after high-risk prescription against which targeted interventions were directed (30). Interventions to reduce PIM prescriptions more often resulted in an improvement in health care processes (14, 26–28), the relevance of which is uncertain for patient-relevant outcomes, however, as PIMs are not the main cause of medication-related hospitalizations (31).

The cluster design of the AdAM study and the exclusive use of health insurance data have the advantage that relevant data is available without omissions and measurement bias is largely avoided (26, 32).

The AdAM study had a number of limitations, especially case number limitations which made the design switch to SWD necessary, as well as unfavorable time effects of the pandemic (21–23). Time effects result from the fact that in the SWD the observations of the control phases took place earlier in the course of the study than those of the intervention phases. The pandemic-related reduction of hospital admissions occurred above all during the intervention phase. A further time effect-related restriction consisted of the need to adjust for covariates, such as disease burden progression and aging. By adjusting the prognostic index of the medication-based Chronic Disease Score (med-CDS) and the care levels in the analyses on a quarterly basis, a potential improvement resulting from the intervention was reduced. In addition, our effect size may have been underestimated because, firstly, the selection criteria favored the inclusion of low-risk patients (low prevalences of PIM prescription, anticholinergic burden, low med-CDS score, eSupplement Table 12), who may not, or may not significantly, benefit from the intervention (33). Secondly, the benefits of the intervention may not have been fully exploited: incomplete learning curves due to low patient numbers in the clusters, possibly incomplete data entry for medication review due to technical barriers and lack of integration of the CDSS into practice management systems, time constraints on the part of medical staff due to pandemic management, and lack of training (16).

Although the AdAM intervention did not show any significant effects in the planned analyses, there is evidence that CDSS-assisted medication reviews and treatment planning could potentially reduce mortality in adult patients with polypharmacy in primary care, given that the effect estimates of all sensitivity analyses conducted before the COVID-19 pandemic support this hypothesis. Together with other studies, in which the quality of prescribing and patient safety were improved, these results justify a regular CDSS-supported medication review. Further studies are needed to improve implementation, integrate CDSS into workflows (34), and intensify training in the use of CDSS and medication optimization (35–38).

In summary, the present study was unable to show any significant effects on the primary endpoint and the principle secondary endpoints in the planned analyses. Due to the pandemic and recruitment difficulties, unplanned analyses were conducted that provided indications of a possible reduction in mortality in the intervention group. Controlled trials with appropriate follow-up and a better implementation strategy are needed to prove that a CDSS has significant effects on mortality.

eTable. Overview of the analyses.

Mixed logistic models						Cox model	Mixed Poisson model
Main analysis	Sensitivity analyses			Subgroup analyses		Post-hoc sensitivity analysis
	1	2	3	Age, sex, main treating physician, cancer, dialysis or palliative patients	Incidental and prevalent cases		Dose–response analysis: ratio between the proportion of patients included and the difference in logarithmic event rates
Without inactive practices	c-RCT	Before COVID-19 pandemic and without inactive practices	Before COVID-19 pandemic and with all randomized practices, i.e. also with inactive practices	Before COVID-19 pandemic and without inactive practices	c-RCT	c-RCT	Dependent variable: – number of events in the primary endpoint Independent variables: – proportion of included patients – mean age – mean care level – mean medication-based Chronic Disease Score per practice – random cluster effects – offset variable that takes into account the different number of quarters

c-RCT, cluster-randomized controlled study