Reduction of antibiotic prescriptions for acute respiratory tract infections in primary care: a systematic review

The systematic literature search was carried out in MEDLINE/PubMed and the Cochrane Library using the following search terms:

((antibiotic*) AND (“respiratory tract infection” OR “respiratory tract infections” OR “res-piratory infection”) AND (communication OR training OR “point of care test” OR “rapid strep test” OR “delayed prescribing” OR intervention* OR “electronic decision support” OR “clinical decision support system” OR “clinical decision support systems” OR “shared decision making”) AND (“primary care” OR “primary health care” OR “medical care”)) OR ((“Anti-Bacterial Agents”[Mesh]) AND (“Respiratory Tract Infections”[Mesh]) AND (“Primary Health Care”[Mesh] OR “Physicians, Primary Care”[Mesh]))

Additionally, the bibliographies of the included trials were screened for relevant intervention trials. We could not search further databases like EMBASE due to limited access and lack of funding.

The literature search included (cluster-) randomised controlled trials (RCTs) investigating the effect of interventions aiming to reduce antibiotic prescriptions for RTIs for patients ≥ 13 years in primary care settings. We excluded pilot trials and non-randomised trials. Eligible interventions were educational seminars, feedback on prescribing behaviour, patient education, communication skills training (CST) for physicians and diagnostic tools such as point-of-care tests (POCT) or (electronic) CDSS. We did not take public campaigns into account. These address a broad audience and use mass media to raise awareness for the problem of inadequate antibiotic prescribing and its influence on bacterial resistance. As our systematic review focuses on primary care, this type of intervention is not considered.

We investigated the primary outcome of Abx prescription rate as well as the number of antibiotic prescriptions for acute upper/lower RTIs (cough and sore throat). Any reported secondary outcome such as patient’s reconsultation rate, days to recovery from RTI, prescribed class of antibiotic, rate of inappropriate antibiotic prescriptions (prescriptions not according to guidelines), Abx prescription rate for specific RTIs or usage of diagnostic devices was of interest to this systematic review.

We only included patients ≥ 13 years as we know that clinical decision-making in paediatric medicine differs from adults. Further, the communication in dyadic consultations between adults differs from “doctor-parent-child triads” and requests other communication styles [12]. In consequence, we excluded children under 13 years. We included primary care physicians working in ambulatory care. Due to different health systems, we included physicians working in practices or primary care clinics.

We only considered publications written in English, German or French dated January 2005 till July 2016.

Two reviewers (AK, SR or JFC) independently reviewed the titles, index terms and abstracts of the identified references and rated each paper as potentially relevant or not. Discrepancies were resolved by consensus.

Relevant information for each trial included into the review was extracted by one reviewer and controlled independently by another one (AK, SR). Discrepancies were resolved by consensus. Three authors had to be contacted due to missing or unclear data about the age of included patients [13‐15]. Gjelstad et al. offered odds ratios (OR) for antibiotic prescriptions for the age group ≥ 13 years, but no Abx prescription rate. It was possible to get unpublished Abx prescription rates for the subgroup of patients ≥ 13 years [14]. Missing data—like p values or absolute numbers of antibiotic prescriptions—was indicated as “not specified” in our tabular summaries of included RCTs (see Tables 1 and 2).

If available, antibiotic use was presented as the absolute number of prescribed antibiotics. In addition, relative Abx prescription rates expressed as a percentage—with the corresponding 95% confidence intervals (= 95% CI) and p values—were indicated as well as the difference of Abx prescription rates between intervention group (IG) and control group (CG) (in percent, OR, relative risks (RR) (Tables 1 and 2). The Abx prescription rates before (T0) and after the intervention (T1 or T2) were shown for trials providing pre-post comparison. As RCTs were heterogeneous in study design and therefore in time points for T1 or T2, we decided to not exclude RCTs by defining a binding point of time for T1 or T2. Details of included RCTs can be seen in a tabular summary of study characteristics.

Two researchers (AK with SR, FB or JFC) independently assessed the risk of bias using the Cochrane Collaboration’s tool for assessing risk of bias in randomised trials (Fig. 1) [16]. We discussed and resolved discrepancies by consensus.

Due to the heterogeneity of study designs, outcome measures and Abx prescription rates, we considered a pooled estimate of the effect size with Cohen’s d to be inappropriate [17].

There is no consensus on which change in Abx prescribing rates marks a meaningful intervention effect. The difference between current Abx prescription rate and optimal prescription rate reflects the range for reducing antibiotic prescriptions to a meaningful extent. But the generalisability for determining an optimal rate is limited by patient age and condition with different likelihoods for bacterial genesis (pneumonia vs. common cold) [18]. One report from the year 2016 states that 44% of all ambulatory antibiotic prescriptions in the USA are due to RTIs. It is said that 50% of these prescriptions are inappropriate leading to an optimal prescribing rate of about 20% for all RTIs [18]. Another publication suggests a prescribing rate of 10–15% for acute cough [19].

As a compromise in this complex field and to simplify the overview of clinically relevant reductions in Abx prescriptions, we will consider an absolute difference of 10% between IG and CG for studies with post-intervention as minimal important change. For studies with baseline and follow-up, we regard a difference in differences of 10% as minimal important change. We acknowledge that this threshold is arbitrary, but a different threshold would not have changed our findings fundamentally.

We searched databases on the 31st of July in 2016 and identified 690 publications. Reviewers independently screened for potentially relevant publications and categorised 215 publications as potentially relevant. After removal of duplicates (n = 84), the remaining 151 titles and abstracts were screened for eligibility and discussed by the reviewers. Disagreements between reviewers were resolved by consensus. Major reasons for exclusion were different trial populations, e.g. including children, trials carried out in non-primary care settings and non-randomised study designs. We excluded 84 out of 151 publications. A total of 67 potentially relevant articles were fully screened for inclusion and exclusion criteria (Fig. 2).

Seventeen trials were finally included in this analysis.

Seventeen RCTs met the inclusion criteria and were included in this review (Tables 3 and 4). Thirteen trials were cluster RCTs based on physician-, practice- or educational group levels as clusters [13‐15, 20‐29]. Four RCTs were randomised at patient level [30‐33]. The majority of trials used a two-arm study design [13‐15, 20, 24, 25, 27, 28, 30‐33]. The remaining trials employed a three-arm [21, 23] or factorial study design [22, 26, 29].

All cluster-randomised trials performed cluster-adjusted data analyses. The number of participating physicians ranged from 6 to 573. Most trials were conducted in Europe [13, 14, 20‐22, 27, 28, 31], six trials in North America [15, 23‐25, 30, 32] and one in Asia [33]. One trial was a multinational project carried out in six European countries (Belgium, Spain, Wales, England, Poland and the Netherlands) [26]. Published baseline data was available for seven trials [14, 20, 23, 26‐29]. Eleven trials assessed the Abx prescription rate right after patients’ initial consultation [13, 20, 22, 23, 26‐31, 33].

Data on Abx prescription rates was collected directly by physicians [20, 21, 26, 27, 31], by pharmacists using faxed or mailed prescriptions [21, 32], by field researchers [33] or by electronic medical records [13, 15, 22‐25, 28, 29]. Four trials used special documentation software [14, 23‐25].

The time period for registration of Abx prescriptions ranged from right after the initial consultation up to 28 days after initial consultation.

Six trials assessed effectiveness of the intervention after a longer period of time, within 1 year [15, 29] or within 18 months after the intervention [30], after 1 year [13, 20] or after 3.5 years (Tables 3 and 4) [34].

Ten trials recruited primary care physicians in private practices [13, 14, 20‐22, 26‐29, 32], and seven trials recruited physicians from primary care clinics [15, 23‐25, 30, 31, 33].

Patients ≥ 13 years with acute upper and lower RTIs were included. The average age of patients was similar across trials and ranged from 40 to 53 years. The number of registered consultations varied from 149 to 1,115,359 [28, 32].

Twelve trials reported statistically significant lowered Abx prescription rates in the IGs compared to CGs [13‐15, 20, 22, 23, 26‐31]. In five RCTs. the Abx prescription rates could not be reduced significantly (see Tables 1 and 2) [21, 24, 25, 32, 33]. Using our definition for a clinically relevant reduction of prescription as criterion for efficacy, only six trials had a meaningful effect on Abx prescription rates [22, 23, 26, 27, 30, 31]. The effect of the interventions cannot be compared directly due to heterogeneous study designs.

Two RCTs implemented interventions addressing all three “target groups”. Both trials could significantly reduce their Abx prescription rates [23, 28]. Gulliford et al. combined a CDSS with patient handouts and DP. The proportion of consultations with antibiotics prescribed declined marginally from 53 to 52% during 12 months after intervention, whereas it remained constant at 52% in the CG. The adjusted mean difference in antibiotic prescriptions was − 1.85% (p = 0.038) [28]. Gonzales et al. observed a reduction of 12% in the first IG (poster with clinical examination algorithm) and a reduction of 13% in the second group (CDSS). Both interventions were combined with patient handouts, feedback on prescribing and seminars for physicians. Compared to CG, both interventions were statistically significant (p = 0.003 or p = 0.01), but not between themselves (p = 0.67) [23].

Five RCTs combined physician- and diagnosis-centred interventions [13, 22, 24, 26, 30], four of them reducing Abx prescription rates to a statistically significant extent [13, 22, 26, 30]: In their 2010 trial, Cals et al. reduced the Abx prescription rate by 12% (RR in the IG = 0.81, 95% CI 0.62–0.99) within a 28-day follow-up with the help of CRP POCT and DP [30]. Bjerrum et al. implemented CRP POCT, seminars for physicians and feedback on prescribing. The Abx prescription rate in the IG was 24% (CG 32%) 1 year after the intervention [13]. In the RCT by Little et al., the combination of CST and CRP-POCT led to a significant reduction of Abx prescriptions compared to CG (58 vs. 32%; p < 0.001) [26]. As the trial of Cals et al. was designed as a factorial trial, there was no testing for significance for the multifaceted intervention consisting of CRP POCT and CST [22]. The Abx prescription rate at index consultation was 23% (95% CI 11.6–34.6) and lower than in the CG (67%; 95% CI 53.9–79.5). Prescription rates for follow-up were not indicated. The RCT by Linder et al. from the year 2010 used CDSS and feedback on prescribing and found no difference between IG and CG [24].

The RCT by Linder et al. from 2009 tested patient handouts in combination with a CDSS. This intervention led to a non-significant reduction of 4% in the IG [25].

In the trial by Altiner et al., the intervention contained patient brochures, a waiting room poster with information on RTI and a CST for physicians. The observed Abx prescription rate after 1 year was 36.7% in the IG compared to 64.8% in the CG. Yet, different baseline rates should be considered (36.4 vs. 54.7%) [20].

Three RCTs used multifaceted interventions focusing at physicians [14, 21, 29]. Gjelstad et al. observed a reduction of 1.5% in the IG (p value = 0.027) [14]. The IG received feedback on prescribing, DP and seminars for physicians. However, they registered an increase of 1.7% in the CG (p value = 0.002). Continuing medical education groups with the corresponding prescription rates served as calculation basis instead of physicians with their individual prescribing rates.

In the trial of Briel et al., physician education alone reduced Abx prescription rates by 5.7%. In combination with CST, the reduction was increased to 7.9% compared to the CG. Both differences were not statistically significant [21]. The combination of interventions in the RCT by Meeker et al. did not result in statistically significant lowered inappropriate Abx prescription rates compared to CG [29].

Table 6 shows the number of studies and intervention concepts with clinically relevant reductions of Abx prescription rates with regard to type of intervention. Six trials had a meaningful effect on Abx prescription rates [22, 23, 26, 27, 30, 31]. Three out of these six RCTs had a three-armed or factorial study design, therefore containing more than one intervention concept [22, 23, 26]. Altogether, our review contains 11 clinically relevant intervention concepts. The majority contained diagnosis-centred elements [22, 23, 26, 27, 30, 31], especially CRP POCT [22, 26, 27, 30].

POCT (CRP and RADT) and CST reduce effectively antibiotic prescriptions alone or in combination [22, 23, 26, 30]. One RCT combined CDSS with knowledge transfer and patient handouts effectively [23].

Predefined secondary endpoints were patient-centred outcomes (e.g. reconsultation rate [21‐23, 25‐27, 29], patient satisfaction [21, 22, 30], patients’ views on RTIs [33] or days to recovery [22, 26, 27, 30, 31]), outcomes related to antibiotic prescribing (e.g. Abx prescriptions according to guidelines [24, 25, 31], class of prescribed antibiotic [13‐15, 31], prescribed antibiotics for specific RTIs [13, 28]) or diagnostics (e. g. number of X-rays [27], number of RADTs [15]). Four RCTs did not find any significant difference in secondary endpoints between IG and CG [20‐23]. One RCT did not provide any information on secondary endpoints [32].

This review updates and summarises current evidence of various interventions in primary care on reducing Abx prescriptions in patients ≥ 13 years for acute RTI. Twelve out of 17 included RCTs showed a statistically significant lower Abx prescription rate in the IG [13‐15, 20, 22, 23, 26‐31]. However, only six of them reported a clinically relevant reduction according to our definition [22, 23, 26, 27, 30, 31]. Due to the three-armed or factorial study design, these six RCTs contained 11 clinically relevant intervention concepts. Interventions focusing at physicians (CST) and at the process of diagnosis-making (CRP POCT, RADT, CDSS) were—alone or in combination—the most effective interventions. Observed reductions for RCTs with baseline ranged from 1.5 to 23.3% and cannot be compared directly due to heterogeneous baseline Abx prescription rates, study designs and settings [14, 26]. For studies with post-intervention measurements, the differences between IG and CG were between 2.9 and − 44% [22, 33]. Studies with reported prescription rates below 20% did not show significant reductions in Abx prescribing rates [21, 33]. Post-intervention observation periods ranged from 2 weeks up to 3.5 years. Conclusions on long-term sustainability of interventions cannot be drawn.

Our findings are in line with other systematic reviews, which reported mixed results regarding interventions to reduce antibiotics with either a larger or narrower spectrum of interventions and setting [7, 8, 10, 38]. We focused on RTIs in primary care excluding other infectious conditions and settings like emergency rooms, hospitals or public campaigns [39, 40].

We included RCTs with single-element [15, 22, 26, 27, 29, 31‐33] and multifaceted interventions [13, 14, 20‐26, 29, 30] focusing at diagnosis-making [13, 15, 22‐31], at physicians [13, 14, 20‐24, 26, 28‐30, 32] or at patients [20, 23, 25, 28, 33]. Nine intervention concepts with meaningful effects on Abx prescription rates contained interventions addressing the process of diagnosis-making [22, 23, 26, 27, 30, 31], five of them in combination with interventions targeting physicians [22, 23, 26, 30]. In contrast to the systematic review of 2005, single-element interventions can be effective (Table 6) [22, 23, 26, 27, 30, 31]. Interventions addressing patients were less likely to reduce Abx prescriptions to a meaningful extent. POCT (CRP and RADT) and CST—alone or in combination—reduce effectively antibiotic prescriptions [22, 23, 26, 27, 30].

We observed large differences in Abx prescription rates between countries ranging from 13.5 and 80% and within a country [21, 23]. There were five trials from the USA [15, 23‐25, 29] where pre-intervention Abx prescription rates ranged from 24 to 80% [23, 29] and post-intervention Abx prescription rates varied from 29 to 47% [15, 24, 25]. There were two studies from the Netherlands with post-intervention Abx prescription rates ranging from 23 to 65% [22, 30]. Two studies from Spain showed post-intervention Abx prescription rates between 24 and 64% [13, 31]. Pre-intervention Abx prescription rates in the trial of Bjerrum et al. were only available for the IG [13]. These large variations within and between countries limit the generalisability of the findings and indicate high possibility of selection bias and regional factors affecting Abx prescription rates.

Five included trials did not demonstrate any reduced Abx prescription rates [21, 24, 25, 32, 33]. Possible reasons for lack of success were possible selection bias in the recruitment of physicians who were already low prescribers [21, 33], low intervention uptake or insufficient implementation [24, 25] as well as lack of power due to low number of participating physicians and patients [32]. We consider five trials reporting statistically significant results as ineffective [14, 15, 20, 28, 29]. Gulliford et al. used a CDSS in a large sample of family practices and reported a small difference of 1.85% (95% CI − 0.1 to 3.59) [28]. IG and CG had a similar baseline prescription rate of roughly 50% reflecting overprescribing. The reported statistical significance of the small observed effect is due to the large sample size and cannot be regarded as efficient (difference in differences: − 1%).

Altiner et al. implemented CST, patient brochures and a waiting room poster in the IG [20]. Despite a large difference in baseline prescription rates and increased Abx prescription rates in the IG (+ 0.3%) and CG (+ 10.1%) within 1 year after baseline, this trial reported statistically relevant reductions after adjusting for seasonal effects and confounding variables such as severity of disease (IG: adjusted OR = 0.72, 95% CI 0.54–0.97, p = 0.028; CG: adjusted OR = 1.31, 95% CI 1.01–1.71, p = 0.044). This reduction does not satisfy our conditions for a meaningful change (difference in differences − 9.8%).

In the trial of Gjelstad et al., the combination of knowledge transfer, DP and feedback on prescribing resulted in a small difference of − 4.1% between IG and CG [14]. Due to the large sample size, this trial reported statistically relevant reductions but cannot be regarded as a meaningful change (difference in differences + 0.2%).

Meeker et al. investigated three interventions separately and in combination within a factorial study design [29]. Although two of the interventions showed a statistically significant and impressive reduction of Abx prescriptions of 17%, this has to be interpreted given that a reduction of 13% in Abx prescriptions was also observed in the CG. An explanation for a reduction without intervention was not given, beside Hawthorne effect. This observation points out the importance of an independent CG and pre- and post-intervention measurements of prescribing rates. Therefore, all trials shown in Fig. 4 lacking pre-intervention measurements have to be interpreted cautiously. For example, McGinn et al. reported a trial with a difference of 9% in Abx prescription rates compared to the CG, just below our arbitrary threshold for minimal significance of 10% [15]. However, for both groups, no baseline prescription rates are available.

Our review suggests that, in countries with relatively low prescription rates like Germany or the Netherlands, CST seems to be the key element for successful interventions. In contrast, interventions focusing on making a diagnosis in terms of POCT and CDSS showed relevant reductions in high-prescribing countries (e.g. Spain, USA, Russia) [13, 23, 27]. The role of electronic decision support systems remains unclear.

This review also adds to our knowledge that interventions aiming at reducing inappropriate antibiotic prescribing can have a positive effect on the prescribing quality. In a number of trials, the number of narrow-spectrum penicillins increased [13, 14], whereas the proportion of broad-spectrum antibiotics declined [13‐15], although the main focus of the interventions was to not prescribe. Effects on patients’ satisfaction were reported in three trials [21, 22, 30]. Only one trial reported a significantly higher proportion of patients satisfied with care in the IG due to CRP-POCT (p = 0.03).

Our systematic literature search was limited to few databases and hand search of references due to lack of access to other databases and funding. Additionally, publication bias and the possible exclusion of some foreign language trials have to be acknowledged. Although we cannot exclude that we have missed few trials, we believe this would not have changed our conclusions or allowed summary statistics given the heterogeneity of the designs and outcome measures.

All included RCTs differed in study design, data collection and time points of measurement, trial quality and baseline prescribing rates. Reporting of trial data was often poor due to missing p values, confidence intervals, absolute number of prescriptions and/or baseline data (Tables 1 and 2). Trials with high risk of bias may have led to a too positive interpretation of reported results. The heterogeneity of trial designs and outcome measurements made it impossible to pool trial data or compare effect sizes (e.g. using Cohen’s d) between trials. Alternatively, we have summarised the baseline and post-intervention Abx prescription rates in Figs. 3 and 4 to illustrate the differences and heterogeneity in between trials. There is no consensus about the effect size on Abx prescription rates considered as minimal important change. Our arbitrary assumptions considering an absolute 10% change as minimal important is based on the impression gained from these figures. The majority of trials did not adjust or balance seasonal effects (winter vs. summer), possibly affecting Abx prescription rates.

CST and POCT alone or in combination have the potential to reduce antibiotic prescriptions for RTIs. Electronic decision support tools showed only mixed results. Eleven out of 17 trials were not successful in reducing Abx prescription rates according to our definition of minimal important change [13‐15, 20, 21, 24, 25, 28, 29, 32, 33]. However, six of them reported a statistically significant reduction [13‐15, 20, 28, 29]. Trials with low initial Abx prescription rates were less likely to be successful. Despite a number of noteworthy current studies, the generated evidence remains disappointingly limited. Only moderate evidence which interventional strategies are successful and how these findings could be generalised beyond the actual setting and the observational period of the trial exist.

We conclude that there is a need to develop a consensus for designing and reporting of trials aiming to reduce inappropriate Abx prescriptions in the near future. It should address (among others) the measurement of pre-intervention prescribing rates, adjustment for seasonal and temporal trends, (minimal) follow-up time, data analysis and reporting.