Inclusion criteria

The target population (P) was adult patients in any setting (e.g. hospital, primary care, or community). Eligible interventions (I) were prescribing or deprescribing of antithrombotics (i.e. anticoagulants or antiplatelet agents) administered by any route for primary or secondary prevention of either thromboembolic events or management of bleeding complications. Comparators (C) were not always included as included studies were trials and beyond. Outcome (O) of interest was objectively measured adherence against explicitly stated and referenced guidelines whether international, national or regional. Articles were eligible if they reported at least one of the indicators (overuse or underuse) where objective outcomes were measured in quantifiable ways. In intervention studies that aimed to increase guideline adherence, we did not limit to those who used STOPP/START criteria but accepted any method of investigating the outcome in relation to the guideline, and only reported baseline outcome data that reflected usual practice. Eligible quantitative study (S) designs included: retrospective record reviews, prospective cohorts, interventional studies whether samples were randomly assigned or not, time series, before-after studies, cross-sectional surveys or audits, and secondary analyses of disease registries.

Exclusion criteria

Excluded were articles using a guideline non-adherence definition which combined both overuse and underuse in a single estimate; having self-reported measures (e.g. clinician surveys of perceived guideline adherence, patient- reported doctor recommendations); failing to mention reference guideline or using internally developed but unpublished guideline. Conference abstracts without sufficient information to assess risk of bias, case studies, study protocols without data, qualitative consultations of perceived inappropriateness, commentaries, and editorial pieces were ineligible.

Search strategy

Our search strategy targeted published English language literature from Medline and EMBASE databases from January 2000 to May 2021 to reflect recent and current practice involving new generation oral anticoagulants. The umbrella search strategy included the terms ‘guideline,’ ‘adherence’ and ‘prescribing’ among others (details in Supplement 1, Table S1.1), was designed jointly by the team based on clinical experience, and subsequently refined by our information specialist JC using polyglot tool [16], word frequency counter, and de-duplicator from the internally developed and tested systematic review accelerator [17].

Screening and data extraction

Subsets of title, abstract and full text screening was conducted by paired authors independently (LC, MC, OB, LH, MO, TH) with discordances resolved by discussion using the disputatron tool [17]. A purpose-built data extraction template including author, publication year, country, sample size, study design, setting, target conditions/guideline topic, and study population was used. Estimates of overuse and underuse were extracted by paired authors (LC, MC, LH, OB) either from the text or tables, and the accuracy of data extraction double checked and confirmed by the statistician (MJ) who subsequently built forest plots. Drivers of overuse and underuse as well as potential solutions were extracted by the lead author (MC) from the discussion sections of eligible papers to enhance the context of our discussion.

Outcome definitions

Our primary outcome was the estimate of percentage overuse or underuse or equivalent terms (excess, under-adherence, underutilisation, overprescribing, non-compliance -if stated the direction, etc). Overuse was defined as follows: prescribing when not clinically indicated (e.g., CHADS₂ score = 0), excess dose/duration, or inappropriate or unnecessary administration route (e.g., IV or subcutaneous medication when an oral alternative would have sufficed). Underuse was defined as non-prescribing when recommended by the guideline such as CHADS₂ score ≥ 2, or lower than recommended dose or shorter duration according to the guidelines used as gold standard. Outcomes were expressed as a percentage of patients managed with or without anti-thrombotics out of eligible patients for prescribing or non-prescribing.

Data synthesis

Estimates of overuse and underuse are presented, where possible, as reported by the published authors, predominantly as mean percentages with 95% confidence intervals. When this information was not reported but absolute numbers for numerators and denominators were presented in the text or tables, we calculated mean percentages with 95% confidence intervals. If the study included an intervention, only pre-intervention data were reported. Given the heterogeneity of study designs, populations, guidelines and outcome measurements, subgroup analysis by setting, region or clinical indication was limited to narrative description. No attempt was made to impute missing data or conduct sensitivity analyses. Forest plots were produced separately for overuse and underuse for ease of visualization, but no meta-analysis was undertaken. Reported findings followed the synthesis without meta-analysis guideline [18].

Risk of bias

Given the anticipated inclusion of trials, cross-sectional and cohort studies, risk of bias was based on Hoy’s published 10-item tool for assessing prevalence studies [19]. We modified the version by using the checklist with the exception of items 5 and 9 which were not directly relevant to our research question.

Paired reviewers (LC, MC, OB, LH, MO) independently assessed each study completing the template with six key questions for all eligible study designs (Supplement 1, Table S1.2). Two senior authors (TH, MC) assisted in resolving risk of bias discrepancies. Trained academics not members of the overuse research team (LA, MB, EL in acknowledgment) also assisted occasionally in the resolution of risk of bias if agreement could not be achieved and senior reviewers were unavailable.

Study selection

Twenty-one contemporary studies (published between 2008 and 2021) met the inclusion criteria. They were conducted in 14 countries across four continents (Europe/UK, n = 11, North America, n = 5, Australia n = 3, and Asia, n = 2) and involved a total of 167,287 participants. Figure 1, the PRISMA diagram shows details of the selection process).

Study characteristics

Study designs included 14 retrospective record reviews [20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33], 6 cohort studies [34, 35, 36, 37, 38, 39] and 1 a case-control [40] (see Table 1). Four studies focused on overuse, 5on underuse, and 12 on both. Information derived predominantly (n = 18) from hospitalized patients or disease registries, with only three studies in primary care. The clinical context included preventive use in atrial fibrillation (AF, n = 8), perioperative bridging (n = 5) and treatment of various conditions (stroke, venous thrombosis, myocardial infarction, excess anticoagulation, trauma management; n = 8).

Risk of bias assessment

Across the studies, risk of bias was generally low in terms of the inclusion criteria matching the guideline’s target population, provision of a clear and acceptable case definition, and objective data extraction methods (Figure 2).

However, in some studies, there was a high risk of selection bias from use of either convenience sub-samples, or unclearly presented selection methods where not all potentially eligible patients were included, incomplete response or final outcome ascertainment of ≥20%, and non-inclusion of full census, consecutive patients, or random subject selection. Six studies did not report the extent of missing follow-up data. No selective reporting was observed as most studies measured a single primary outcome.

Overuse estimates

Large variation in the rates of antithrombotic overuse was observed across geographic regions with studies from Asia reporting the lowest levels of overuse (1.3–7.6%) [27, 33], followed by Australia (24.1–42.9%) [20, 23, 24], and Europe (1.4–72.0%) [21, 25, 28, 29, 30, 33, 36, 37, 38, 39, 40], through to the highest variation in North America (6.6–73.0%) [26, 31, 32, 34, 35]. No clear time trend of increasing or decreasing overuse was observed over the period of the studies (2008–2020). Estimates in 10 (of 17) studies had narrow confidence intervals due to moderate to large sample sizes (Figure 3).

The most commonly reported indication of antithrombotic overuse was in prophylaxis for people with atrial fibrillation (AF) at low risk of stroke or other thrombotic events. As shown in Figure 3, overuse ranged from a low of 1.3% for stroke prevention in a Japanese cohort with AF [33] and a 1.4% in a Polish study [25], to a high of 72.2% prophylactic use in non-valvular AF in Romania [22]. Overall, median overuse for this indication was 24% (IQR 5.0%–42%).

The retrospective studies of AF patients conducted in primary care or using community registries tended to report the lowest levels of overuse in this collection, such as 1.3% in Japan [33], 4.0% in Canada [32], 6% in the largest sample (n = 70,498) in the US [31], and 13.1% in the UK [33]. The exception was a case-control study of AF patients within six months of diagnosis in multiple Belgian general practices which found overprescribing of 57.9% [40].

There was a wide range in overuse in venous thromboembolism (VTE) treatment in hospitals, from 7.6% in a study in Iran [27] to 73% in a US study of patients admitted for VTE treatment contrary to guideline recommendation to manage them as ambulatory patients [26].

Among the three studies of perioperative bridging anticoagulation reporting overuse, estimates varied from 21% in a Polish cohort of patients with AF undergoing percutaneous coronary intervention [39], through to 34% in a retrospective review of Dutch patients receiving Vit K antagonists before acute or elective surgeries [21], and up to 59.6% of French patients in people undergoing elective or emergency procedures [29]. A North American cohort study of patients presenting with prolonged prothrombin time INR ≥4.5 reported as 22.3% due to unjustified administration route or excess dose [35]. Median overuse across these studies was 34% (IQR 27.5 – 46.8%).

Underuse Estimates

Variation in underuse of appropriate anticoagulation was observed across clinical indications with estimates of greater than 30% in 6 of the 17 studies (Figure 4) [21, 24, 27, 33, 37, 40]. The highest rate of underuse was in adults with AF at high risk of stroke who were not prescribed anticoagulant therapies in a tertiary referral hospital, with similar underuse rates in both Indigenous Australians (72%) and non-Indigenous patients (69%) [24].

Substantial underuse of prophylactic anticoagulants in AF was also reported (51.4%) by a Belgian case-control study in primary care [40], as well as the AF community registries in Japan (40%) [33] and the UK (36%) [33]. Moderate estimates of prophylactic underprescribing (27% and 21%) were derived from retrospective record reviews in Australia and Romania respectively [22, 23]. The lowest estimates of underuse for this indication were found by two retrospective chart audits in Poland hospitals (9.0%) [25] and primary care in Canada (10%) [32]. Overall, median prophylactic underuse in AF was estimated at 36.0% (IQR 21.0–51.4%).

Underuse of VTE treatment for people with clinical indication was reported in over a third (36%) of Iranian patients admitted to a teaching hospital [27]. Two recent hospital studies reported the extent of antithrombotic undertreatment for management of transient ischaemic attack (28%) or acute myocardial infarction (44%) in Italy and Norway respectively [37, 41].

Underuse of bridging coagulation in the perioperative period was found to be present in 33.0% cases in a Canadian hospital cohort [34] and wide variation between 9.0 and 50.0% was found in Europe across three retrospective hospital audits [21, 29, 30] and one hospital cohort [39]. A US cohort of patients presenting with excess coagulation found 6.2% under-administration of vitamin K [35], and a retrospective analysis of management or traumatic injuries to prevent coagulopathy and thromboembolic events in a French hospital reported 14.3% underprescribing of thrombolytic agents [28]. Median underuse in perioperative bridging across the five studies was 25.0% (IQR 13.1 – 32.0%). While no time trend was visually apparent, underuse estimates from five retrospective chart reviews [22, 25, 26, 30, 32] and four cohort studies [34, 36, 38, 39] since 2020 in Europe and North America have remained below 33% (9–32%) by comparison with 36%–72% reported between 2015 and 2019, but we did not formally test this.

Determinants of guideline non-adherence (underuse and overuse)

The reasons for underuse were generally not mentioned in our identified articles. Among the one third of included articles reporting reasons, it was clear that justifications for underuse were largely clinician-driven by fear of complications or insufficient knowledge of specific guidelines. By contrast, most of the determinants of overuse when reported were system related, with minority attributed to patient preference (Table 2).

None of the studies specifically examined or reported the influence of patient preferences for anticoagulant type. However, a study comparing patient and clinician preferences for anticoagulation found difference in preferences: the minimum number of strokes that needed to be prevented to justify anticoagulation was significantly lower for patients than for physicians (1.8 versus 2.5 per 100 patient years), and the maximum number of excess bleeds that was acceptable was significantly higher for patients than for physicians [42]. No study reported using decision aids to inform the risk of bleeding events versus the benefits of stroke prevention.

Potential solutions

Brief solutions listed in some included studies without further elaboration and other relevant recommendation from different specialties are shown in Table 3. Addressing overuse, underuse and practice variation is likely to need strategies at the patient, clinician and system levels. Some voluntary or mandatory adoption policies may play a role in reducing variation but the proliferation of guidelines from different groups and updated recommendations for treatment across countries such as in the case of antithrombotics for AF may generate confusion for clinicians managing different at-risk populations [43].

It can be argued that guideline non-adherence can be caused by the knowledge that some guidelines are flawed if based on expert consensus rather than on high level of evidence [52]; or even when trial evidence exists the recommendations only apply to a limited type of patients [53]. The single-disease guideline focus may also contribute to non-adherence as they may not apply to patients with multiple comorbidities [54]. In sum, estimates of overuse and underuse based on comparison against static practice guidelines with data collections not linked to reasons for clinical decisions can only provide an indication, and may be prone to error rates in both directions.

Comparisons with other published studies

Other literature reviews that were either non-systematic or conducted on single medications reported much higher rates (40.0%–85.6%) of under-prescribing of anticoagulants for AF internationally [9, 55]. Similarly, a large cohort study conducted in 28 countries [56] reported variation in prevalence of initial therapy for and undertreatment of VTE at 46.5% during pregnancy. This study did not meet our eligibility criteria because the comparator for four continents was the USA guidelines, which renders some of the overuse or underuse estimates invalid for other national contexts in light of different risk stratification approaches and local management guidelines [57].

While a previous review reported increasing overuse of multiple practices over the years [58], the estimates in our study did not indicate a clear trend in guideline adherence improvement or worsening overtime, but we did not formally examine this. These mixed results across all years could be due to clinician’s deliberate decisions (i.e. discretionary care) [58], availability of medications and tests in different health systems, as well as changes in patient attitudes or clinician awareness of guideline updates [59].

Such inconclusive results of non-guideline concordant use of prescribing must be interpreted with caution as some studies spanned several years and guideline recommendations changed in that period, where uptake in routine care is known to take years. This may have led to re/misclassification of guideline adherence and non-adherence and made previously accepted practices non-concordant in more recent times. Further, some alarming overestimates were derived from small denominators, and other small estimates emerged from large registries or hospital chart reviews.

Limitations

Our search strategy was applied to only two major databases in English language and some studies may have been missed as a result. However, the search strategy underwent extensive review and iteration to ensure search output efficiency and accuracy. Our search strategy did not include the terms ‘omission’ or ‘commission’ that might have been used by others to indicate underuse. We also excluded studies using local or single institution guidelines and studies of single treatments or single conditions as we were interested in population-wide impact of guideline non-adherence of recognised guidelines. Our decision may have resulted in an underestimate; however, it makes these findings more generalisable at population level. Heterogeneity of study methods, sources and outcome definitions precluded meta-analysis. We chose guidelines as a generally accepted reference of appropriateness. Yet, we acknowledge that not all guidelines have been developed through a comprehensive review of the scientific evidence [13] and some of the recommendations do not apply to patients of all ages, with multiple underlying conditions [60], allow for clinical flexibility, or consider all organisational constraints in all settings. Further, recommendations in clinical guidelines also change overtime, and their publication can be delayed, leading to true adherence estimates not being reflected [61]. While guideline adherence was objectively measured from available data sources, the authors of included papers did not generally report how clinical decisions were made or how individual exemptions were taken into account. Since the data were largely not linked to reasons for clinical decisions, it cannot be unequivocally claimed that all reported cases of non-adherence were true overuse or underuse if justifiable exemptions existed. In sum, level of adherence to guidelines can only be considered a best surrogate indicator of medication safety [62].

Implications for practice and research

Overprescribing in the absence of a clinical indication, hospitalising for anticoagulation management when ambulatory treatment is considered safe, and underprescribing for patients with high-risk profile identified in this review can expose patients to further risk of complications such as bleeding or preventable thromboembolic events. While getting the balance right between too much and not enough takes expertise, guidelines are a best available reference point. Importantly, many factors may have contributed to the wide variations in practice found in this review: use of a reference guideline from a different country to the study setting; differences in availability of medicine types across health systems; co-payment schemes influencing prescribing patterns or patient preference; true unawareness of the guideline by prescribers; clinician’s personal biases, preferences or caution based on previous experience with adverse events; and the actual methodology of individual studies.

Our findings highlight some critical issues for attention of future guideline developers in this area, such as reporting of explicit exemptions. We were unable to find examples of introduction of machine learning algorithms for retrospective assessments of non-adherence using clinical exemptions in decision-making. The usefulness of integrated pharmacy and health service records in predicting future patient adherence [63] or estimating individualised optimal heparin doses [64] is promising. There are also high hopes for machine learning to contribute identification of statistical patterns of prescribing quality [65] but their effectiveness in identifying meaningful and credible conclusion on clinician’s guideline concordance and decision parameters are still under investigation [66]. So far prospective attempts to encourage and enhance adherence to guideline targets in cardiovascular disease management through computerised clinical decision making tools have not achieved substantial gains [67].

A gap identified in the included articles of this review were the impact of guideline non-adherence on health outcomes for patients, and the possible impact of using decision aids before antithrombotics prescribing on patient behaviour. Both were beyond scope, and future qualitative research could assist in improving understanding of the reasons for overuse or underuse by different stakeholders. Ascertainment of the reasons for unwarranted variations – whether clinician or patient driven – is important to inform the development of tailored de-implementation interventions. Future studies could investigate the economic impact of over/underprescribing on health service utilisation, and the avoidable cost of over-treatment and propose a lower/acceptable limit of what is ‘acceptable/unavoidable healthcare resource waste.’