Electronic health record-based behaviour change interventions aimed at general practitioners in the UK: a mixed methods systematic review using behaviour change theory

Search strategy

Databases were searched for quantitative and qualitative articles based on three concepts:

1. EHR systems,

2. General practice/practitioners,

3. UK.

MEDLINE, EMBASE, CENTRAL and APA PsycINFO were searched up to March 2023 (online supplemental material 1). Quantitative studies were limited to UK-based randomised controlled trials (RCTs), controlled before-and-after studies and interrupted time-series (ITS) analysis. EHR interventions needed to be at the point-of-care and integrated into the EHR system with the aim of influencing the behaviour of the GP. Outcomes had to capture the impact of the EHR intervention. Qualitative studies were limited to those which were part of controlled experimental or ITS studies. Studies were excluded if interventions were not based/supported in the EHR system; were complex interventions, for example, the EHR-based element was part of a wider intervention but not evaluated separately; were targeted at several types of people (eg, secondary and primary care) or only measured outcomes at patient level. Studies to develop EHR-based tools to change clinicians’ behaviour but not used or evaluated in a controlled study were also excluded, alongside observational studies (cross-sectional, case–control, cohort) and conference abstracts, opinion pieces, protocols, unpublished/not peer-reviewed articles.

The Preferred Reporting Items for Systematic Reviews and Meta-Analyses checklist was followed12 and the review was registered with PROSPERO.

Study selection and data extraction

Title and abstract screening and full-text review were conducted independently by two reviewers (JS and LP). Any discrepancies were resolved through discussion. Reference and citation searching of included articles was undertaken.

For all included studies, the Template for Intervention Description and Replication (TIDiER) checklist was used to aid data extraction.13 Where needed, study authors were contacted to complete the TIDieR checklist. Where available, study protocols or online supplemental information were used to improve understanding of the intervention (online supplemental material 2). Full details of data extracted are found in online supplemental material 3.

Interventions were categorised by behaviour change intervention using two recognised theory-based frameworks: (1) an adapted version of the UK government’s Behavioural Insights Team MINDSPACE framework, focusing on cognitive biases14; (2) the Behaviour Change Wheel’s (BCW) intervention functions.15 Both were chosen because of their applicability to EHR-based behaviour change interventions and because they have been successfully applied to develop and evaluate other behaviour change interventions in clinical settings.16 17 By using both frameworks, we also aimed to compare how each could be applied to EHR-based interventions and understand how this aligned with previous mapping exercises.18 Both frameworks and examples of how they can be applied to EHR-based interventions are presented in table 1.

Critical appraisal

For quantitative studies, the online Cochrane risk-of-bias tool for randomised trials and cluster randomised trials (RoB 2) was used.19 For qualitative studies, the Critical Appraisal Skills Programme Qualitative Studies Checklist was used.20 Critical appraisal was carried out independently by two authors (JS, LP) with any disagreement resolved through discussion.

Data synthesis

We used a mixed methods synthesis based on a ‘convergent segregated approach’, whereby quantitative and qualitative syntheses were conducted separately, and results synthesised to provide an overall argument.21

Quantitative synthesis was based on elements of Cochrane’s Synthesis without Meta-analysis (SWiM) approach22 (online supplemental material 4). Specific characteristics of studies to explain variation in results included quality, interventions used, behaviour change approach, and study design. Inductive thematic synthesis was undertaken for qualitative studies.23

Quantitative syntheses used vote counting based on direction of effect as the heterogeneity of outcomes precluded meta-analysis and comparison of magnitude of effect. The results of the primary and secondary outcomes of each study were recategorised into outcome domains depending on whether they measured a change in GP or patient behaviour; this separation was chosen because of the logic model behind behaviour change interventions aimed at clinicians, that is, clinician behaviour has to change before more distal factors can be affected, and because clinical changes in patients may be more susceptible to confounding factors. The results in the outcome domains were then transformed into positive, mixed or negative effects.24 Vote counting was not based on significance of result as recommended by Cochrane guidance.25 For completeness, however, this information was captured and is available in online supplemental material 5. An effect direction plot was used to summarise results and a harvest plot was generated based on the effectiveness metric, study quality and the behaviour change category.

Study characteristics

Table 2 lists the characteristics of included studies. Publication of articles ranged from 2002 to 2021. Data collection for all studies took place seven or more years prior to March 2023 when the literature was searched. Six quantitative studies were cluster RCTs at practice level, two were randomised at patient level. The number of GP practices in trials ranged from 15 to 104, with a total of 412 and a mean of 52. Intervention duration and follow-up varied from 18 weeks to 24 months. There was large variation in outcome measures, from antibiotic prescribing rates to changes in patients’ blood pressure. Qualitative studies involved semistructured interviews; a total of 27 GPs were interviewed, about the usability, acceptability and implementation of interventions during the associated trial. No other clinical staff were interviewed.

Critical appraisal

Quantitative studies were of mixed quality; four were judged low risk of bias and four had some concerns (online supplemental files 6, 7). Qualitative studies were of high quality (online supplemental file 8).

Intervention characteristics

Overall, three areas of clinical activity were targeted: reducing overtreatment, encouraging review of patients and controlling at-risk patients. These covered various clinical domains including antibiotic prescribing, cardiovascular disease, respiratory disease and neurocognitive disorders.

Table 3 lists the behaviour change approaches as we classified them using the MINDSPACE and BCW categories. The studies used five of nine MINDSPACE categories: increasing salience, using affect, increasing commitment, defaults and the messenger. No interventions fitted under the MINDSPACE categories of ego, norms, incentives or priming. Five of nine BCW intervention functions were used: environmental restructuring, education, persuasion, training and enablement. No interventions were classed under the BCW’s modelling, incentivisation, restriction or coercion intervention functions. The number of intervention types identified per study varied from one to five.

All interventions used automated on-screen prompts or pop-ups that were categorised as a form of salience from MINDSPACE because it caught GPs’ attention; in the BCW, this was deemed a type of environmental restructuring because the normal clinician workflow was modified. As part of this, some interventions (five of eight) highlighted patients at-risk of hospitalisation/overprescribing or in need of a health check, which was considered a form of affect (MINDSPACE) by appealing to clinicians’ emotions as they would not want to be responsible for a deterioration in patients’ health. The use of the word ‘free’ was also considered to utilise affect, rather than an incentive as there was no direct benefit/loss for the clinician. These interventions were categorised as persuasion in the BCW, alongside studies using a persistently appearing prompt. One study was categorised under the messenger (MINDSPACE) because it used a university-led study banner. Defaults (MINDSPACE) in the form of templates requiring data entry were used in addition to prompts in four studies. Three studies were classed as having an education (BCW) element, by providing information on risk and what to do. Training (BCW) was used in three studies via videos to introduce the intervention, and enablement (BCW) used in four studies as recommendations presented to the clinician to increase the possibility of action. Only one study used commitment (MINDSPACE) by requiring GPs to indicate their intended management.

Quantitative study findings

Considering outcome domains by clinician or patient change, four of seven studies measuring clinician outcomes found a positive direction of effect, the remaining three were mixed (table 4, full details in online supplemental material 5). Of the six studies measuring patient outcomes, three were positive26–28 and three were mixed.29–31

Only one study had mixed results in both clinician and patient outcomes aiming to reduce blood pressure and cholesterol and influence prescribing based on guidelines,30 although the authors’ primary outcome was positive. Two studies showed mixed results in the clinician domain and for their primary outcome, failing to show increased adherence to guidelines32 33; neither measured patient outcomes.

Qualitative study findings

Qualitative findings are synthesised under the themes of limitations and strengths of interventions from the GPs’ perspective.

Limitations of EHR-based interventions

Strengths of EHR-based interventions

Synthesis of quantitative and qualitative findings

Quantitative results indicated that EHR-based behaviour change interventions for UK GPs may have some overall effect on clinician behaviour, but the effect on patient-level outcomes is mixed. Changes in GP behaviour were supported by qualitative findings as GPs described how interventions prompted them to review patients and make positive changes to their care. However, contradictory findings were noted as clinicians described wanting to have guidelines and information available to them via the EHR, whereas some findings suggested these were underused and not fully engaged with. Only three of eight studies reported on intervention usage and those that did, reported low use.26 30 33 This may reflect the limited time in consultation, the clinician prioritising the patient over the computer or a mismatch between what clinicians say they like/want and are able to do in practice.34

Effect of behaviour change intervention characteristics

Complexity of intervention

Interventions varied in their level of complexity. Simple interventions only required one action with limited options; extensive ones had various options/scenarios and included multiple tools (guidelines, recommendations, GP/patient information, etc). Those that had mixed effects provided extensive information requiring an active, although not compulsory, decision-making process to use it. This was illustrated by the two interventions that provided links to guidelines that needed to be opened and read but which failed to achieve positive results in the clinician outcome domain,30 33 whereas effective interventions focused on prompts that were harder to ignore, more succinct and required less cognitive work from the clinician. Reminders that persisted until actively dismissed achieved positive direction of effect in the clinician outcome domain.28 29 31 Three of the four more extensive interventions failed to influence clinician outcomes.30 32 33

Figure 2 presents a harvest plot of results and behaviour change interventions.

Figure 2

Harvest plot of results.

Effect of quality and study design

Quality varied between studies with positive effect and those with mixed effect; there did not appear to be an association. Six studies were practice-level cluster RCTs, two were patient-level RCTs; the effect of study design did not explain results.

Summary

This is the first systematic review of controlled intervention studies and associated qualitative evidence on the effectiveness of any point-of-care EHR-based behaviour change interventions aimed at UK GPs. Despite widespread use of EHR systems and associated EHR-based ‘nudges’, our review found a small evidence base (eight studies) where the direct effect of the EHR intervention was measured. All data collection took place over seven years ago and the oldest study was from 2002 when EHR systems were markedly different.

EHR-based interventions were found to have a positive or mixed direction of effect and clinician-targeted outcomes were marginally more likely to capture positive change compared with patient-level outcomes. Studies’ short follow-up period, distal outcomes or insufficient statistical power may have contributed to this weakness.

The use of affect when clinicians were presented with ‘at-risk’ patients was a common feature of interventions with positive effect. No studies were found which used interventions that fitted under the MINDSPACE categories of ego, norms, priming or incentives suggesting areas for future research (see table 1). Like the MINDSPACE category of affect, studies that highlighted ‘at-risk’ patients as well as those that used persistent pop-ups fitted under the persuasion category of the BCW intervention function as they used communication to induce positive (or negative) feelings to stimulate action. The BCW intervention function categories provided more scope to include different aspects of interventions, because the categories were broader and open to interpretation, whereas MINDSPACE was based on specific human behaviours, which did not suit the lack of detail provided in the studies regarding the intervention design.

Simple, persistent pop-up reminders with clear instructions requiring action seemed more effective compared with multifaceted, guideline-based interventions, the exception being Gulliford et al, although the intervention was underutilised. Effect was seen more often in specific process activities, for example, recording (dementia, blood pressure) or prescribing (antibiotics, oral anticoagulants), compared with concordance with recommended clinical management, which often was more complex and also required patient behaviour change.

Intervention design

There was a lack of behaviour change theory in the description of the intervention design and analysis of findings. Most interventions were basic and based on ‘prompts’ without explanation or justification for the design (eg, wording, colour, position on screen, size, timing, etc). There was a notable lack of detail surrounding logic models to enable an understanding of the intervention’s anticipated mechanisms of action.

Outcomes

Some authors were ambitious with primary outcomes (eg, cardiovascular disease events), which would likely require a larger population or longer follow-up to observe an effect. Others were unclear which outcomes they defined as primary or secondary. A large range of outcomes were used within some studies; some attempted to analyse a complex group of indicators to see some effect33 or assess ‘concordance with guidelines’, which was challenging to assess.32

Authors did not always state which direction of behaviour change was favourable in their methods, as this can depend on the baseline relative to good practice and the individual patient.29–31 For example, it was unclear if authors viewed a reduction in prescribing short-acting beta-agonists or theophylline as a sign of improved management,31 or whether an increase or fall in consultation rate indicated improved quality.33 Most studies did not question the quality of coding and data within the EHR, which was used to generate the EHR-based prompt, or the GPs’ views on this which may have affected the reliability of the intervention and the willingness of the GP to respond.

Limitations

The lack of evidence found may be a function of the inclusion criteria, which only considered controlled and ITS studies, point-of-care EHR-based behaviour change interventions and studies where the effect of the EHR intervention could be measured separately to other concurrent interventions. Notably, we did not find any studies that assessed EHR-based interventions linked to a financial incentive, such as the Quality and Outcomes Framework, despite financially driven EHR prompts being widespread in the UK.36 This review highlights the lack of published evidence in this important area, which would be suited to rapid roll out of RCTs via the EHR, with the potential to explore the effect of financially linked EHR-based interventions compared with EHR-based interventions using other forms of behaviour change theory. Future research could investigate the role of EHR-based behaviour change interventions on other clinicians in primary care and consider other forms of evidence, including stand-alone qualitative studies or grey literature such as reports from EHR providers.

The assignment of behaviour change approaches to interventions was subject to the review authors’ interpretation and details provided of the interventions’ design were limited. This made it challenging to categorise interventions into the various dimensions of MINDSPACE and BCW intervention functions, and, therefore, some of the interpretations may be considered subjective.

Comparison with existing literature

Existing literature surrounding EHR-based interventions often refers to ‘computerised decision support systems’ (CDSS).37 Our review found small positive effect sizes; other systematic reviews of on-screen point-of-care reminders found similarly modest effects or that reminder ‘nudges’ did not lead to significant changes in clinician behaviour.38–40 The finding that simple prompts requiring action were reasonably effective is supported by a study which found the most successful interventions were those requiring clinicians to provide a reason to over-ride them41; another review also noted that CDSS were more effective when automatically displayed on-screen, which lends itself to EHR-based intervention delivery.42 Our review found that effect was generally seen on clinician-dependent processes of care compared with patient outcomes, which is supported by a 2020 CDSS meta-analysis, which noted improvements in processes of care43; another study on asthma and Chronic Obstructive Pulmonary Disease (COPD) tools found positive changes in management processes.8 Being realistic about what type of change can be achieved is also important. Two studies in our review with effective interventions aimed to change prescribing which is directly within the clinician’s control.26 29 Electronic decision support for antibiotic prescribing has been effective in the USA, demonstrating that EHR-based interventions are likely suited to the domain of prescribing,44 45 with prompts a common feature of other prescribing interventions.46

Systematic reviews exploring interventions targeting clinicians have found peer group norms and use of social reference points to be effective drivers of behaviour change.44 47 48 These previous reviews included clinicians outside of general practice and settings outside the UK. This contrasts with our review which found no evidence of norms (MINDSPACE) being used and evaluated. Therefore, future research should evaluate incorporating norms into EHR-based interventions in UK general practice at the point-of-care, ideally including other general practice clinicians, for example, nurses, pharmacists and physician associates. For example, prompts drawing on local clinicians’ behaviour could refer to desirable norms.

The same BCW group of intervention functions found in this review were found in a primary care deprescribing scoping review where environmental restructuring was most used followed by persuasion and enablement, but not coercion and restriction.49

The qualitative finding that EHR-based interventions provide reassurance to clinicians and/or patients is reflected in a recent review exploring patient views of CDSS in asthma, which found that such tools can help with shared decision-making.50 Another review found that coproducing tools/interventions with clinicians may aid their uptake51; future studies should consider this. Furthermore, the finding of clinical judgement taking precedent over what the computer says is reflected in a qualitative review, which concluded that clinicians are likely to go with their ‘gut instinct’ unless the tool is carefully developed.9

Implications for research

When developing EHR-based interventions, researchers should consider the theory and logic model underpinning the intervention.52 How more nuanced design choices, such as the wording, colours, associated images, timing of messages and the use of different aspects of behaviour change theory should be considered. Evidence from the use of computer-based nudges in other settings can inform this.53 Researchers should ensure outcomes are clear and feasible within the study’s timescale. Future research could include quasi-experimental trials with the existing EHR systems.

Implications for policy and practice

This review found few controlled experimental studies, most of which would be considered old given the rapidly changing digital landscape. Prompts seem to be a simple, cost-effective solution to potentially influence clinicians’ behaviour—in particular, when these are brief, persistent and require action for at-risk patients, rather than providing extensive information during busy consultations. Similarly, EHR-based interventions should be as integrated as possible within the existing EHR, recognising that links to external sources will likely be underutilised. Policymakers, researchers and clinicians should endeavour to work with software providers to codesign, test and optimise EHR systems to positively influence clinician behaviour for the benefit of patients. Importantly, the findings of such research should be transparent and openly available as the role of EHRs expands both in the UK and elsewhere.