What works to increase attendance for diabetic retinopathy screening? An evidence synthesis and economic analysis

Types of studies

We considered RCTs, both individually randomised and cluster RCTs, conducted in a primary or a secondary care setting. To investigate cost-effectiveness, we included full economic evaluations (cost-effectiveness analyses, cost–utility analyses and cost–benefit analyses), cost analyses and comparative resource utilisation studies conducted alongside or as part of an included RCT.

Types of participants

We included participants with type 1 and type 2 diabetes mellitus who were eligible for DRS.

Types of interventions

We included RCTs that used any planned strategy or combination of strategies to improve attendance for DRS targeted at individuals with diabetes, HCPs or the health-care system. Interventions included those specifically targeting DRS as well as those that were part of a general QI intervention for diabetes care. Comparator interventions were as specified in the included studies.

Types of outcome measures

Search methods for identifying studies

Study selection and data extraction

Two review authors (JGL and JMB) independently screened the titles and abstracts of studies identified in the electronic searches. We obtained full copies of research papers in cases of uncertainty and resolved any differences of opinion between review authors by discussion. Two review authors (JGL and EGR) worked independently to extract data from the included studies using a modified version of the EPOC Group data collection form,47which incorporates information on study design, type and duration of interventions, participants, setting, methods, outcomes and results. For the extraction of data on the sociodemographic characteristics of participants that are known to be important from an equity perspective, we used the PROGRESS (place, race, occupation, gender, religion, education, socioeconomic status, social status) framework48and also recorded whether or not any interventions were aimed at disadvantaged or low- and middle-income country populations (defined using theWorld Bank Atlas Method – Detailed Methodology49).

Studies judged to potentially include economic data were identified and further assessed by an economics reviewer (PA). Data from included economic evaluations were extracted by one reviewer (PA) and checked by a second reviewer. Data collection was adapted from the format and guidelines used to produce the structured abstracts of full economic evaluations for inclusion in the NHS EED,50which were redesigned to accommodate specific data required for the review. Economic evaluations were classified based on their analytical framework and were coded appropriately.

Coding of intervention content

We coded intervention descriptions from all of the included studies using a validated taxonomy to characterise the constituent components of each intervention. The Cochrane EPOC Group has developed a comprehensive taxonomy to classify interventions for use in systematic reviews. 43We used a subset of the EPOC taxonomy that had been previously used by members of the review team in a review of the effectiveness of general QI implementation strategies for diabetes care29(seeAppendix 4). This adapted taxonomy incorporates 12 components that target health-care systems (case management, team changes, electronic patient registry, facilitated relay of information to clinicians, continuous QI), clinicians (audit and feedback, clinician education, clinician reminders, financial incentives) or patients (patient education, promotion of self-management and reminder systems). Two review authors (JGL and EGR) independently coded QI components as ‘present’ or ‘absent’ for all intervention and control arms. Any discrepancies were resolved by discussion.

To better characterise intervention content we also coded BCTs within interventions using the BCTTv1. 36Describing an intervention in terms of BCTs (i.e. ‘active ingredients’) provides a useful level of detail for synthesis and comparison44(seeAppendix 5for a description of the BCTs identified in the included studies). We coded BCTs as ‘present’ or ‘absent’ separately for patient and HCP recipients. 44We contacted all authors of included studies to ask for further information on the content of the intervention (e.g. a trial protocol, letters sent to patients, written or audio-visual materials) to clarify the BCT coding. We coded these materials using the BCT taxonomy in the same manner as for the corresponding published reports. Two review authors independently conducted BCT coding (EGR and FL), resolving discrepancies by discussion and, when necessary, by the involvement of a third reviewer (JJF).

Resource requirement needed to deliver interventions

We developed an ordered ranking scale to quantify the level of resource needed to deliver each intervention, based on the description of the intervention components in each included study (seeChapter 6for further details of the methodology). To determine the feasibility of this approach, we initially piloted the scale on a sample of 10 included studies using two members of the review team. Each intervention was initially graded between 1 (least resource intensive) and 5 (most resource intensive), or as 0 (unable to determine), with a record of how the reviewer graded each study also provided.

An algorithm was developed to derive the ordered rank. This mapped resource components and their intensity to the ordered rank. The following resource components were incorporated into the algorithm:

• face-to-face minutes

• telephone calls

• patient home visits

• printed materials/software

• training.

We defined, a priori, a criterion for success of the ranking scale as reviewer scores from nine out of 10 studies being within one grade of each other following discussion. This criterion was achieved and the notes about how each study was graded were used to produce a reproducible description of the resource input associated with each grade on the ranking scale. The resource components and their intensity levels were then used to extract the resource use required to deliver the interventions in all included studies. This was conducted by two reviewers independently (JGL and EGR).

For the algorithm that mapped resource use onto the ordered ranking, the weights given to each resource component and their intensity were subsequently revised based on cost analyses (seeChapter 6). The revised grading system produced 18 ranks and these were recategorised into five ordered categories for the purposes of analysis.

Assessment of risk of bias

Two review authors (JGL and JMB) independently assessed study quality using the EPOC Group risk-of-bias tool,51which uses nine standard criteria.

1. Was the allocation sequence adequately generated?

2. Was the allocation adequately concealed?

3. Were baseline outcome measurements similar?

4. Were baseline characteristics similar?

5. Were incomplete outcome data adequately addressed?

6. Was knowledge of the allocated interventions adequately prevented during the study?

7. Was the study adequately protected against contamination?

8. Was the study free from selective outcome reporting?

9. Was the study free from other risks of bias?

For cluster RCTs we considered particular biases, including (1) recruitment bias, (2) baseline imbalance, (3) loss of clusters and (4) unit of analysis errors. For each domain, two review authors performed the risk-of-bias assessment independently and assigned a judgement of low risk, high risk or unclear risk of bias. The review authors resolved any discrepancies between them by discussion.

The reliability of the data outputs from any full economic evaluation are in part predicated on the reliability of the estimates of the relative treatment effects (for benefits or harms) of the alternative courses of action [i.e. intervention(s) and comparator(s)) under investigation]. As the identified economic studies were a subset of the studies included in the review, the risk of bias was already assessed. However, assessment of the overall methodological quality of the economic component was still required and was carried out by one reviewer (PA) using the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) statement52together with the Consensus on Health Economic Criteria (CHEC). 53In assessing the methodological quality of economic evaluations, the main objective is to assess the applicability of the scope of the analysis in terms of costs and outcomes. This helps to highlight the applicability and relevance of each economic evaluation. The checklists used differed from those that were originally included in our published Cochrane protocol because of the recent updates of the methods for the incorporation of economic evidence into Cochrane intervention reviews (seeReport Supplementary Material 1,Completed checklists for methodological quality assessment of economic evaluations, for the completed checklists for each included economic evaluation).

Data synthesis

Attendance at screening post intervention is a dichotomous outcome and our measure of intervention effect was the risk difference (RD), that is, the actual difference in the observed events between experimental and control interventions. For individual RCTs the unit of analysis was the individual participant. For cluster RCTs we analysed data after adjustment for clustering. In the case of cluster RCTs, when outcomes were presented at the patient level, we used an established method to adjust for clustering. 54This involved dividing the original sample size by the design effect, which was calculated from the average cluster size and the intraclass correlation coefficient (ICC). When the ICC was not reported, we imputed the most commonly reported value from studies in which it was reported. We contacted authors of included studies when important data were missing; we did not impute missing data if data were not available.

We conducted meta-analyses in Review Manager 5 (The Cochrane Collaboration, The Nordic Cochrane Centre, Copenhagen, Denmark) using random-effects models to estimate the pooled RD across studies. Data from patient RCTs and cluster RCTs that were adjusted for clustering were included in the same meta-analyses. In the case of multiple intervention groups, as recommended in theCochrane Handbook for Systematic Reviews of Interventions,55we combined groups to create a single pairwise comparison.

A summary of the results of the included economic evaluations is available inReport Supplementary Material 1[seeFurther details of the review of economic evidence (phase 1 review)]; this is supplemented by a narrative description in the results and discussion sections. Costs for each study were adjusted to 2016 UK pounds using a web-based conversion tool based on implicit price deflators for gross domestic product (GDP), a measure of the wealth of a country, and GDP purchasing power parities. 56The tables inReport Supplementary Material 1[seeCharacteristics of included, ongoing and excluded studies (phase 1 review)] present the original currency and price year used in each included study. Users of this review who might want to adjust costs to another currency and price year suitable for their needs should use the costs for each study presented inReport Supplementary Material 1and not the adjusted costs presented in the main text of this report.

Assessment of heterogeneity

We assessed heterogeneity between studies by visual inspection of forest plots and by formal statistical tests of heterogeneity (chi-squared test and theI²statistic). TheI²statistic is the proportion of variation between studies not due to chance and can take a value from 0% to 100%. Heterogeneity may be the result of variation in the ‘true’ effects underlying the studies but may also be the result of clinical diversity (participants, interventions, outcomes) and methodological diversity (varying degrees of bias).

We performed the following prespecified subgroup analyses to investigate whether or not the presence or absence of a particular covariant explained the variability in effect size:

• QI intervention components/BCTs

• resource requirements to deliver the intervention.

In our analyses, QI components (coded using the modified EPOC Group taxonomy) and BCTs of each intervention were assessed separately. When a study used multiple QI components and/or BCTs, the same effect size was applied to each component for the analysis. We compared effect estimates for subsets of studies that used a particular QI component/BCT or resource intensity and calculated a pooled effect size.

We further investigated associations between DRS attendance and effect size by meta-regression for a number of covariates including type of study design (individual/cluster RCT), baseline DRS attendance and QI component/BCT used in the intervention. For meta-regression we used a prespecified random-effects model and compared the RD of studies containing a particular explanatory variable with that of studies in which the variable was absent.

Subgroup analyses and meta-regression were conducted using Stata^®14 (StataCorp LP, College Station, TX, USA) using the metan and metareg commands. We performed a sensitivity analysis to determine the impact on the pooled effect estimate of imputing the lower and upper range values for the ICC.

Methods used to assess the quality of the evidence for outcomes included in the summary of findings tables

We assessed the quality of the evidence using the evidence grading system developed by the Grading of Recommendations Assessment, Development and Evaluation (GRADE) collaboration57and described in section 12.2 of theCochrane Handbook for Systematic Reviews of Interventions. 55One review author (JGL) initially applied the GRADE evidence rating system and discussed the rating with other members of the review team. A final decision was reached by discussion and consensus.

We took the following into account when deciding whether or not to downgrade the quality of evidence for each outcome:

• risk of bias

• inconsistency of results

• indirectness of evidence

• imprecision of results

• publication bias.

Study selection

The database searches yielded 7244 titles and abstracts. We identified a further 33 studies from additional sources. After removing duplicates we screened 7277 studies and reviewed 130 full-text articles. We excluded 64 studies with reasons [seeReport Supplementary Material 1,Characteristics of included, ongoing and excluded studies (phase 1 review)] and included 66 studies that met our inclusion criteria [Figure 4provides a Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram]. Of the included studies, 5058–107(75.8%) reported general QI interventions and evaluated the impact of the interventions across a range of outcomes, including DRS uptake; in 16 studies108–123(24.2%), the primary target of the intervention was to improve attendance for DRS. From the search of trials registers we identified nine ongoing studies [seeReport Supplementary Material 1,Characteristics of included, ongoing and excluded studies (phase 1 review)].

FIGURE 4.

Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram.

From the database searches we identified 22 references containing potential economic evaluations. Following full-text review, eight58,83,94,103,114,115,122,123of these references were excluded and 14 references59,62,67,69,70,85,87,89,95,105,117,124–126reporting economic outcomes were included in the review.

Characteristics of included studies

The included studies were conducted between 1988 and 2013. Thirty-five studies (53%) were parallel-group patient RCTs enrolling 237,025 patients59,61–63,71,72,75,80,82,85–89,95,96,98,99,102,103,107–109,111–122and 31 (47%) were cluster RCTs in which the HCP or the health-care setting was the unit of randomisation. 58,60,64–70,73,74,76–79,81,83,84,90–94,97,100,101,104–106,110,123These included 6126 clusters (range 6–4125 clusters). Fifty-nine studies (89%)58–63,65–78,80–93,95–105,107–116,118,120–122had two arms, six studies (9%)64,79,94,106,117,119had three arms and one study (2%)123had more than three arms. An overview of the included studies is provided inTable 1[for further details seeReport Supplementary Material 1,Characteristics of included, ongoing and excluded studies (phase 1 review)].

Intervention content in terms of quality improvement components (coded using the modified Effective Practice and Organisation of Care Group taxonomy)

Interventions either were specifically targeted at improving attendance for DRS (n = 16) or were part of a general QI intervention to improve diabetes care (n = 50). For studies comparing any intervention with usual care, the majority of studies provided no description of usual care, which precluded coding of the comparator arm. All 12 QI intervention components, as defined by the modified EPOC taxonomy, were used in at least one study. Generally, interventions were multifaceted with several QI components per intervention arm (median 3, range 1–7). For interventions specifically targeting DRS attendance, the most commonly used QI components were ‘patient reminders’ (56% of studies) and ‘patient education’ (75%) (Figure 5). For general QI interventions, a greater number and range of strategies were used, including ‘patient education’ (48% of studies), ‘promotion of self-management’ (40%), ‘case management’ (40%), ‘clinician education’ (38%) and ‘team changes’ (36%).

FIGURE 5.

Quality improvement components used in the intervention arms of included studies. GQI, general quality improvement. Reproduced from Lawrensonet al.,127with permission from John Wiley & Sons, Ltd. Copyright © The Cochrane Collaboration. Published by John Wiley & Sons, Ltd.

Reproduced from Lawrensonet al.,127with permission from John Wiley & Sons, Ltd. Copyright © The Cochrane Collaboration. Published by John Wiley & Sons, Ltd.

Intervention content in terms of behaviour change techniques (coded using the Behaviour Change Technique Taxonomy version 136)

Overall, 39 out of the possible 93 BCTs (42%) were identified as targeting change in patient or HCP behaviour in at least one trial. Interventions specifically targeting DRS primarily used techniques aimed at patients, particularly ‘instruction on how to perform the behaviour’ (75% of studies), ‘prompts/cues’ (69%) and ‘information about consequences’ (56%) (Figure 6). Relatively few of these studies used BCTs that were aimed at HCPs (Figure 7). By contrast, the following HCP-directed strategies were more widely used in general QI interventions, in particular ‘instruction on how to perform the behaviour’ (66%), ‘restructuring the social environment’ (52%) and ‘feedback on outcomes of behaviour’ (36%). Table 2provides definitions and illustrative quotations for each BCT.

FIGURE 6.

Behaviour change techniques targeting patients used in the intervention arms of included studies. GQI, general quality improvement. Reproduced from Lawrensonet al.,127with permission from John Wiley & Sons, Ltd. Copyright © The Cochrane Collaboration. Published by John Wiley & Sons, Ltd.

Reproduced from Lawrensonet al.,127with permission from John Wiley & Sons, Ltd. Copyright © The Cochrane Collaboration. Published by John Wiley & Sons, Ltd.

FIGURE 7.

Behaviour change techniques targeting HCPs used in the intervention arms of included studies. GQI, general quality improvement. Reproduced from Lawrensonet al.,127with permission from John Wiley & Sons, Ltd. Copyright © The Cochrane Collaboration. Published by John Wiley & Sons, Ltd.

Reproduced from Lawrensonet al.,127with permission from John Wiley & Sons, Ltd. Copyright © The Cochrane Collaboration. Published by John Wiley & Sons, Ltd.

TABLE 2 - Definitions of and illustrative quotations for BCTs identified in the included studies

Reproduced from Lawrensonet al.,127with permission from John Wiley & Sons, Ltd. Copyright © The Cochrane Collaboration. Published by John Wiley & Sons, Ltd.

BCTs, abbreviated definitions and DRS-specific examples	Illustrative quotation from existing DRS interventions
Goals and planning
Goal-setting (behaviour)	Practice nurses planned independent consultations with patients. The monitoring tool guided them through the consultations, and provided the opportunity to help the patient in selecting appropriate, concrete, behavioural goalsFreiet al.69(pp. 1040–1)
Set or agree a goal defined in terms of the behaviour to be achieved (e.g. set targets for how often patients should attend DRS, or general diabetes self-management, such as the frequency of blood glucose testing or the amount of carbohydrates to consume at each meal)
Problem-solving	the health educator . . . offered one-on-one, interactive education and counselling. Having established rapport, she worked to identify and understand each subject’s reasons for and/or barriers to having a dilated retinal examination. Focused problem-solving then guided the subject toward making an informed choice about receiving an ophthalmic examinationBaschet al.109(p. 1879)
Analyse, or prompt the person to analyse, factors influencing the behaviour and generate or select strategies that include overcoming barriers and/or increasing facilitators (e.g. support patients to identify reasons for wanting or not wanting to attend DRS and help them select potential strategies for overcoming these barriers to screening attendance)
Goal-setting (outcome)	During the case management sessions, patients and providers set management goals that were reasonable to achieveBarcelóet al.60(p. 147)
Set or agree a goal defined in terms of a positive outcome of wanted behaviour (e.g. agree with the patient target glycated haemoglobin level, blood pressure or cholesterol levels or target range for blood glucose levels)
Action-planning	Behavioural activation for diabetic retinopathy prevention combined the principles of education about diabetes mellitus, behavioural therapy, and the health belief model to assist participants in identifying barriers to obtaining DFEs [dilated fundus examinations], problems-solving solutions to surmounting barriers, formulating action plans to facilitate DFEs, and gauging the success of action plansWeisset al.121(p. 1007)
Prompt detailed planning of performance of the behaviour (e.g. support the patient to develop a plan for how often they will attend DRS, where the DRS will occur and how they will get to their appointment)
Review behaviour goals	Care managers were trained to use a patient-centred self-management approach that included review of the medical care needs and self-care goals that the patient identified and brainstorming additional strategies that patients could use to overcome barriers to their goalsGlasgowet al.73(p. 35)
Review behaviour goal(s) jointly with the person and consider modifying goal(s) or the behaviour change strategy in light of achievement [e.g. during scheduled diabetic review consultations, discuss with patients how they are progressing with their agreed self-management behavioural goals (e.g. frequency of blood glucose testing, attendance for DRS); when patients are not meeting agreed goals, either discuss how to adjust goals if needed to increase feasibility or engage in problem-solving to overcome any barriers to goal attainment]
Discrepancy between current behaviour and goals	Physicians in the IG [intervention group] received a monthly report of their care quality with the top 10% quality of diabetes care score for all physicians being the achievable benchmarkHayashinoet al.77(p. 599)
Draw attention to discrepancies between a person’s current behaviour and the person’s previously set outcome goals, behaviour goals or action plans (e.g. provide feedback to HCPs on the proportion of patients who have received DRS in the previous 12 months and compare this against a gold standard for clinical practice based on clinical guidelines)
Review outcome goal(s)	The telephone call was structured to first review the patient’s goals, followed by medication use, symptoms, glucose monitoring, blood pressure monitoring and self-management/care activitiesTayloret al.102(p. 1059)
Review outcome goal(s) jointly with the person and consider modifying goal(s) in light of achievement (e.g. review or alter target blood glucose levels towards a more feasible/achievable intermediate target)
Behavioural contract	Care guides asked patients to sign a contract (which was scanned into the EHR [electronic health record]) agreeing to work toward their disease-specific goalsAdairet al.59(p. 176)
Create a written specification of the behaviour to be performed, agreed by the person and witnessed by another (e.g. ask the person with diabetes to sign a contract in their self-management plan or diary, undertaking to attend DRS once)
Commitment	The initial goal was to elicit a verbal commitment to schedule an eye examinationBaschet al.109(p. 1879)
Ask the person to affirm or reaffirm the statement indicating commitment to change the behaviour (e.g. ask the person with diabetes to verbally affirm or reaffirm that they are committed to attending DRS at the agreed frequency and location)
Feedback and monitoring
Monitoring of behaviour by others without feedback	Foot examinations, blood pressure, and eye examinations were recorded on the reminder by clinic staff, collected after the patient visit and entered manuallyPetersonet al.58(p. 2239)
Observe or record behaviour with the person’s knowledge as part of a behaviour change strategy (e.g. record the proportion of patients who attend for a DRS examination as part of clinical audit, but do not feed back the results to the HCPs whose practice has been audited)
Feedback on behaviour	In addition, diabetic members who did not have a record of a diabetic retinopathy exam received educational materials and a report of their current DRE [diabetic retinal examination] status directly from the HMO [health maintenance organisation] 2 weeks laterHalbertet al.114(p. 753)
Monitor and provide information or evaluative feedback on performance of the behaviour (e.g. form, frequency, duration, intensity) (e.g. provide a feedback report to HCPs, stating the proportion of their patients who have attended a DRS examination, had their blood pressure taken and had a foot examination)
Self-monitoring of behaviour	We prepared feedback sheets for adherence to these eight indicators using data from the physicians’ self-report forms, as the physicians monitored and promoted these indicators to improve adherenceHayashinoet al.77(p. 601)
Establish a method for the person to monitor and record his or her behaviour(s) as part of a behaviour change strategy (e.g. a person with diabetes maintains a self-management diary in which they record their daily food intake and exercise and tick off on a checklist when they have attended the annual DRS examination)
Self-monitoring of outcomes of behaviour	In general, case managers were directed to encourage patient self-management, including diet and exercise, provide reminders for recommended screening/tests, help with appointment scheduling; monitoring home glucose and blood pressure levelsKreinet al.85(p. 734)
Establish a method for the person to monitor and record the outcome(s) of their behaviour as part of a behaviour change strategy (e.g. a person with diabetes records in their self-management diary the results of their latest glycated haemoglobin test and DRS examination)
Monitoring of outcomes of behaviour by others without feedback	The nurse case manager used behavioural goal setting, established individualized care plan, provide patient self-management education and surveillance of patients . . . ordered protocol-driven laboratory tests, tracked the outcomes using the computerized data registryGabbayet al.71(p. 30)
Observe or record outcomes of behaviour with the person’s knowledge as part of a behaviour change strategy (e.g. a person attends a DRS examination but is not provided with the results of the examination)
Feedback on outcomes of behaviour	All persons who attended the screening clinics received a dilated eye exam by a volunteer community-based ophthalmologist. The eye exam included visual acuity . . . and a fundus examination through a dilated pupil . . . immediately after receiving the dilated eye exam, the patient was told the results by the examination ophthalmologistAndersonet al.108(p. 41)
Monitor and provide feedback on the outcome of performance of the behaviour [e.g. informing the person with diabetes of the results of the DRS examination (i.e. presence/absence of retinopathy)]
Biofeedback	Immediately after receiving the dilated eye exam, the patient was told the results by the examination ophthalmologistAndersonet al.108(p. 41)
Provide feedback about the body (e.g. physiological or biochemical state)using an external monitoring device as part of a behaviour change strategy
Social support
Social support (unspecified)	Overall, the intervention included . . . and self-management support (provided by the practice nurse)Freiet al.69(p. 1041)
Advise on, arrange or provide social support (e.g. from friends, relatives, colleagues, ‘buddies’ or staff) or non-contingent praise or reward for performance of the behaviour – includes encouragement and counselling (e.g. provide general encouragement or reassurance to a person with diabetes to attend their DRS appointment)
Social support (practical)	Referrals were facilitated to other clinicians when indicated, including ophthalmology, podiatry, nutrition and primary care for follow-up of acute or other chronic issues or when requested by patientsJacobset al.82(p. 616)
Advise on, arrange or provide practical help (e.g. from friends, relatives, colleagues, ‘buddies’ or staff) for performance of the behaviour (e.g. provide practical help for a patient with diabetes to attend DRS; this can include, for example, arranging a referral to DRS, arranging or providing transport to the clinic)
Shaping knowledge
Instruction on how to perform the behaviour	A direct mail reminder was sent to patients to reinforce the importance of annual eye examinations and included the following text:If you don’t have an eye doctor, ask your regular doctor to refer you to onePrelaet al.118(p. 258)
Advise or agree on how to perform the behaviour (includes ‘skills training’) (e.g. provide advice to a person with diabetes on how often guidelines recommend attending DRS, where they can obtain DRS and how to schedule an eye examination)
Natural consequences
Information about health consequences	A tailored telephone intervention was delivered by bilingual interventionists:Risk communications, such as the frequent lack of symptoms of retinopathy and that early treatment for retinopathy decreases the risk of blindness, were includedWalkeret al.120(p. 187)
Provide information (e.g. written, verbal, visual) about the health consequences of performing the behaviour (e.g. provide advice to the person with diabetes on the negative health consequences of retinopathy and the benefits of early detection.)
Salience of consequences	The videotape used emotional appeals through storytelling to increase motivation to have a yearly dilated retinal examinationBaschet al.109(p. 1879)
Use methods specifically designed to emphasise the consequences of performing the behaviour with the aim of making them more memorable (e.g. give a person with diabetes a leaflet containing testimonials from other people with diabetes who suffer from retinopathy to emphasise the benefits of attending DRS and of early detection)
Information about social and environmental consequences	A take-home reminder (aimed at patients, to remind them to make an appointment for an eye examination), to be given to patients by their family practitioner, included the following text:OHIP [Ontario Health Insurance Plan] covers annual eye checks for patients with diabetes so you will not have to payZwarensteinet al.123(p. 90)
Provide information (e.g. written, verbal, visual) about the social and environmental consequences of performing the behaviour (e.g. provide information on the costs of having a DRS examination)
Information about emotional consequences	Group visit content, though patient-guided, was physician-directed to cover educational topics . . . and the emotional aspects of diabetesClancyet al.62(p. 621)
Provide information (e.g. written, verbal, visual) about the emotional consequences of performing the behaviour (e.g. provide a leaflet recognising the potential negative effects on emotional and mental health of managing a chronic illness such as diabetes)
Comparison of behaviour
Demonstration of the behaviour	The newsletter consisted of six sections, including a testimonial designed to model eye examination behaviourEllishet al.113(p. 1593)
Provide an observable sample of the performance of the behaviour, directly in person or indirectly (e.g. through film, pictures), for the person to aspire to or imitate (e.g. play a video demonstrating the DRS procedure)
Social comparison	The system presented register data on their Type 2 diabetes population, giving them the option either to use the data during individual diabetes consultations or to gain an overview of the quality of their diabetes care and compare it with the corresponding quality in their colleagues’ practicesGuldberget al.74(p. 326)
Draw attention to others’ performance to allow comparison with the person’s own performance (e.g. provide HCPs with feedback on the proportion of their patients who have had a DRS examination and benchmark this in comparison to other hospitals or HCPs)
Information about others’ approval	One of the messages in the targeted newsletter read:Even though you’ve been thinking about getting a dilated eye exam, we hope you’ll make the call nowEllishet al.113(additional information provided by the author)
Provide information about what other people think about their behaviour. The information clarifies whether others will like, approve of or disapprove of what the person is doing or will do (e.g. tell the person with diabetes that their family members would likely be keen for them to attend their DRS appointment)
Associations
Prompts/cues	For those who made an appointment, a reminder letter was mailed 3 weeks prior to the scheduled appointment. Additionally, there was an automated reminder call the day before the scheduled appointmentPizziet al.117(p. 255)
Introduce or define environmental or social stimulus with the purpose of prompting or cueing the behaviour (e.g. telephone the person with diabetes to remind them of their upcoming DRS appointment)
Reduce prompts/cues	Recommendations for regular telephone follow-ups for diabetes patients, which will be monthly in the 1st half year and then will probably decreaseJansinket al.83(coded from protocol128)
Withdraw gradually prompts to perform the behaviour [e.g. decrease the frequency with which a person with diabetes is sent a reminder about their DRS attendance (i.e. from weekly to fortnightly, to monthly, to quarterly reminders)]
Repetition and substitution
Behavioural practice/rehearsal	During a 2-day training session, case managers received instruction on collaborative goal setting, with case examples and role-playing used to familiarize them with the treatment algorithmsKreinet al.85(p. 734)
Prompt practice or rehearsal of the performance of the behaviour one or more times in a context or at a time when the performance may not be necessary, to increase habit and skill (e.g. provide an opportunity for trainee HCPs to practise delivering a DRS examination to an actor role-playing a patient with diabetes)
Graded tasks	Theoretically, this form of facilitation should be necessary for only a relatively short period of time, with the practice improvement team progressively assuming responsibility for the ongoing improvement efforts after the initial facilitationDickinsonet al.64(p. 10)
Set easy-to-perform tasks, making them increasingly difficult, but achievable, until the behaviour is performed (e.g. initially allocate a HCP responsibility for one component of the DRS examination and progressively increase their level of responsibility)
Comparison of outcomes
Credible source	Participants in the print-intervention group received a mailing of a colourful, 14-page booklet on preventing diabetes eye problems called Keep Your Eyes Healthy, in English or Spanish, developed by the National Institutes of HealthWalkeret al.120(p. 187)
Present verbal or visual communication from a credible source in favour of or against the behaviour (e.g. include the logos for national health institutes or cite published clinical guidelines to endorse information provided in leaflets regarding DRS)
Reward and threat
Material incentive (behaviour)	The automated system offered a live telephone call back to assist in scheduling test and also offered to send participants the following items: 1) a voucher that would allow the provider to waive the co-payment for a dilated eye examination . . .Simonet al.98(p. 1452)
Inform that money, vouchers or other valued objects will be delivered if and only if there has been effort and/or progress in performing the behaviour (e.g. advise the person with diabetes that they will receive a shopping voucher if they attend their upcoming DRS appointment)
Social reward	When a subject reported having a dilated retinal examination a congratulatory letter was sentBaschet al.109(p. 1879)
Arrange a verbal or non-verbal reward if and only if there has been effort and/or progress in performing the behaviour (e.g. verbally praise the person with diabetes if they attend their DRS appointment)
Non-specific reward	CME [continuing medical education] credits were given to the participating physicians in the workshopsVidal-Pardoet al.104(p. 752)
Inform that a reward will be delivered if and only if there has been effort and/or progress in performing the behaviour (e.g. inform the HCP that they will be rewarded for conducting a DRS examination with a target proportion of their patients)
Antecedents
Restructuring the physical environment	Care guide workstations were located in the clinic waiting rooms, to facilitate face-to-face interactions with patients, providers, and nursesAdairet al.59(p. 177)
Change, or advise changing, the physical environment to facilitate performance of the wanted behaviour or create barriers to performance of the unwanted behaviour (e.g. introduce mobile DRS vans in geographically remote areas to increase access to screening facilities)
Restructuring the social environment	Three multi-lingual Link Workers already employed by Coventry Primary Care Trust (PCT) were trained in diabetes management and care and assigned to work with specific intervention GP [general practice] surgeriesBushet al.110(p. 295)
Change, or advise changing, the social environment to facilitate performance of the wanted behaviour or create barriers to performance of the unwanted behaviour (e.g. change a health-care team and team working, such as introducing a new specialist diabetes nurse role responsible for monitoring screening rates and telephoning people with diabetes to remind them to attend their DRS appointment)
Adding objects to the environment	In addition 4500 diabetes passports were made available at the four hospitalsDijkstraet al.65(p. 128)
Add objects to the environment to facilitate performance of the behaviour (e.g. introduce new computerised software to a general practice to help monitor and remind HCPs about which patients need to be prompted to attend DRS)
Scheduled consequences
Behaviour cost	We were interested to find out whether a small copayment would be an important deterrent to the uptake of screening for diabetic retinopathy (DR) . . . We conducted a randomized trial in which one group was charged a small fee for DR screening and the other was provided with free accessLianet al.115(p. 1247)
Arrange for withdrawal of something valued if and only if an unwanted behaviour is performed (e.g. charging people with diabetes a fee for failing to attend a DRS examination)
Self-belief
Verbal persuasion about capability	Diabetes is a serious, lifelong condition, but there is so much that you can do to protect your health. Take charge of your health, not only for today, but also for the years to comeLafataet al.86(p. 523)
Tell the person that they can successfully perform the wanted behaviour, arguing against self-doubts and asserting that they can and will succeed (e.g. encourage or reassure the patient to attend a DRS examination, providing information as needed to address any concerns or self-doubts that they may have about attending for DRS)
Focus on past success	A comprehensive programme that integrated lifestyle counselling based on motivational interviewing principles was integrated into structured diabetes careJansinket al.83(p. 119)In additional file 2, which is linked to the protocol for this study,128the authors describe the staff training programme. Within the text they explain that one guiding principle of motivational interviewing is to ‘support self-efficacy’. They state that ‘self-efficacy can be strengthened by affirming past success (i.e. reinforcement)’
Advise thinking about or listing previous successes in performing the behaviour (or parts of it) (e.g. help the person with diabetes to remember the last time that they attended DRS and use this as an opportunity to reassure them of the benefits of attending)

For studies comparing any intervention with usual care, the majority of studies provided no description of usual care, which precluded coding of the comparator arm.

Excluded studies

SeeReport Supplementary Material 1,Characteristics of included, ongoing and excluded studies (phase 1 review).

Ongoing studies

SeeReport Supplementary Material 1,Characteristics of included, ongoing and excluded studies (phase 1 review).

Risk of bias in included studies

A risk-of-bias assessment was conducted using the Cochrane EPOC Group risk-of-bias tool (Figure 8). 51Overall, trials were judged to be at low or unclear risk of bias. However, 33 studies (48.5%)58–60,62,65,68,69,71,72,74,76,78,81–84,88,91–97,99,100,103–107,119,121were judged to be at high risk of bias in at least one domain. The domains most commonly at high risk of bias were ‘incomplete outcome data addressed’ [15 studies65,68,72,76,78,81–84,88,93,94,100,103,105(22.7%)], ‘protected against contamination’ [seven studies60,62,96,99,104,107,119(10.6%)] and ‘similar baseline outcome measurements’ [five studies58,69,92,95,106(7.6%)]. It was possible to judge if a study was free from selective outcome reporting for only 17 studies (25.7%)58,59,63,64,67,69,72,77,78,83,84,94,96,99,100,121,123because of a lack of availability of a prospectively published trial registration or protocol. Although studies were rarely judged to be at a high risk of bias for adequate sequence generation and adequate allocation concealment, the methods associated with these domains were frequently poorly reported.

FIGURE 8.

Risk-of-bias graph.

We explored publication bias using a funnel plot for the main comparison of any intervention with usual care (56 studies58–63,65–72,74–78,80–89,92–105,108–112,115–118,120–123). Although there were few data points for the less precise studies, those with greater precision were evenly distributed.

The studies that reported economic outcomes are a subset of the studies included in the systematic review and the risk of bias of these studies was very similar to the risk of bias of the main body of included studies. With respect to methodological quality, only five67,117,124–126of the 14 included economic studies reported full economic evaluations. One of these studies124was published as an abstract and lacked important methodological details. Only three96,117,125of the studies with full economic evaluations reported a sensitivity analysis to explore changes in the costs and outcomes under different scenarios. Discounting in economic evaluations is necessary to adjust future costs and outcomes of an intervention to their present values but was reported in only one96of the full economic outcomes. Its use would have been appropriate in those other studies with a stated follow-up of > 12 months. 67,70,85,89,105The methodological quality of the full economic evaluations was considered to be moderate. Full details of the methodological quality assessment for each of the included economic evaluations are available inReport Supplementary Material 1[seeFurther details of the review of economic evidence (phase 1 review)].

Effects of the interventions

Primary outcome

One or more visits for diabetic retinopathy screening within a 2-year period following implementation of the intervention

All 66 trials provided data for this outcome. There were two types of comparison: 56 (85%)58–63,65–72,74–78,80–89,92–105,108–112,115–118,120–123of the 66 studies compared an intervention against current usual care whereas 10 (15%) studies64,73,79,90,91,106,107,113,114,119compared a more intensive QI intervention or group of QI interventions against a less intensive intervention. As these addressed different questions, meta-analyses were conducted separately on the 56 and 10 studies.

Thirty-one (47%)58,60,64–70,73,74,76–79,81,83,84,90–94,97,100,101,104–106,110,123of the 66 trials were cluster RCTs. Only nine65,67,69,70,73,83,93,101,123of these reported an ICC and the ICC reported typically did not relate specifically to DRS outcomes. Of the nine cluster RCTs reporting an ICC, the most commonly reported value was 0.05 and so this was the value imputed for studies with no estimates of ICCs. The smallest value reported was 0.01 and the largest value was 0.2. A sensitivity analysis was conducted to investigate the impact on the computed effect estimates of using the lower and upper range values.

Comparison A: any quality improvement intervention compared with usual care

Of the 56 studies that compared any intervention against usual care, 13 (23%)108–112,115–118,120–123evaluated interventions specifically targeting DRS. The remaining 43 studies (77%)58–63,65–72,74–78,80–89,92–105evaluated interventions directed towards improving the general quality of diabetes care (including DRS attendance). Although there was substantial heterogeneity in the intervention effects (I² > 90%), 48 out of the 56 studies showed an improvement in DRS attendance (Figure 9). As it may be argued that it is better to examine clinical differences in a meta-analysis rather than use them as a reason for not conducting one, we computed pooled estimates for each of these subgroups. A random-effects model was adopted, which can accommodate statistical heterogeneity between studies by assuming that different studies have different true effect sizes (Figure 9), but we acknowledge that use of the random-effects model does not in itself deal with heterogeneity. We assessed whether or not there was evidence of a subgroup effect and, as there was not (p = 0.15), subsequent analyses were conducted on the 56 studies. The pooled RD for comparisons between the intervention and usual care was 0.12 (95% CI 0.10 to 0.14). This corresponds to a 12% absolute increase in DRS attendance.

FIGURE 9.

Meta-analysis of any QI interventions compared with usual care. df, degrees of freedom; M-H, Mantel–Haenszel.

Comparison B: more intensive (stepped) intervention compared with less intensive intervention

Examples of studies in this comparison included a tailored (individualised) newsletter compared with a generic patient education newsletter113and a comparison between audit and feedback to HCPs and audit and feedback combined with a diabetes outreach service. 91Ten studies64,73,79,90,91,106,107,113,114,119contributed to this analysis (Figure 10). Three (30%)113,114,119evaluated interventions specifically targeting DRS, whereas seven (70%)64,73,79,90,91,106,107evaluated interventions directed towards improving the general quality of diabetes care. The pooled RD for comparisons between a more intensive (stepped) intervention and a less intensive intervention was 0.05 (95% CI 0.02 to 0.09). This corresponds to a 5% absolute increase in DRS attendance for participants receiving the stepped intervention.

FIGURE 10.

Meta-analysis of stepped QI interventions compared with less intensive QI interventions. df, degrees of freedom; M-H, Mantel–Haenszel.

Exploration of heterogeneity

Substantial heterogeneity was observed (I² > 90%), which was investigated by subgroup analysis and meta-regression.

Subgroup analysis

Comparisons of studies grouped according to the QI component/BCTs present in the intervention arm were conducted as well as comparison of studies at high compared with low risk of bias and according to the resource intensity scale.

Although we attempted to code for the QI component/BCT in the control arm in all studies, we did not adjust for this in the analysis as, for the majority of these studies, usual care was not specified and therefore could not be coded.

Quality improvement components

A sufficient number of studies were available to investigate the effectiveness of nine out of the possible 12 QI components (seeAppendix 4for definitions of the QI components). Insufficient data were available to analyse ‘continuous quality improvement’, ‘financial incentives’ and ‘facilitated relay of information to clinicians’. Interventions incorporating all nine QI components evaluated in the subgroup analysis were associated with improvements in DRS attendance, with higher pooled effect estimates for interventions directed at patients (promotion of self-management and patient education) or the organisation of the health system (team changes or the establishment of an electronic patient registry) (Figure 11).

FIGURE 11.

Subgroup analysis of QI components.

Component behaviour change techniques

Sufficient studies were available to investigate the effectiveness of interventions containing particular BCTs (including 10 BCTs aimed at patients and seven aimed at HCPs) (Figures 12and13). Interventions incorporating all 17 BCTs included in the subgroup analysis were all shown to be effective at improving DRS attendance. For BCTs aimed at patients, higher pooled effect estimates were found for ‘goal-setting (outcome)’ and ‘credible source’; for BCTs aimed at HCPs, higher pooled effect estimates were found for ‘restructuring the social environment’ and ‘credible source’ (seeTable 2andAppendix 5for BCT descriptions).

FIGURE 12.

Subgroup analysis of BCTs aimed at patients.

FIGURE 13.

Subgroup analysis of BCTs aimed at HCPs.

Resource utilisation

Analysis of subgroups according to resource intensity did not show a relationship between effect size and increasing resource intensity (Figure 14).

FIGURE 14.

Subgroup analysis based on resource utilisation.

Meta-regression

The results of the meta-regression analysis are provided inTables 3–5. There was some evidence of an association between effect size and baseline risk, with larger effects in studies with poorer baseline DRS attendance (Figure 15). Because of regression to the mean, this association might be spurious, so a permutation test was conducted to allow for this (with 1000 permutations;p = 0.055).

TABLE 3 - Results of the meta-regression analysis of type of QI intervention

Explanatory variable	Regression coefficient (95% CI)	p-value	ResidualI2(%)
Baseline DRS attendance	–0.208 (–0.419 to 0.004)	0.054	94
High vs. low risk of bias	0.008 (–0.136 to 0.094)	0.079	92
Cluster vs. individual RCT	–0.049 (–0.136 to 0.039)	0.268	93
QI component
Audit and feedback	–0.029 (–0.136 to 0.079)	0.597	92
Case management	0.001 (–0.092 to 0.095)	0.979	93
Team changes	0.089 (–0.001 to 0.178)	0.052	91
Electronic patient registry	0.051 (–0.061 to 0.164)	0.363	92
Clinician education	–0.011 (–0.107 to 0.085)	0.822	93
Clinician reminders	–0.012 (–0.127 to 0.103)	0.835	92
Patient education	0.075 (–0.009 to 0.160)	0.081	92
Promotion of self-management	0.085 (–0.002 to 0.172)	0.055	93
Patient reminders	–0.033 (–0.128 to 0.063)	0.493	93
Resource requirement	0.013 (–0.015 to 0.042)	0.356	93

TABLE 4 - Results of the meta-regression analysis of BCTs aimed at HCPs

Explanatory variable (HCP-BCTs)	Regression coefficient (95% CI)	p-value	ResidualI2(%)
Feedback on outcomes of behaviour/biofeedback	–0.035 (–0.129 to 0.060)	0.465	92
Social support (practical)	–0.013 (–0.128 to 0.102)	0.821	93
Instruction on how to perform the behaviour	–0.022 (–0.110 to 0.066)	0.622	93
Prompts/cues	–0.029 (–0.126 to 0.069)	0.559	92
Credible source	0.028 (–0.075 to 0.132)	0.586	93
Restructuring the social environment	0.085 (–0.001 to 0.172)	0.053	91
Adding objects to the environment	–0.003 (–0.102 to 0.095)	0.946	92

TABLE 5 - Results of the meta-regression analysis of BCTs aimed at patients

Explanatory variable (patient BCTs)	Regression coefficient	p-value	ResidualI2(%)
Problem-solving	0.042 (–0.073 to 0.158)	0.466	92
Goal-setting (outcome)	0.162 (0.070 to 0.254)	0.001	90
Feedback on outcomes of behaviour/biofeedback	0.081 (–0.016 to 0.179)	0.102	92
Social support (unspecified)	0.063 (–0.037 to 0.162)	0.211	91
Social support (practical)	0.012 (–0.079 to 0.103)	0.789	92
Instruction on how to perform the behaviour	0.031 (–0.059 to 0.120)	0.497	92
Information about health consequences	–0.023 (–0.114 to 0.069)	0.624	93
Prompts/cues	–0.050 (–0.137 to 0.037)	0.251	93
Credible source	0.097 (–0.016 to 0.211)	0.092	92
Restructuring the social environment	0.051 (–0.043 to 0.145)	0.284	93

FIGURE 15.

Bubble plot showing the relationship between the RD and the baseline percentage screened. The RD represents the difference in DRS attendance between the intervention arm and the control arm. Each bubble corresponds to a different study and the size of the bubble reflects the number of participants. The regression line shows a trend towards reduced screening attendance with increasing baseline compliance. Reproduced from Lawrensonet al.,127with permission from John Wiley & Sons, Ltd. Copyright © The Cochrane Collaboration. Published by John Wiley & Sons, Ltd.

Reproduced from Lawrensonet al.,127with permission from John Wiley & Sons, Ltd. Copyright © The Cochrane Collaboration. Published by John Wiley & Sons, Ltd.

The effect sizes from studies at high risk of bias were slightly (but not statistically significantly) higher than those from studies at low risk of bias. Similarly, no statistically significant difference was found between individual RCTs and cluster RCTs (p = 0.268).

For the meta-regression analyses comparing studies containing particular QI components/BCTs with studies not including these components, no statistically significant differences were found, although there was evidence of an association for ‘team changes’ (p = 0.052) and ‘promotion of self-management’ (p = 0.055). The patient-targeted BCT ‘goal-setting (outcome)’ was associated with a significantly greater improvement in DRS attendance than in studies without this BCT (p = 0.001) and there was some evidence of an association for the HCP-targeted BCT ‘restructuring the social environment’ (p = 0.053).

Overall quality of the evidence

The summary of findings for the primary outcome of this review is presented inTable 7. For the main comparison (any intervention vs. usual care), as most of the information came from studies at low or unclear risk of bias, we did not downgrade for risk of bias. We downgraded one step for inconsistency as there was substantial unexplained heterogeneity between the studies. We found no reasons to downgrade the evidence on the basis of indirectness, imprecision or publication bias. Figure 16provides a funnel plot for the main comparison.

FIGURE 16.

Funnel plot of comparison: any QI intervention compared with usual care. SE(RD), standard error of the risk difference. Reproduced from Lawrensonet al.,127with permission from John Wiley & Sons, Ltd. Copyright © The Cochrane Collaboration. Published by John Wiley & Sons, Ltd.

Reproduced from Lawrensonet al.,127with permission from John Wiley & Sons, Ltd. Copyright © The Cochrane Collaboration. Published by John Wiley & Sons, Ltd.

Summary of the main results

Quality of the evidence

Overall, we judged the quality of the evidence to be moderate using the GRADE evidence rating system. 57The level of evidence was downgraded one step because of inconsistency of findings in studies comparing QI interventions with usual care and one step for imprecision or inconsistency in studies comparing a less intensive with a more intensive intervention. Using the Cochrane EPOC risk-of-bias tool,51we found that nearly half of the included studies were at high risk of bias in at least one domain. This was most often because of incomplete outcome reporting. For many domains it was not possible to judge the risk of bias because of poor reporting. For example, because many of the RCTs did not have a prospectively published protocol, it was not possible to make a judgement whether or not outcomes were selectively reported. A subgroup analysis found that, although studies that were at a high risk of bias had slightly higher effect estimates than those at a low risk of bias, this difference was not statistically significant. The consensus of the review team was not to downgrade the quality of the evidence for risk of bias.

Of the 22 potential ‘economic’ studies identified by the review team, 14 studies were eligible for the review as partial or full economic evaluations. We identified a publication bias in two of the eight excluded studies. 94,123These studies failed to report the planned economic evaluations as no evidence of intervention effectiveness was identified. Such an approach could be considered as selective outcome reporting, such that potentially negative economic findings are not reported. This phenomenon of reporting bias has been recognised previously,129when studies with unfavourable effectiveness results are not published or are published later in low-impact journals. Furthermore, analytically, such an approach is substandard as these studies conflate absence of evidence with a finding of evidence of absence (of an effect).

Most of the economic evaluations had limitations in their reporting, with few providing a breakdown of the costs associated with delivering the different components of the intervention. There was also insufficient evidence to show whether or not part of the direct costs of the intervention and care may be offset by reduced productivity costs. However, it is important to note that the expected findings of an effective intervention would be gains in health and reductions in the costs of treating diabetes. The overall methodological quality of the included economic studies was mixed. The partial economic evaluations identified lack, by their nature, the methodological characteristics expected of an economic evaluation. The methodological quality of the full economic evaluations was considered to be moderate.

Agreements and disagreements with other studies or reviews

Only one systematic review22has previously investigated the effectiveness of interventions to increase the uptake of DRS. Although this review included 48 studies, only 12 of these were RCTs (10 of which were included in the current review). The authors similarly concluded that a variety of interventions can be effective in improving screening uptake, including increasing patient and provider awareness of diabetic retinopathy, introducing a computer-based registration/reminder programme and developing a community-based health-care system.

Compared with the paucity of systematic reviews of the impact of interventions to improve DRS outcomes, many reviews have evaluated the impact of general QI interventions to improve the overall quality of diabetes care. 28A systematic review published by members of the current team29included 48 cluster RCTs and 94 patient RCTs and found improvements in many important quality outcomes for patients with diabetes. A meta-analysis of a subset of 23 RCTs reporting the uptake of DRS found a similar pooled relative effect size to that in the current review (RR 1.22, 95% CI 1.13 to 1.32;I² = 80.4%).

Limitations of the review

Our systematic review has some limitations. Coding of intervention content was challenging given the paucity of primary data sources, although in some cases this was offset by obtaining further information from researchers on intervention content, who also provided the materials used to deliver the interventions. We were not able to assess the impact of some QI intervention components because of too few trials being available for our subgroup and meta-regression analyses. Furthermore, we could not control for all potential confounding factors. Given the complexity of the interventions under investigation, which incorporate multiple QI components, it is likely that other covariates may have interacted synergistically or antagonistically with them. The short duration of the included RCTs (typically ≤ 12 months) also meant that we were unable to assess the effect of the QI interventions on ongoing DRS attendance or their impact on the progression of diabetic retinopathy.

Implications for practice

The results of this review provide evidence that QI interventions targeting patients, HCPs or the health-care system are likely to be associated with meaningful improvements in DRS attendance compared with usual care. There was no statistical difference between interventions specifically aimed at DRS and those that were part of a general QI strategy for improving diabetes care. This is a significant finding because of the additional benefits of general QI interventions in terms of improving glycaemic control, vascular risk management and screening for other microvascular complications. 28,29It is likely that further (but smaller) improvements in DRS attendance can also be achieved by increasing the intensity of a particular QI component or adding further components.

One of the main objectives of the review was to identify the ‘active’ components of successful interventions by using validated taxonomies to describe the content of the interventions. All of the QI components as defined by the modified EPOC taxonomy were associated with improvements in DRS attendance. To better characterise intervention content we coded the interventions in terms of patient and provider BCTs. For BCTs aimed at patients, higher effect estimates were found for interventions incorporating goal-setting; for BCTs aimed at HCPs, higher effect estimates were found for interventions involving environmental restructuring. However, only 42% of the 93 possible BCTs were reported in the included interventions and, although not all BCTs in the taxonomy might be appropriate for DRS, the findings of this review suggest that there may be opportunities to assess the potential of additional BCTs in future trials of novel interventions to improve screening attendance.

Although only three of the studies were conducted in the UK, we believe that the results of the review are applicable to a UK context. Two-thirds of the included studies were conducted in North America, where guideline recommendations include a DFE by an ophthalmologist or optometrist at least once every 2 years (or more frequently in high-risk individuals). Although the North American system for DRS is different from the national screening programme that operates in the UK, the two care pathways are similar in that DRS is generally conducted in a different health-care setting from that for other aspects of diabetes care.

Implications for research

The review highlighted a number of gaps within the evidence base. There was limited evidence on the relative effectiveness of QI interventions in particular population subgroups according to demographic characteristics that could have an impact on health equity, for example ethnicity, level of education or socioeconomic status. There is also insufficient evidence on the cost-effectiveness of QI interventions to improve DRS attendance.

Most of the included studies, whether targeting DRS or general QI, enrolled patients not achieving diabetes-relevant quality indicators. For example, five studies specifically targeting DRS exclusively recruited patients who were not meeting guideline recommendations for screening. It is not clear if the interventions would be as effective in populations with a higher level of screening attendance (> 80%). There was some evidence from our meta-regression analysis that the effectiveness of the interventions is negatively correlated with baseline level of DRS.

Although we have been able to show that interventions containing particular BCTs have a greater likelihood of success, given the multicomponent nature of interventions, it is likely that the presence of other BCTs or other effect modifiers in the intervention arm may also have an impact on effectiveness. The analysis conducted as part of this review did not attempt to isolate the impact of individual QI/BCT components, although exploratory analysis was conducted using multivariate meta-regression methods to determine cost-effectiveness (reported inChapter 6). Further research is needed to identify which components of interventions or combinations of components can optimally improve DRS attendance at an acceptable cost.

Aim

In the evidence synthesis reported in this chapter we set out to complement the evidence on the effectiveness of QI interventions used in RCTs to improve attendance for DRS (seeChapter 3). We aimed to clarify how QI interventions might work and gain an understanding of DRS attendance behaviour by identifying the theoretical determinants of DRS attendance.

The specific objectives were to:

• identify the published and grey literature reporting perceived barriers and enablers associated with DRS attendance

• extract reported barriers/enablers and categorise them according to TDF/CFIR domains

• identify key content themes within domains, regarding barriers to and enablers of DRS attendance

• apply prespecified criteria to identify the likely importance of TDF/CFIR domains in influencing DRS attendance.

Study eligibility criteria

Search strategy

Study selection process

Following deduplication, one member of the research team (EGR) screened the titles and abstracts of all of the identified references against the inclusion/exclusion criteria to decide whether or not the full-text manuscripts should be retrieved. Each study was judged as either (1) not meeting the eligibility criteria for inclusion or (2) potentially meeting the eligibility criteria for inclusion. A second report author (FL) screened the titles of approximately 10% of the titles and abstracts. Inter-rater reliability statistics were calculated. The full-text screening was conducted using similar methods as used to screen titles and abstracts.

Data extraction and analysis materials/tools

We developed a data extraction form to extract study characteristics, including author/date, method of identification (e.g. database search, Google search), country where study was conducted, research objectives, topics of investigation (e.g. knowledge, attitudes, beliefs or perceptions of barriers/enablers), methodological/theoretical approach, data collection methods, data analysis methods, participant characteristics and sample size.

We prepared a codebook with the definitions of each of the 14 theoretical domains from the TDF and the 39 distinct constructs from the five domains of the CFIR, to facilitate coding consistency and reliability. Two report authors (EGR and FL) collaboratively coded three included studies, making reference to the codebook and amending it when appropriate.

Quality assessment

One review author (EGR) assessed the quality of the included studies using items selected from the Critical Appraisal Skills Programme (CASP) qualitative checklist139and the Mixed Methods Appraisal Tool (MMAT). 140Studies that reported mixed methods were appraised using both the quantitative and the qualitative appraisal tools. A quantitative survey that reported open-ended questions but which did not analyse the responses qualitatively was defined as a quantitative study. Studies presented as abstracts only were not subjected to quality appraisal. A second review author (JGL) assessed a random sample of studies (20%) and differences of opinion regarding quality were resolved by discussion.

Analysis

To the extent that it was possible in a secondary analysis context, we followed analysis methods that have been used in previous studies applying the TDF to interview transcripts from semistructured interviews. 141–143These methods typically follow a combined content and framework analysis approach and involved five steps: (1) data extraction, (2) deductive analysis (TDF coding), (3) inductive analysis (thematic synthesis), (4) CFIR coding and (5) identifying important domains. Figure 17presents a flow diagram of the steps in the analysis.

FIGURE 17.

Flow diagram of steps in the analysis.

Search results

Searches of MEDLINE, EMBASE, PsycINFO, Web of Science, CENTRAL and ProQuest generated 3251 records. Searches in OpenGrey, PsycEXTRA and the HMIC database generated a further 62 records. A further 144 records were identified through the other sources, totalling 3457 records. After deduplication, a total of 3010 records were screened based on title and abstract. In total, 2800 records were excluded, leaving 210 studies for full-text screening. When the full text of a record was not available but the abstract presented meaningful results, these were also screened against our inclusion/exclusion criteria. We identified 65 studies24,147–210that met our eligibility criteria. Figure 18presents the PRISMA flow diagram.

FIGURE 18.

Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram.

Characteristics of included studies

Report Supplementary Material 2[seeCharacteristics of the 65 included studies (phase 2 review)] shows the characteristics of the 65 included studies. Forty-one (63%)147,148,150,151,153,156,162–166,171,175–180,182,184–187,190,192,194–199,201–210used quantitative methods only (e.g. questionnaires, surveys), 18 (28%)149,152,154,155,157–160,168,169,172–174,181,188,189,191,200used qualitative (e.g. interviews/focus groups) methods only and six (9%)24,161,167,170,183,193used mixed methods. Forty-seven studies used quantitative methods in full or as part of a mixed-method design. Of these 47 studies, 37 (79%) used a cross-sectional survey design, nine (19%) used a follow-up design and one (2%) was a report of a questionnaire development study. Twenty-four studies used qualitative methods in full or as part of a mixed-method design. Of these 24 studies, 19 (79%) were descriptive studies that had no specific analytical or theoretical approach, one (4%) was part of an evaluation of a QI intervention, one (4%) used a theoretical framework (TDF), one (4%) used a grounded theory approach, one (4%) used a case-base approach and one (4%) was a service evaluation.

Twenty-nine (45%) of the studies were carried out in the USA, 12 (18%) in the UK, four (6%) in Australia, three (5%) in Canada, 10 (15%) in Asia, four (6%) in Africa, two (3%) in Europe and one (2%) in South America. Forty-eight studies (74%) were published as full texts in peer-reviewed journals, five (8%) were reported as unpublished full-text reports/dissertations and 12 (18%) were published in abstract/poster form only.

Fifty studies (77%) reported barriers/enablers from the perspective of the people with diabetes only, 14 (21%) reported barriers/enablers from the perspective of both people with diabetes and HCPs and one (2%) reported barriers/enablers from the perspective of HCPs who were diagnosed with diabetes. Twenty-nine studies (45%) reported barriers/enablers from a subset of patient samples. These included 10 studies (15%) that reported barriers/enablers from the perspective of specific ethnic groups (e.g. African Americans, American Indians, Aboriginal Canadians, people of South Asian or Hispanic origin), six studies (9%) that reported from the perspective of people who were classified as either non- or late attenders, four studies (6%) that reported from the perspective of older adults, two studies (3%) that reported from the perspective of younger adults, two studies (3%) that reported from the perspective of women only, two studies (3%) that reported from the perspective of people who had been diagnosed with diabetic retinopathy, one study (2%) that reported from the perspective of people with diabetes receiving treatment, one study (2%) that reported from the perspective of people with diabetes who had participated in a blindness prevention programme and one study (2%) that reported from the perspective of a Medicare population. Report Supplementary Material 2[seeExcluded studies (phase 2 review)] presents a description of articles that were assessed for inclusion at the full-text stage but which were subsequently excluded.

Quality of included studies

The quality appraisals of the included studies (n = 5324,147–161,163–170,172–175,177–180,182,184,185,187,190–192,194,195,197,199,201–210) are presented inReport Supplementary Material 2[seeQuality assessment (phase 2 review)]. Overall, studies were judged to be at low, medium or unclear risk of bias. However, studies were often poorly reported, which made the process of quality appraisal difficult. Limitations in the qualitative studies included (1) limited reporting or rigour of the data analysis and (2) limited reporting or consideration of the relationship between the authors and the participants. Generally, the strengths of these studies were (1) the research design (although there was often a lack of information justifying the approach), (2) the recruitment strategy and (3) the data collection procedures. Limitations in the quantitative studies related to (1) poor reporting of questionnaire validity and (2) low response rates.

Deductive analysis

In total, 612 units of data were extracted, of which 378 were qualitative (126 quotations from study participants and 252 from authors’ conclusions) and 234 were quantitative [e.g. percentages of participants agreeing with a questionnaire item or odds ratios (ORs)].

Reported barriers were identified in all but one of the TDF domains (‘skills’). Enablers were identified in all but two domains (‘beliefs about capabilities’ and ‘skills’). However, overall, there were almost three times as many themes/subthemes identified as barriers only than themes/subthemes identified as enablers only (60 vs. 22). Twenty-one themes/subthemes represented both barriers and enablers. Table 8reports the frequencies of barriers and enablers identified within each of the 15 TDF domains.

Inductive analysis

Report Supplementary Material 2(seeThemes/subthemes within each of the 14 domains from the Theoretical Domains Frameworks) presents all themes and subthemes identified within each domain, alongside frequencies, relevant studies and sample quotations. A narrative description of the themes, within domains, is presented later in this chapter, organised into two groups: (1) domains that were identified as being ‘high’ in importance (seeThematic synthesis for domains identified as having high importance) and (2) domains that were identified as being less important (seeDomains identified as being less important).

Consolidated Framework for Implementation Research: organisational analysis

Themes identified in the four TDF domains that arguably relate to the organisational level (as specified in step 4 of the methods) were recoded into CFIR constructs. The full results are presented inReport Supplementary Material 2(seeCoding of themes/subthemes into Consolidated Framework for Implementation Research constructs). The studies included in this review predominantly reported barriers and enablers for the person with diabetes, as reported by the patient, HCP or study author, rather than for the organisation or HCP, as conceptualised by the CFIR. Therefore, the majority of the themes were coded into the construct ‘needs and resources of those served by the organisation’ which, according to the CFIR, is defined as ‘the extent to which the needs of those served by the organisation (e.g. patient), as well as barriers and facilitators to meet those needs, are accurately known and prioritized by the organisation’ [reproduced from the CFIR. 211This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See:http://creativecommons.org/licenses/by/4.0/]. The focus of the CFIR on the HCP as the individual of interest and not the person with diabetes means that, in this study, it did not offer any additional insights over and above the TDF. As a consequence, no further analysis was conducted using the CFIR.

Importance of the Theoretical Domains Framework domains

This section reports the findings regarding the three importance criteria (frequency, elaboration and expressed importance).

Thematic synthesis for domains identified as having high importance

The content themes in the domains that were identified as being potentially important factors influencing screening attendance are described in further detail in the sections below.

Domains identified as being less important

The content themes in the domains that were not identified as being factors of high importance in influencing screening attendance are described in further detail in the sections below.

UK studies

Twelve24,149,155–159,162,169,174,183,189of the 65 studies (18%) were carried out in the UK, three162,183,189of which were reported in abstract form only. Most of the key issues (themes) identified from studies in other countries were also identified in at least one study carried out in the UK. Table 10lists the themes/subthemes that were identified in (1) UK studies only, (2) UK studies and studies from other countries and (3) non-UK studies only.

The themes and domains most frequently identified within the UK studies were (1) scheduling appointment issues (environmental context and resources), identified in eight24,149,155–159,169of the 12 UK studies (67%), (2) accessibility of the screening clinic (environmental context and resources), identified in eight149,156–159,162,169,183of the 12 UK studies (67%), (3) time – competing demands (environmental context and resources), identified in eight149,155–158,174,183,189of the UK studies (67%), (4) confusion between screening and routine eye tests (knowledge), identified in six155–159,169UK studies (50%), (5) symptoms (memory, attention and decision processes), identified in six149,155,158,159,169,189UK studies (50%), (6) competing health problems (memory, attention and decision processes), identified in six24,149,156–158,169UK studies (50%), (7) awareness of illness (knowledge), identified in five155,158,159,169,174UK studies (42%), (8) doctor–patient communication (social influence), identified in five24,149,156,159,169UK studies (42%), (9) fear and anxiety (emotions), identified in five157–159,174,183UK studies (42%) and (10) forgetting (memory, attention and decision processes), identified in five24,149,157,159,183UK studies (42%).

Patient and health-care professional perspectives

All studies included in this review either reported findings from the patient’s perspective or tested a hypothesis based on patient data. Only 14 studies (22%) reported barriers and enablers from the perspective of the HCP. Most of these studies did not distinguish between the two perspectives and integrated the findings from people with diabetes and HCPs. Only four studies noted any clear differences between the two perspectives. 152,160,169,174In one study the authors reported that HCPs had cited inadequate patient education as the major barrier, yet claimed that participating patients largely reported that diabetes education was adequate and instead felt that financial concerns were their biggest barrier. 160In another study the author reported that, for the most part, patients’ and HCPs’ perspectives mirrored each other. 169However, they did note that views on whether or not language was a barrier differed. Although people with diabetes reported to some degree that language was a barrier, most reported that there were systems in place to overcome these issues. In contrast, HCPs (in this case screeners) reported that language was an issue that they faced in everyday practice. In a third study, people with diabetes and HCPs agreed that regular attendance could be inconvenient but suggested that HCPs underestimated the difficulties that people could face when taking time off work. 174In the fourth study the author explained that HCPs and people with diabetes face different barriers to annual screening. For the person with diabetes the biggest barrier was a lack of understanding of the importance of screening and a lack of awareness of insurance coverage. The HCPs, on the other hand, identified the need for patient education on eye examinations and suggested that greater assistance was needed with tracking the patient through the system. 152

Summary of findings

The combined content and framework analysis identified six TDF domains [(1) ‘environmental context and resources’, (2) ‘social influences’, (3) ‘knowledge’, (4) ‘memory, attention and decision processes’, (5) ‘beliefs about consequences’ and (6) ‘emotions’] as representing the most important factors potentially influencing screening attendance. Hence, the hypothesis arising from this review is that interventions that target these domains are more likely to be effective at increasing screening attendance. This process also identified three TDF domains that potentially have the least influence on screening: (1) ‘optimism’, (2) ‘reinforcement’ and (3) ‘skills’. Hence, we propose that interventions targeting these three domains are less likely to be effective at increasing screening attendance (seeTables 8and9).

Inductive analysis within domains generated specific content themes that may help to identify potential targets for future QI interventions. The content themes were identified at multiple levels, including the person with diabetes (e.g. confusion between screening and routine eye tests), the HCP (e.g. absence or presence of a HCP recommendation), the health-care system (e.g. inaccurate register) and the wider community (e.g. lack of media coverage) (seeTable 10). Recoding the themes from the TDF domains into CFIR constructs did not offer any further insights and, as a consequence, we did not apply the importance criteria to the CFIR constructs. The large majority of barriers and enablers were identified in the UK and also in the context of health systems in other countries (seeTable 10).

Strengths, limitations and challenges

The combination of deductive coding (informed by theoretical frameworks to guide barrier identification) and inductive analysis (to allow more granular content themes, unanticipated findings and patient insights to emerge) is a strength of this review. Furthermore, this relatively novel approach of applying multilevel theoretical frameworks within the systematic review context acknowledges the potential influence of organisational and contextual factors on screening attendance. We were able to code all extracted data from the 65 studies into the TDF domains, thus demonstrating that the TDF framework provided a comprehensive coverage of barriers and enablers.

Another strength of this review is its inclusiveness. We included published and grey literature, qualitative and quantitative methodologies, patients’ and HCPs’ perspectives and any context and/or screening model. We reasoned that limiting the search to UK studies only or to one methodological perspective would risk overlooking potentially important and relevant findings. Although not all barriers and enablers will be relevant to all settings, this review gives a comprehensive overview of potential factors that may influence screening attendance.

The studies in this review predominantly identified barriers and enablers from the perspective of the person with diabetes rather than from the perspective of the organisation or HCP. Even the data that we had from the HCPs mostly focused on their views regarding patient barriers. Therefore, the CFIR did not offer additional insights over and above those obtained from the TDF. The CFIR could potentially contribute an interesting perspective but this would involve a slight reconceptualisation of the components of the framework. For example, it is plausible that, by conceptualising the patient as part of the ‘system’ (rather than the ‘outer context’) and describing DRS in terms of ‘innovation characteristics’ (seeAppendix 7), the CFIR could offer additional insights over and above those obtained from the TDF. This kind of reconceptualisation would need further investigation and methodological work, which is outside the scope of this study.

Although there were no major methodological concerns that we believe would have biased the outcomes of the included studies, it was unfortunate that a number of the studies were poorly or thinly described. This was problematic as it hampered our ability to differentiate between HCPs’ and patients’ perspectives or to distinguish between different patient subgroups. Furthermore, the data extracted and analysed in the present review were the data that were reported, analysed and interpreted by the study authors. It is possible that our data set may have been biased in that authors may have selectively reported findings on perceived barriers/enablers that were more prevalent or interesting or that had a better fit with the stated research question. There may have been additional views on barriers/enablers in the full, original data set that were not represented in the present findings. A final limitation of this review worth noting is that the theoretical frameworks used are limited in that they do not specify relationships between domains and hence the strength of the direct impact of barriers on behaviour is not known.

Implications for research

The findings of this review have important implications for the design of effective interventions to promote DRS attendance as they can be used to assess the extent to which the active components (BCTs) identified in existing QI interventions are actually targeting the theoretical domains that are important in determining attendance. This can be achieved by mapping the BCTs identified in the review reported inChapter 3against the theoretical domains identified in the current review using established mapping tools. The results of the mapping study are reported inChapter 5.

Implications for practice

The findings from this review can help to inform a set of recommendations on which theoretical domains to target in interventions (i.e. ‘environmental context and resources’, ‘social influences’, ‘knowledge’, ‘memory, attention and decision processes’, ‘beliefs about consequences’ and ‘emotions’) and which to avoid (i.e. ‘optimism’, ‘reinforcement’ and ‘skills’), thus enhancing the likely effectiveness of DRS programmes. Furthermore, within the TDF domains judged to be of high importance, the findings from the inductive analysis provide more detailed insight into which specific barriers and enablers may be associated with attendance behaviour. Findings from this level of analysis suggest potential targets for future QI interventions. Four key recommendations based on the findings from the thematic synthesis are presented in the following sections.

Recommendations for future research

In this review only 12 studies were conducted in the UK. Additional high-quality studies are therefore needed from the UK, from the perspective of both HCPs and people with diabetes. Ideally, such additional research would focus on sampling the subgroups that are potentially at the highest risk of blindness or that are least likely to attend for screening, as the pattern of important domains may differ between these subgroups depending on specific patient characteristics (e.g. location, age or socioeconomic status).

Future research could endeavour to identify which theoretical domains are most important for the subgroups that have been identified as being most at risk and/or least likely to attend for screening. For example, only two studies identified in this review explored factors that affect young adults. 188,189This patient group is not only under-researched but also at high risk of vision loss/blindness from diabetic retinopathy.

It is possible that the less frequently identified domains (e.g. ‘skills’, ‘intentions’, ‘optimism’) may simply not be as commonly assessed in quantitative work or are less likely to be spontaneously mentioned in response to open questions in interviews. Although such an occurrence may itself be an indication of low importance, it may be interesting to carry out a comparison of the results of this study against the results of more comprehensive assessments that identify not only omissions of these factors but provide evidence which shows that they have been assessed empirically and have been found to have no association with DRS attendance.

Concluding remarks

This review employed a systematic, theory-informed and replicable approach to identify barriers and enablers associated with DRS attendance. Six TDF domains were identified as having high importance and therefore are the theoretical domains most likely to be key mediators of DRS attendance behaviour. Thematic synthesis identified key content themes that offer further insights into which specific issues need to be addressed. Future research is needed to identify which domains are most important for subgroups of people with diabetes who have been identified as being most at risk.

Aims

This study aimed to integrate the findings fromChapters 3and4by applying a mapping tool to assess the theoretical coherence of existing DRS interventions, that is, the extent to which BCTs used in interventions (seeChapter 3) target the identified important theoretical determinants of screening attendance (seeChapter 4).

The specific research questions were:

1. What is the level of theoretical coherence of the intervention literature relating to DRS, that is, to what extent do BCTs frequently used in existing interventions to increase DRS attendance target the important barriers to attendance?

2. To what extent is observed theoretical coherence associated with intervention effectiveness, that is, higher screening attendance?

3. Are there any potential ‘missed opportunities’ for intervention design in terms of BCTs that are not frequently used in existing interventions but which are theoretically coherent with important barriers?

Sample

Data from the first systematic review (seeChapter 3) relevant to the work presented in this chapter included (1) the frequency with which each BCT was identified as present in intervention arms targeting people with diabetes and/or HCPs and, for those BCTs that were frequently identified (i.e. in ≥ 10 intervention arms), (2) the observed association between the BCT and intervention effectiveness (i.e. higher DRS rates in terms of RD, 95% CI). Findings from the review of perceived barriers to and enablers of screening attendance (seeChapter 4) relevant to the work presented in this chapter included the domains identified as important to screening attendance. Important domains (ranked in the top six) were ‘environmental context and resources’, ‘social influences’, ‘knowledge’, ‘memory, attention, and decision processes’, ‘beliefs about consequences’ and ‘emotions’ (seeTable 9).

Materials: mapping tool

A mapping tool was developed by integrating two previously published tools214,216(seeReport Supplementary Material 3). The integrated tool was used to identify the BCTs that are theoretically coherent with important domains (i.e. the domains that, arguably, should be used; presented inReport Supplementary Material 3,Behaviour change techniques theoretically coherent with important domains). The tool was also applied to the DRS literature to identify the domains that have been targeted by frequently used BCTs (i.e. the barriers that have been targeted in interventions; presented inReport Supplementary Material 3,Domains targeted by frequently used behaviour change techniques).

Procedure and analysis

Exploration of theoretical coherence between BCTs identified in the first review (seeChapter 3) and TDF domains identified in the second review (seeChapter 4) was conducted through two, bidirectional mapping processes, structured around the three research questions.

Research question 1: theoretical coherence of existing diabetic retinopathy screening interventions

Overall, high levels of theoretical coherence between frequently used BCTs in existing interventions and important theoretical domains were observed. The majority (80%) of frequently used BCTs in interventions targeting people with diabetes and all (100%) frequently used BCTs in HCP interventions were paired with at least one important domain (fromChapter 4). Only two BCTs in interventions targeting people with diabetes were paired exclusively with domains of lesser importance and were thus deemed to have low theoretical coherence: ‘goal-setting (outcome)’ and ‘problem-solving’ (seeTable 11).

Research question 2: association between theoretical coherence and intervention effectiveness

The subgroup analyses reported inChapter 3(seeFigures 12and13) identified that all BCTs frequently used in patient and/or HCP interventions were associated with higher DRS attendance (seeTable 11). Higher effect estimates were observed for the BCTs ‘goal-setting (outcome)’ and ‘credible source’ in interventions targeting people with diabetes and for the BCTs ‘restructuring the social environment’ and ‘credible source’ in interventions targeting HCPs (seeFigures 12and13). However, with the exception of ‘goal-setting (outcome)’ (p = 0.001), differences in effect estimates between BCTs were not statistically significant (seeTable 5). All frequently used BCTs were associated with effectiveness, regardless of whether they had high or low theoretical coherence. However, the majority (n = 15/17) of statistically significant BCTs had high theoretical coherence according to the integrated mapping tool.

Research question 3: ‘missed opportunities’ for intervention design

Table 12presents the frequency with which BCTs paired with important TDF domains were identified in existing DRS interventions. Overall, the BCTs that paired with the six most important domains have not been used frequently in existing interventions targeting DRS attendance. This finding indicates numerous missed opportunities for intervention design. Opportunity seized was highest for the domains ‘memory, attention and decision processes’ (50% of the theoretically coherent BCTs were frequently used in interventions) and ‘knowledge’ (43% of the theoretically coherent BCTs were frequently used). The most missed opportunities were for the domain ‘emotions’ (no BCTs paired with this domain were frequently used in existing DRS interventions (range 0–3 intervention arms) (seeTable 11).

Research question 1: theoretical coherence of existing diabetic retinopathy screening interventions

Overall, there is a high degree of theoretical coherence between the components of existing DRS attendance interventions and the important theoretical determinants of screening attendance. Nearly all component BCTs frequently included in existing interventions were paired with at least one of the six most important theoretical domains representing the key barriers to or enablers of DRS attendance. This suggests that existing interventions may work as they are targeting important factors influencing whether or not people with diabetes attend DRS.

Research question 2: association between theoretical coherence and intervention effectiveness

It was not possible to formally evaluate the association between theoretical coherence and effectiveness as all frequently used BCTs were associated with improved screening attendance. This suggests that all BCTs are effective regardless of whether they have high or low theoretical coherence. However, the majority (88%) of frequently used BCTs with significant effect estimates were classified as having high theoretical coherence. It is possible that intervention designers more frequently select BCTs with high coherence because they are aware of theory and/or evidence or because they intuitively select BCTs that are more likely to address relevant barriers to and enablers of screening attendance.

Research question 3: ‘missed opportunities’ for intervention design

Our second mapping process investigated the extent to which theoretically coherent BCT–domain pairs have been used in existing interventions to increase DRS attendance. We identified that the inclusion of BCTs that are potentially relevant to the important domains is limited and variable in existing interventions. All six important domains had at least one theoretically coherent BCT that was not frequently used in existing interventions. Such BCTs thus represent potential ‘missed opportunities’ for intervention design. Addressing these missed opportunities may thus increase intervention effects.

Specifically, all BCTs paired with the domain ‘emotions’ were used not at all, or only very infrequently, in existing interventions. There is thus a specific priority for the inclusion in future interventions and research of BCTs that target the emotional determinants of screening attendance. For instance, an important theme within the ‘emotions’ domain was fear/anxiety about vision loss. For some people with diabetes, the fear of losing their vision acted as a strong incentive to attend screening (e.g. Applebee,159Hartnettet al. 160and Dervanet al. 165). However, for others, fear of a diabetic retinopathy diagnosis and of the screening procedure were perceived barriers to attendance. 174,185,190This barrier to screening attendance could be addressed through the delivery of the BCTs ‘reduce negative emotions’, ‘information on health consequences of behaviour’ and ‘anticipated regret’, by providing information on treatment options for diabetic retinopathy. Such information could focus on providing reassurance by emphasising that effective treatments are available following early detection. The ‘missed opportunities’ identified through this second mapping process therefore represent a theoretically informed means of identifying BCTs that may be of greater relevance and effectiveness.

Strengths

The strengths of the current study include the transparent approach to systematically assessing the extent to which intervention components target theoretical determinants of behaviour. The use of available BCT × TDF domain mapping tools in a systematic review context represents a methodologically innovative aspect of this work and provides an example of how these recently developed tools may be further applied in secondary analyses of existing interventions. This systematic approach also demonstrates how these frameworks may be applied to identify limitations of, and missed opportunities in, the design of existing interventions,221and thus offers clear implications for future intervention refinement, design and research. In turn, this systematic approach has clear implications for future intervention design, refinement and research.

Additionally, to our knowledge, this is the first study to provide empirical findings that can be used to evaluate the validity of the proposed BCT and domain pairings in the mapping tools, which are based on expert opinion and consensus. Although based on a limited sample size, the frequently used BCTs that were effective and that map onto key TDF domains provide empirical support for the agreed BCT and domain pairings. Conversely, frequently used BCTs that were effective but that do not map onto key TDF domains with high coherence potentially refute the recommendations in the mapping tools. Given the high proportion of effective BCTs achieving high theoretical coherence, there is a trend towards a supportive pattern of findings for the proposed pairings in the available tools.

Limitations

An important caveat when interpreting these findings and a limitation to this methodological approach concerns the validity and applicability of the mapping tools. There are discrepancies across, and limitations of, the published tools132,216on which we based the integrated mapping tool. Hence, low observed coherence may not reflect the ‘true’ level of theoretical coherence and instead may be a product of limitations regarding the validity of the materials used to identify coherence. For instance, one may query if the BCT ‘goals’ is conceptually applicable to the domain ‘skills’ or if the BCT ‘feedback on outcomes of the behaviour’ is conceptually applicable to the domains ‘beliefs about consequences’ and ‘beliefs about capabilities’ (seeReport Supplementary Material 3,Behaviour change techniques theoretically coherent with important domains). It is also surprising that the BCTs involving ‘feedback’ are not paired with the domain ‘behavioural regulation’, given that feedback is regarded as the archetypal self-regulation strategy. 218,222,223Further research is thus needed to empirically evaluate and refine these mapping tools.

Furthermore, current BCT–domain mapping tools do not distinguish between different contexts, populations and types of behaviour, whereas theoretical coherence may depend on the nature of the behaviour and the population being investigated.

Conclusions

The first aim of this study was to assess the theoretical coherence of existing DRS interventions, that is, the extent to which BCTs used in interventions (seeChapter 3) target the identified important theoretical determinants of screening attendance (seeChapter 4). We identified high levels of theoretical coherence between frequently used BCTs in interventions to increase DRS attendance and important barriers to and enablers of screening attendance. The second aim was to identify whether high theoretical coherence was associated with intervention effectiveness (i.e. higher screening attendance). Although the large majority (88%) of frequently used BCTs with significant effect estimates had high theoretical coherence, it was not possible to formally evaluate the association between theoretical coherence and effectiveness, as all frequently used BCTs were associated with improved screening attendance. The third aim was to identify potential ‘missed opportunities’ for intervention design, in terms of infrequently used BCTs that are theoretically coherent with important barriers. A number of additional BCTs may be relevant to addressing important theoretical determinants of screening attendance that are not included in existing interventions. There are particular gaps around the inclusion of intervention components to address the emotional barriers to and enablers of screening attendance. Inclusion of such components in the design of future interventions, or the enhancement of existing interventions, may help to further improve DRS attendance and in turn ensure the timely referral and treatment of patients who test positive for diabetic retinopathy. There is also a need for further empirical research to evaluate the validity of mapping tools to facilitate the selection of coherent BCTs for future research and intervention design.

Resource use and cost estimates

The level of the cost of resource use in each included study was of interest for two purposes in this study. First, in the economic analysis, we were interested in estimating the cost of the individual components of the complex interventions included in each study rather than the cost of the complex interventions. This requires a form of multivariable regression with the QI and BCT components as explanatory variables, as used in the intervention effectiveness analysis. Second, inChapter 3the association between resource use and intervention effectiveness was investigated. InChapter 3the average effectiveness was calculated at different levels of resource use and resource use intensity was added as a covariate in a meta-regression.

A measure of resource use intensity was needed for each study. As there were 56 studies with a usual care comparator, an efficient method was required to derive this measure. The approach taken was to design a data abstraction method for recording predefined key categories of resource use and the levels of each category. The details are provided inReport Supplementary Material 4. The process of determining the resource categories and levels involved agreement between reviewers on an ordered ranking of resource use for each study. The cost of each level of each resource category was then estimated through random sampling of at least three studies with each category and level of resource use, costing the resource use associated with the intervention description using national cost estimates and calculating the average. The total cost of a complex intervention in each study is the product of the resource categories and estimated costs for each category. The cost variable was therefore categorical and the applicable regression method to estimate the incremental cost of each QI/BCT component was an ordinal logistic regression, with the ordered resource use ranking as the dependent variable.

The costs of the interventions from the 56 included studies were distributed across a total of 16 ordered ranks, which were utilised in the economic model. InChapter 3these 16 ranks were put into five groups to facilitate the analysis of an association between resource use and intervention effectiveness. The grouping was as follows: ranks 1–3 became rank 1; ranks 4–6 became rank 2; ranks 7–9 became rank 3; ranks 10–12 became rank 4; and ranks 13–16 became rank 5.

The cost analysis for the economic model differs from the work conducted in the review of economic outcomes reported inChapter 3in that the economic model utilises unit cost estimates for England and Wales to value the described interventions rather than converting the costs reported in the studies from one currency to another. Arguably, such data would be more transferable as the resources required to provide an intervention may not vary between settings.

It was assumed that all people with diabetes would be eligible for DRS and that BCT/QI interventions would target all eligible people. It is noted that some of the evidence on the effectiveness of BCT/QI interventions came from studies targeting those most likely not to attend DRS. It was not possible to control for this factor in the meta-regression data and we have assumed a constant effect. However, we have explored the impact of different baseline uptake rates of DRS to illustrate the relative impact of BCTs/QI components when the baseline uptake is lower (seeBase-case and sensitivity analyses).

Resource use may be incurred at different organisational levels. It may differ between individuals or at the GP or practice level or even be determined at the level of the NHS. It was assumed that all software development and the design of educational pamphlets would be financed at a higher organisational level than, for example, a general practice, for simplicity taken as the national level. This assumes that a country takes advantage of economies of scale. The total numbers of people with diabetes (n = 2,913,538),230general practices (n = 8151) and GPs (n = 40,265)231in England were used to derive a cost per patient for each intervention. The price year was 2016.

Details of how the ordered resource use and cost estimates were incorporated in the model are provided inReport Supplementary Material 4(seeEconomic model).

Analyses and interpretation

Quality improvement results

The base-case analysis includes a baseline probability of DRS attendance of 0.7. The probability that each QI component is more cost-effective than no QI intervention at improving attendance at DRS is presented inTable 15. All of these results are based on effectiveness and cost analyses that adjust for the other QI components. They provide an indication of which QI components are worth prioritising over which others when designing a DRS attendance intervention, which is likely to consist of a combination of QI components. The results of the sensitivity analyses including baseline DRS attendance probabilities of 0.5, 0.6, 0.8 and 0.9 are also reported inTable 15. These results assume a societal willingness-to-pay threshold for a QALY of £20,000. The results assuming a threshold of £30,000 per QALY are reported inAppendix 10.

The results are presented as the probability that a QI component is cost-effective at a specific willingness-to-pay threshold. SeeAnalyses and interpretationfor an explanation of these results.

Table 15shows that there is considerable uncertainty in the results when considering which interventions are cost-effective compared with no intervention. In pairwise comparisons, a probability of > 0.95 could be considered highly likely to be cost-effective and a probability of > 0.7 could be considered to show some evidence of cost-effectiveness, but with significant uncertainty. There are areas of uncertainty that are not captured in these results but the results indicate which QI components are worth prioritising when developing interventions from a cost-effectiveness perspective. In the base case, with a baseline probability of DRS attendance of 0.7, there are no QI components with a probability of cost-effectiveness of > 0.9. Patient education has the highest probability of being cost-effective (0.64).

The greater scope for increasing DRS attendance with a lower baseline probability of screening attendance is reflected in the higher probabilities of cost-effectiveness for each QI component compared with no QI intervention at lower baseline probabilities of screening attendance. Even at a baseline probability of DRS attendance of 0.5, only ‘team changes’ and ‘patient education’ have a probability of being cost-effective of > 0.7. As the baseline probability of DRS attendance reduces from 0.7 to 0.5, the probability of being cost-effective increases more for ‘team changes’ than for ‘patient education’. This is because resource use is estimated to be much less for ‘patient education’ than for ‘team changes’ (see proportional RRs of resource use ranking inTable 13).

The probability of each QI intervention being cost-effective when all QI factors are compared simultaneously is reported inTable 16. The QI components with the highest probabilities of cost-effectiveness inTable 15also have the highest probabilities of cost-effectiveness inTable 16. WhatTable 16shows is that, among the QI components with the highest probabilities of being cost-effective, there is no single QI component that stands out as being far more likely to be cost-effective. Patient education has the highest probability of being cost-effective and this is only 0.31 in the base case.

Sensitivity analyses for scenarios including different probabilities of monitoring following HES referral attendance, different DRS frequencies and different ages of the baseline cohort and including and excluding social care costs are presented inTable 17.

Monitoring may occur if treatment is not provided after a positive diagnosis following HES referral attendance. In the base-case analysis it was assumed that 78% of people who are not treated following a positive diagnosis of preproliferative diabetic retinopathy, proliferative diabetic retinopathy or diabetic maculopathy are monitored. If this proportion is reduced to 40% then the probability of a given QI component being cost-effective compared with no QI intervention increases when DRS is scheduled every year.

The probability of an intervention being cost-effective is higher when DRS is carried out every 3 years than when it is carried out every 2 years, which in turn is higher than when DRS is carried out every year. This is consistent with the results of the Scanlonet al. 224cost–utility analysis, which identified screening every 3 years as being the most cost-effective frequency of screening.

In the base-case analysis it was assumed that the age of the cohort being screened was 64 years. If either a younger (age 56 years) or older (age 70 years) cohort is considered then the probability of an intervention being cost-effective compared with no intervention decreases. The cost-effectiveness of the intervention decreases at age 70 years because the potential cumulative benefit associated with the intervention decreases. Specifically, the likelihood of attending a referral to the HES decreases, the utility of an individual decreases with age and mortality increases. The quality of life benefits and potential cost savings increase with age. Starting DRS at an earlier age increases the duration that a benefit may be realised for an individual, but the greatest benefits occur a longer time after the start of DRS and these are discounted more heavily.

For other scenarios considered, the results described inTable 17indicate that there remains considerable uncertainty. However, for a cohort of people aged 64 years, increasing the screening interval to 2 or 3 years means that the probability that ‘team changes’ and ‘patient education’ would be considered cost-effective would be > 0.8 compared with no intervention. Similarly, increasing the screening interval to 3 years for the same cohort would mean that the probability that ‘promotion of self-management’ and ‘clinician education’ are cost-effective compared with no intervention would increase to > 0.7.

Behaviour change technique results

The base-case analysis includes a baseline probability of DRS attendance of 0.7. The probability that each BCT component would be more cost-effective than no BCT intervention in terms of improving attendance at DRS is presented inTable 18. The BCT components are listed as being either patient targeted or HCP targeted. All of these results are based on effectiveness and cost analyses that adjust for the other BCT components. They provide an indication of which BCT components are worth prioritising over which others when designing a DRS attendance intervention, which is likely to consist of a combination of BCTs. The results of the sensitivity analyses including baseline DRS attendance probabilities of 0.5, 0.6, 0.8 and 0.9 are also reported inTable 18. These results assume a societal willingness-to-pay threshold for a QALY of £20,000. The results assuming a threshold of £30,000 per QALY are reported inAppendix 10.

Similar interpretations of the probabilities presented for the cost-effectiveness of BCT components can be made as for the QI components. For the patient-targeted BCTs, at a probability of screening uptake of 0.7, ‘goal setting (outcome)’, ‘feedback on outcomes of behaviour’, ‘social support (unspecified)’ and ‘information about health consequences’ have extremely high probabilities of being cost-effective compared with no BCT intervention. To put this into context, this is crudely analogous to saying that a comparison is statistically significant in a one-sided test at conventional (5% levels) of statistical significance. The BCT ‘prompts/cues’ is also associated with a high probability of 0.84 of being cost-effective compared with no intervention.

For the HCP-targeted BCTs, only ‘adding objects to the environment’ has a probability of being cost-effective of > 0.9, with ‘credible source’ having a probability of being cost-effective of > 0.7.

When the baseline probability of DRS is decreased, then each BCT component has a higher probability of being considered cost-effective compared with no BCT intervention. At a baseline probability of DRS attendance of 0.5, ‘social support (practical)’ and ‘credible source’ are the only BCTs that have a probability of being cost-effective that rises above 0.8 compared with scenarios with a higher baseline probability of DRS attendance. When the baseline probability of DRS attendance is 0.9, however, all BCTs are less likely to be cost-effective and only the patient-targeted ‘information about health consequences’ has a probability of being cost-effective of > 0.9.

The probability of each BCT component being cost-effective when all BCT components are compared simultaneously is reported inTable 19. The BCT components with the highest probabilities of being cost-effective inTable 18also have the highest probabilities of being cost-effective inTable 19. WhatTable 19shows is that, among the BCT components with the highest probabilities of being cost-effective, there is no single BCT component that stands out as being far more likely to be cost-effective. Patient-targeted ‘information about health consequences’ has the highest probability of being cost-effective and this is only 0.47 in the base case. Patient-targeted ‘goal-setting (outcome)’, ‘feedback on outcomes of behaviour’, ‘social support (unspecified)’ and ‘information about health consequences’ all have a significant probability of being the most cost-effective isolated BCT component.

Sensitivity analyses for scenarios including different probabilities of monitoring following HES referral attendance, different screening frequencies and different ages of the baseline cohort and including and excluding social care costs are presented inTable 20.

Monitoring may occur if treatment is not provided after a positive diagnosis following HES referral attendance. As in the QI analysis, in the base-case analysis it was assumed that 78% of people who are not treated following a positive diagnosis of preproliferative diabetic retinopathy, proliferative diabetic retinopathy or diabetic maculopathy are monitored. If this proportion is reduced to 40% then the probability of a given BCT component being cost-effective compared with no BCT intervention increases when DRS is scheduled every year.

As the frequency of DRS is reduced, the probability of a BCT component being considered cost-effective compared with no BCT intervention increases. When the screening interval is extended to 3 years, all BCTs have a probability of being cost-effective of > 0.8 except for ‘Instruction on how to perform the behaviour’ targeted at both the patient and the HCP, ‘credible source’ targeted at the patient, ‘restructuring the social environment’ targeted at the patient, ‘instruction on how to improve behaviour’ targeted at HCPs and ‘feedback on outcomes of behaviour’ targeted at HCPs. This is consistent with the results of the Scanlonet al. 224cost–utility analysis, which identified screening every 3 years as the most cost-effective screening frequency. However, when the age of the cohort is reduced from 64 years (the base case) to 56 years then the BCT components all become less cost-effective compared with no BCT intervention, although for some BCTs, such as ‘information about health consequences’ targeted at the patient, the change is very small (the probability that this BCT is cost-effective remains at 0.998, which is extremely high). When the age of the cohort is increased to 70 years then the probability that a BCT is cost-effective compared with no intervention also decreases. The quality-of-life benefits and potential cost savings increase with age. Starting DRS at an earlier age increases the duration that a benefit may be realised for an individual, but the greatest benefits occur a longer time after the start of DRS and these are discounted more heavily. The results do not change greatly in the other sensitivity analyses.

Threshold cost curve

The threshold intervention cost per patient was derived for a range of hypothetical RD effect estimates, using the base-case model and assuming a societal willingness-to-pay threshold for a QALY of £20,000 (Figure 21). The base-case analysis excludes the social care cost implications but the results of a model including the social care cost implications are also presented inFigure 21.

FIGURE 21.

The maximum cost per-patient-screened threshold below which a hypothetical DRS uptake intervention remains cost-effective compared with no intervention for a range of RD effect estimates for the intervention compared with no intervention.

Similar results but with effectiveness reported as ORs are provided inFigure 22. As withFigure 21, the results for scenarios excluding and including social care cost implications are presented.

FIGURE 22.

The maximum cost per-patient-screened threshold below which a hypothetical DRS uptake intervention remains cost-effective compared with no intervention for a range of OR effect estimates for the intervention compared with no intervention.

The cost thresholds for a range of RDs and RRs are presented inTable 21. The cost thresholds are presented using three different units: per patient, per general practice and for England as a whole. This table represents a lookup table indicating a plausible cost range over which an intervention with a specific level of effect may be cost-effective at a societal willingness-to-pay threshold for a QALY of £20,000.

The maximum costs for the mean RD estimates for the QI and BCT components are presented inAppendix 11.

Summary of findings

In the base case there was considerable uncertainty in the cost-effectiveness of the QI components. The QI component most likely to be cost-effective at a societal willingness-to-pay threshold of £20,000 per QALY was patient education, which is perhaps the closest QI component to the most cost-effective BCT components. The probabilities of being cost-effective at a societal willingness-to-pay threshold of £20,000 per QALY were higher for BCTs than for QI components. Patient-targeted ‘goal-setting (outcome)’, ‘feedback on outcomes of behaviour’, ‘social support (unspecified)’ and ‘information about health consequences’ had extremely high probabilities of being cost-effective compared with interventions not containing this BCT. For the HCP-targeted BCTs, ‘adding objects to the environment’ had a probability of being cost-effective of > 0.9. The BCT components with the highest probabilities of being cost-effective were patient targeted. The sensitivity analyses that had a significant effect on the results were varying the baseline probability of screening uptake and reducing the frequency of screening. The change in the results was much more pronounced for the BCT components than for the QI components. A lower baseline probability of DRS attendance increased the probability of an intervention being cost-effective, whereas a higher baseline probability of DRS attendance reduced the probability of an intervention being cost-effective. Reducing the frequency of screening also increased the probability of an intervention being cost-effective.

The second objective was to estimate the cost threshold below which an intervention might be cost-effective over a range of effectiveness estimates. For a RD of 0.2, the cost threshold was £63 per patient or £4559 per general practice when including social care costs in the analysis.

Strengths, limitations and challenges

This economic analysis was concerned with estimating the cost-effectiveness of interventions designed to increase DRS attendance. Long-term cost and health outcomes (valued using QALYs) were incorporated using a published model to derive more useful estimates of the likelihood that QI and BCT components are cost-effective and of the cost threshold below which an intervention may be cost-effective. A number of assumptions were adopted within the Scanlonet al. 224model that we have not been able to review. In addition, the model used in our work was not a perfect reproduction of the Scanlonet al. 224model as simplifications were made. We relied on sensitivity analyses to test particularly significant changes to the model. It is possible that there are uncertainties that have not been evaluated, such as uncertainty associated with the outcomes of DRS attendance for different combinations of screening frequencies and interventions to encourage DRS attendance.

Although the objectives of the economic analysis were to evaluate the likelihood that individual QI components and BCTs are cost-effective, the interventions to increase DRS attendance that were studied are complex interventions consisting of multiple QI components and BCTs. The evidence on the effectiveness of the complex interventions was identified through a systematic review of the literature. The individual treatment effects and marginal costs of each QI component and BCT were estimated using multivariable meta-regressions. Although this is an appropriate method to estimate the individual effects, there is a possibility of confounding and spurious results when conducting meta-regression. Furthermore, there was heterogeneity in the study characteristics and most studies were not conducted in the UK, making firm conclusions around the cost-effectiveness of specific combinations of BCTs and QI components in specific UK populations difficult. Consequently, the economic analysis was limited to providing evidence on the likelihood that individual QI components and BCTs could be cost-effective. The focus of this was solely to help prioritise future research by focusing on QI components and BCTs that are most likely to be cost-effective when designing and studying interventions. It is possible that a complex intervention designed to incorporate a BCT or QI component with a high probability of being cost-effective may need to include a BCT or QI component that is less likely to be cost-effective.

Future research

This analysis suggests that future research should focus on designing and studying interventions consisting of one or more of the following patient-targeted BCTs: ‘goal-setting (outcome)’, ‘feedback on outcomes of behaviour’, ‘social support (unspecified)’ and ‘information about health consequences’. The HCP-targeted BCT, ‘adding objects to the environment’, should also be considered. The interventions studied may also include other BCTs.

Based on evidence from randomised controlled trials, how effective are quality improvement interventions for increasing diabetic retinopathy screening attendance? (seeChapter 3)

We conducted a systematic review of the evidence from RCTs that investigated any QI intervention to improve attendance for DRS. Interventions could be directed at patients with diabetes, HCPs, the health-care system or any combination thereof. The primary outcome was attendance for one or more visits for DRS within a 2-year period following randomisation. We included data from 66 studies (analysing data for 352,879 participants). Fifty of the included studies (76%) were RCTs of general QI interventions that sought to improve a number of diabetes quality indicators and 16 studies (24%) tested interventions that specifically targeted DRS. The majority of studies were conducted in North America, with only three studies conducted in the UK. The interventions were usually multifaceted, consisting of multiple QI components (e.g. ‘patient education’, ‘team changes’ and ‘audit and feedback’). We used two validated taxonomies to describe the content of the interventions: a modified version of the taxonomy29originally developed by the Cochrane EPOC Group43to provide a general description of the intervention and the BCTTv136to describe the specific QI components present in each intervention. Interventions were further described in terms of their component BCTs, which provides a level of granularity that is better suited to describing the active ingredients of the interventions. BCTs were coded using the BCTTv136in terms of whether they targeted the patient, the HCP, or both.

Although our meta-analysis for the main comparison identified that the pooled effect estimate for studies that specifically targeted DRS was larger (17% absolute increase in DRS attendance) than that in studies of general QI interventions (12% increase), the difference between these subgroups was not statistically significant. This is an important finding given the previously demonstrated effectiveness of general QI interventions for the overall quality of diabetes care. For example, a systematic review of the effectiveness of such interventions by members of the current review team (NMI, JMG),29which incorporated 26 of the trials included in the current review, found improvements in glycated haemoglobin (HbA_1c), diastolic and systolic blood pressure, low-density lipoprotein cholesterol and medication use and increased monitoring for other diabetes complications, for example foot screening.

The pooled effect estimates for studies comparing a more intensive (stepped) intervention with a less intensive intervention were smaller (5% increase in DRS attendance) than the pooled effect estimates for the comparisons with non-specified usual care. In terms of health economic outcomes, there was too much methodological variability in studies reporting comparative cost-effectiveness outcome measures to make a judgement on the relative cost-effectiveness of QI interventions compared with each other or with usual care. The effectiveness of interventions that included specific QI components or BCTs was also calculated for level of resource use, as measured on an ordered ranking scale, to explore whether or not more resource-intense interventions were more likely to be effective. The results indicate that there is no evidence of a clear relationship between effectiveness and intensity of resource use.

Sufficient studies were available to investigate the effectiveness of nine out of 12 QI components and 17 BCTs by subgroup analysis and meta-regression. All QI components analysed were associated with significantly higher DRS attendance. Larger pooled effect estimates were found for interventions directed at patients (promotion of self-management and patient education) or the organisation of the health system [team changes (e.g. use of multidisciplinary teams or expansion of professional roles) and introducing an electronic diabetes register]. All frequently used BCTs identified in at least 10 patient and/or HCP intervention arms were similarly associated with higher DRS attendance. Higher effect estimates were observed for the BCTs of ‘goal-setting (outcome)’ and ‘credible source’ (information on the importance of the behaviour from individuals/organisations with expertise in the field) in interventions targeting patients and for the BCTs of ‘restructuring the social environment’ (e.g. introducing diabetes link workers to facilitate the behaviour) and ‘credible source’ (e.g. persuasive argument from a respected colleague) in interventions targeting HCPs. Using univariate meta-regression we found evidence for larger effects in studies with poorer baseline DRS attendance.

The review highlighted a number of gaps within the evidence base. No data were available to assess the effectiveness of QI interventions to improve ongoing adherence to screening and limited data were available to assess the relative effectiveness of interventions in particular population subgroups, for example type 1 compared with type 2 diabetes and those of different ethnicities or socioeconomic status. Although we were able to show that interventions containing particular BCTs have a greater likelihood of success, given the multicomponent nature of interventions, it is likely that there is interaction with other BCTs or that other effect modifiers in the intervention arms may also be having an impact on effectiveness. The analysis conducted as part of this review did not attempt to evaluate the impact of individual BCTs in isolation; however, this was evaluated as part of the exploratory analysis (reported inChapter 6) to determine the likely cost-effectiveness of interventions in which multivariate meta-regression was used. This approach allows for more flexible modelling of arm-level covariates.

What are the theoretical determinants (e.g. barriers/enablers) of diabetic retinopathy screening attendance? (seeChapter 4)

We aimed to clarify how QI interventions might work and gain an understanding of factors influencing DRS attendance behaviour. This was achieved by identifying studies reporting primary data on barriers to and enablers of DRS attendance from the perspectives of people with diabetes and HCPs and framing these in terms of theoretical domains from the TDF (an integrated framework of behaviour change theories) and the CFIR (a comprehensive listing of constructs thought to influence implementation). We hypothesised that interventions that target these theoretical domains are more likely to be effective at increasing screening attendance.

From the analysis we identified the TDF domains ‘environmental context and resources’, ‘social influences’, ‘knowledge’, ‘memory, attention and decision processes’, ‘beliefs about consequences’ and ‘emotions’ as representing the most important factors potentially influencing screening attendance. We also identified that the other eight TDF domains – ‘social/professional role and identity’, ‘beliefs about capabilities’, ‘goals’, ‘intentions’, ‘behavioural regulation’, ‘optimism’, ‘reinforcement’ and ‘skills’ – may have less influence on screening behaviour.

These findings have two important implications for the design of effective interventions to promote DRS attendance: (1) they identify which theoretical domains to target in future QI interventions and which to avoid and (2) they can be used to assess the extent to which the active components (BCTs) identified in existing QI interventions are actually targeting the theoretical domains that are important in determining attendance.

The studies in this review predominantly identified barriers and enablers from the perspective of patients rather than the perspective of organisations or HCPs. Even the data from HCPs were mostly focused on patient barriers. Therefore, the CFIR did not offer additional insights over and above those gained from the TDF. Thematic synthesis within the TDF domains identified as being of high importance resulted in the identification of specific content themes that may help to identify potential targets for future QI interventions. There are four key recommendations based on these findings:

1. Reduce inconvenience to patients. Many of the barriers and enablers identified within the TDF domain ‘environmental context and resources’ relate to perceptions of convenience. Providing local screening facilities, ‘one-stop shops’, flexible appointment systems, childcare facilities and transportation might all be advantageous.

2. Increase awareness of the importance of DRS. HCP recommendations and information provision was reported to be an enabler within the TDF domain ‘knowledge’. Furthermore, within the TDF domain ‘social influences’, use of local media and local community networks was identified as another potential way to raise awareness.

3. Increase patients’ sense of comfort and support. A number of barriers were identified within the TDF domains ‘emotions’, ‘social influences’ and ‘beliefs about consequences’ that highlight the need to increase support for patients who struggle with issues such as communicating with HCPs, negative emotions (e.g. fear), a lack of social support and a lack of trust in the HCP or system. These barriers may be overcome by providing community-based clinics, ensuring that there is some social/cultural compatibility between the patient and the HCP and the HCP showing compassion to the patient.

4. Improve message content. Within the TDF domains ‘memory, attention and decision processes’, ‘beliefs about consequences’ and ‘knowledge’ we found a number of barriers that relate to health misconceptions. Many of the issues identified in this review, including those captured within recommendations 2 and 3, can perhaps be addressed through improved message content. Messages that highlight the asymptomatic nature of diabetic retinopathy, the safety of the screening procedure, the benefits of early detection (e.g. providing piece of mind) and the potential negative consequences if diabetic retinopathy is left unchecked are all potentially beneficial. Furthermore, messages that clarify the differences between routine eye tests and DRS, that offer reminders to attend for screening and which emphasise that DRS is a routine part of diabetes care could also be helpful.

The combination of deductive coding (informed by theoretical frameworks to guide barrier identification) and inductive analysis (to allow more granular content themes, unanticipated findings and patient insights to emerge) is a strength of this review. However, it is worth noting that the theoretical frameworks used in this review are limited in as much as they do not specify relationships between domains and hence the strength of the direct impact of barriers on behaviour is not known.

The majority of the studies included in this review were from North America. Additional UK studies are needed from the perspective of both patients and HCPs. Ideally, such additional research would focus on sampling the subgroups least likely to attend for screening.

Do the components of existing quality improvement interventions target the important theoretical determinants of diabetic retinopathy screening attendance? (seeChapter 5)

We hypothesised that QI interventions to increase DRS attendance would be more effective if they targeted the key barriers to/enablers of DRS attendance. To investigate this we integrated BCT components identified in existing QI interventions (seeChapter 3) with the identified important theoretical determinants of DRS attendance (seeChapter 4). Certain BCTs may be of greater or lesser relevance to addressing barriers/enablers within a given theoretical domain (i.e. may be of higher or lower theoretical coherence). We therefore examined the extent to which component BCTs in existing QI interventions are theoretically coherent with domains representing key barriers to/enablers of DRS attendance. This was achieved by applying published mapping tools that pair BCTs from taxonomies with domains from the TDF that experts agree represent theoretically coherent pairings.

We examined the extent of theoretical coherence for frequently used BCTs, identified in at least 10 intervention arms of existing QI interventions that target patients and/or HCPs (seeChapter 3). BCTs were classified as having high theoretical coherence if at least one domain with which they were paired in available mapping tools was classified as being important in the second review (seeChapter 4). Eight of the ten frequently used BCTs targeting patients had high theoretical coherence: (1) ‘feedback on outcomes (i.e. consequences) of behaviour (including biofeedback)’, (2) ‘credible source’ (i.e. persuasive argument presented by a respected person), (3) ‘prompts and cues’, (4) ‘social support (unspecified)’, (5) ‘restructuring the social environment’ (e.g. introducing a diabetes link worker), (6) ‘instruction on how to perform the behaviour’ (i.e. provide further details about how to make, accept or attend a DRS appointment), (7) ‘social support (practical)’ (e.g. assist with transport) and (8) ‘information about health consequences’ (e.g. provide an information leaflet on how DRS assists with the early detection and treatment of retinopathy). All seven frequently used BCTs targeting HCPs had high theoretical coherence: (1) ‘restructuring the social environment’, (2) ‘credible source’, (3) ‘adding objects to the environment’ (e.g. reminder systems), (4) ‘social support (practical)’, (5) ‘instruction on how to perform the behaviour’, (6) ‘prompts and cues’ and (7) ‘feedback on outcomes of behaviour’ (including biofeedback).

Our findings thus suggest that the majority of components that are frequently used in existing QI interventions do target the important determinants of DRS attendance. Although it was not possible to formally examine the association between theoretical coherence and effectiveness, the observation that nearly all frequently used BCTs associated with higher DRS attendance also had high theoretical coherence supports the notion that interventions are more likely to be effective if they target the key determinants of behaviour and behaviour change. Therefore, the design of future QI interventions, or the refinement of existing QI interventions, may benefit from the inclusion of the aforementioned BCTs with demonstrated effectiveness and high theoretical coherence.

However, existing QI interventions in this review were found to use a limited number of BCTs (median 4). Our mapping investigation revealed that there are numerous ‘missed opportunities’ for intervention design, in the form of BCTs that the mapping tools suggest are theoretically applicable to addressing the important domains representing identified barriers to/enablers of DRS attendance, but which are not frequently used in existing QI interventions. Missed opportunities for intervention design were identified for all six important TDF domains identified inChapter 4. These findings point to theoretically informed candidate BCTs that could be included in future QI interventions to increase the targeting of the important domains to DRS attendance. This could in turn increase intervention effectiveness and patient benefits. This is particularly true for the important theoretical domain ‘emotions’. None of the five BCTs that are agreed to be theoretically coherent with this domain was frequently identified in existing QI interventions (range 0–3 intervention arms). This suggests that there is a critical need for future QI interventions to include components that target the emotional influences on DRS attendance, such as the BCTs ‘reduce negative emotions’ (e.g. reassure the patient that any identified retinal pathology will be treated without delay), ‘information on emotional consequences’ (e.g. mention that a negative test will provide reassurance about eye health), ‘self-assessment of affective consequences’ (e.g. ask the person with diabetes to record their feelings after the DRS examination), ‘social support (emotional)’ (e.g. ask a friend to comfort the person with diabetes) and ‘conserving mental resources’ (e.g. arrange a DRS appointment at a time when the person with diabetes is less likely to be busy).

However, an important caveat to interpreting these findings concerns the applicability and validity of the methods and mapping tools used to pair BCTs with theoretically coherent TDF domains. Although this study is one of the first to provide an empirical test of the validity of available mapping tools, a number of limitations and discrepancies were identified in the tools, suggesting a need for further research to refine and expand these tools to enhance their validity and usability. The validity of any BCT and domain pairing recommended in each mapping tool depends on the nature of the behaviour being investigated and what intervention components are likely to be feasible, appropriate and acceptable to deliver in the context of DRS attendance. Lastly, our methods and available data did not enable investigation of which component BCTs may be more or less theoretically coherent for addressing barriers and enablers for different subgroups of patients attending DRS. As future evidence emerges concerning which barriers/enablers are of greater importance to different subgroups, there is the potential to apply the present mapping methods to develop more tailored interventions that include component BCTs that are theoretically coherent with the determinants of DRS attendance for specific patient subgroups.

Which quality improvement interventions designed to increase diabetic retinopathy screening attendance are most likely to be cost-effective and what are the potential cost consequences? (seeChapter 6)

At a baseline probability of DRS attendance of 0.7 and when the maximum that society is willing to pay for a QALY is £20,000, no QI components (coded using the modified EPOC Group taxonomy29) are likely to be cost-effective (judged on a probability of > 0.7) compared with no QI intervention. When the baseline probability of DRS attendance is decreased to 0.5, ‘team changes’ and ‘patient education’ have a probability of being cost-effective of > 0.7.

The probabilities of being cost-effective at a societal willingness-to-pay threshold of £20,000 per QALY were higher for BCTs (patient targeted or HCP targeted) than for QI components. Of the patient-targeted BCTs, ‘goal-setting (outcome)’, ‘feedback on outcomes of behaviour’, ‘social support’ and ‘information about health consequences’ had extremely high probabilities of being cost-effective compared with no BCT intervention. Of the HCP-targeted BCTs, ‘adding objects to the environment’ (e.g. decision support or monitoring tools such as diabetes passports) had a probability of being cost-effective of > 0.9. Nevertheless, the BCT components with the highest probabilities of being cost-effective were patient targeted. Likewise, the QI component with the highest probability of being cost-effective was ‘patient education’.

In sensitivity analyses, as the frequency of DRS was reduced, the cost-effectiveness of a BCT component compared with no BCT intervention increased. When the screening interval was increased to 3 years, all BCTs had a probability of cost-effectiveness of > 0.8 except for ‘instruction on how to perform the behaviour’ targeted at both patients and HCPs, ‘credible source’ and ‘restructuring the social environment’ targeted at patients, and ‘instruction on how to perform the behaviour’ and ‘feedback on outcomes of behaviour’ targeted at HCPs.

When the age of the cohort was reduced from 64 years (the base case) to 56 years then the BCT components all became less cost-effective compared with no BCT intervention. However, for some BCTs, such as ‘information about health consequences’ targeted at the patient, the change was very small (the probability that this BCT is cost-effective remained at 0.998, which is extremely high).

The economic analysis utilised the best available effectiveness and resource use data to evaluate the likelihood that individual QI and BCT components are cost-effective compared with no intervention and with each other at willingness-to-pay thresholds for a QALY normally considered to be acceptable. Resource use associated with the main intervention in each included study was recorded and valued so that both the effects of the QI and BCT components on DRS attendance and resource use could be estimated. Data from a recently published cost–utility model of screening frequency, conducted as part of a previous NIHR HTA study,224were used to model the long-term outcomes of screening attendance and non-attendance.

Interventions were assumed to target all people eligible for screening, that is, all people with diabetes. It may be possible to target a subset of the eligible population who would benefit from encouragement to attend screening. Included studies may have targeted different groups. This would require a baseline probability of screening uptake specific to that population, an intervention effect estimate specific to that population and cost estimates that account for a smaller targeted population. These data were not available from this study. 224

The economic analysis explores which individual BCTs and QI components are the most likely to be cost-effective given the available evidence; it does not consider the cost-effectiveness of multicomponent interventions. In reality, the interventions designed to encourage screening uptake are likely to be complex, incorporating several BCTs or QI components. The results of the economic analysis are therefore less useful in informing which interventions to adopt but provide an indication of which QI and BCT components are worth considering as a starting point in the design of a DRS attendance intervention.

The effectiveness evidence used in the economic model was derived from a multivariate meta-regression with all of the QI components and BCTs included as explanatory variables. These analyses are essentially based on observational data. The potential for the existence of uncontrolled confounding factors may affect the power of the statistical model to predict outcomes in other studies and settings. Furthermore, an additive model is assumed and the interaction effects between QI components and BCTs are unknown and are not captured in the model. This statistical approach is different from the analysis conducted as part of the phase 1 systematic review. The two approaches have different interpretations. The univariate meta-regression analysis conducted as part of the phase 1 intervention review compared studies that used multifaceted interventions that incorporate a particular QI/BCT component with studies using other multifaceted interventions that do not contain that component. This analysis estimates the average effect across a range of complex interventions. The additive (multivariate) model used in the economic analysis estimates the effects of individual QI components by adjusting for the other components in the intervention. There is far greater variation in the univariate model than in the multivariate model and some effects are significantly different. For example, using a RD outcome variable, the patient-targeted BCT ‘restructuring the social environment’ is ranked six out of 17 in the univariate analysis of BCTs with a coefficient of 0.051, whereas it is ranked 17 out of 17 in the multivariate model with a coefficient of –0.19.

There is considerable heterogeneity in the data and most of the studies were not conducted in the UK; therefore, it is difficult to make specific conclusions regarding people with diabetes in the UK. Consequently, the economic analysis was limited to providing evidence on the likelihood that individual QI components and BCTs are cost-effective to help prioritise future research by designing and studying interventions based around the QI components and BCTs most likely to be cost-effective. It is possible that a complex intervention designed to incorporate a BCT or QI component with a high probability of being cost-effective may need to include a BCT or QI component that is less likely to be cost-effective. It may be possible to develop an intervention based around a specific QI component or BCT that is more cost-effective than indicated here; the cost thresholds below which an intervention may be cost-effective may be a guide to the cost per patient of an intervention.

The evidence on the resources used to provide the QI/BCT components of interventions was based on an ordered logistic regression with all of the QI components and BCTs included as explanatory variables. A similar model was used to estimate the effectiveness of QI/BCT components and the same limitations apply. With respect to resource use, the weights used were developed in an explicit fashion and care was taken to check the scorings provided by the original raters. However, the resource use ordered rankings were derived using a small sample of cost estimates based on the intervention descriptions. Uncertainty associated with these cost estimates was not included in the analysis and further work could usefully be undertaken to provide a wider sample of cost estimates to place a value on the resource use rankings.

Overall, the economic evaluation does have a number of limitations (described in more detail inChapter 6), primarily linked to the limitations of the evidence base. However, the economic analysis identified the most promising BCT/QI components to take forward to enable further development of the evidence and theory-based interventions to increase attendance for DRS.

Recommendations for practice and applicability to the UK setting

The findings suggest that, among interventions that target the patient, promotion of self-management (provision of equipment or resources to promote self-management) and patient education (interventions designed to promote greater understanding of diabetic retinopathy) are more likely to be effective at increasing levels of DRS attendance. For interventions at a system or organisational level, team changes (changes to the structure or organisation of the primary health-care team, e.g. adding a team member or shared care, use of multidisciplinary teams or expansion or revision of professional roles) and the establishment of an electronic diabetes registry (e.g. general electronic medical record system or electronic tracking system for patients with diabetes) are associated with higher levels of DRS attendance.

Including behavioural interventions to support the uptake of DRS services could improve the uptake of DRS. Such interventions include providing feedback on the consequences of attendance or non-attendance; encouraging social (interpersonal) support to attend; providing more information on diabetic eye disease, including information about the health consequences, and on the process of screening; introducing reminder systems; and ensuring that patient information is provided by a credible source, such as respected national clinical guidelines. These interventions can be delivered at a patient level and at the level of the health system, including at the level of HCPs. We identified that these interventions are effective and are also likely to target the important factors associated with attendance at DRS.

Considering cost-effectiveness and theoretical coherence (seeChapter 5), we have identified BCTs that could be worthwhile for services with lower attendance levels. These are ‘feedback’ and ‘information about health consequences’, for example providing information on diabetic retinopathy and the consequences and benefits of DRS, explaining the difference between DRS and attendance for regular eye tests (e.g. enhanced patient information sheets), introducing processes to facilitate social support to attend and improving the screening environment [introducing processes to improve convenience for patients, e.g. online management/booking systems or monitoring tools (e.g. diabetes passports)].

Interventions should be judged by their cost-effectiveness, but it is also useful to know the relative resource implications that are likely to be associated with QI components and individual BCTs. The QI component with the highest probability of being cost-effective was ‘patient education’, but in general QI components were unlikely to be considered cost-effective. In terms of BCTs, among patient-targeted techniques, ‘problem-solving’, ‘goal-setting (outcome)’, ‘social support (practical)’ and ‘instruction on how to perform the behaviour’ were associated with relatively high resource use; ‘feedback on outcomes of behaviour/biofeedback’ and ‘restructuring the social environment’ were associated with moderately high resource use; and ‘social support (unspecified)’, ‘information about health consequences’, ‘prompts/cues’ and ‘credible source’ were associated with relatively low resource use. Among the HCP-targeted BCTs, ‘social support (practical)’, ‘prompts/cues’ and ‘restructuring the social environment’ were associated with relatively high resource use; ‘instruction on how to perform the behaviour’ and ‘credible source’ were associated with moderately high resource use; and ‘feedback on outcomes of behaviour/biofeedback’ and ‘adding objects to the environment’ were associated with relatively low resource use.

Implications for the UK NHS

In the UK, on average, 20% of those offered a DRS appointment fail to attend, with wide geographical variations in screening uptake. 20,23,24Living in areas of high social deprivation, a younger age and having a longer duration of diabetes are associated with lower attendance rates. Ethnicity is also an important determinant of DRS attendance, with black, Asian and minority ethnic groups less likely to attend for screening. How can attendance be improved for the missing 20%?

Integrating our findings to the specific context of the UK National Diabetic Eye Screening Programme, during a knowledge exchange event (seeAppendix 2), we agreed that interventions to increase the uptake of DRS should focus on low or non-attenders, especially those with multiple missed appointments, rather than attempting to achieve 100% for all. Special consideration is needed of working people, people with mental health issues, vulnerable groups including those on low incomes or those with special needs and people with multiple conditions and thus competing demands on their time. It was noted that there is great diversity across the UK and that initiatives should be tailored to local needs. The feasibility and ethical issues around focusing on low attenders and vulnerable groups need to be explored in future research.

Based on the study findings on key determinants of attendance, potentially effective and likely cost-effective interventions and the views of stakeholders, the following strategies could improve attendance at DRS.

• Patient-focused interventions: improve patient information regarding DRS to improve patients’ knowledge about the condition and the screening process, in particular the differences between a regular eye test and a DRS appointment; work with other health professionals to use every potential health-care visit to provide information about eye health; work closely with diabetes teams and patient groups to review and update existing diabetes education programmes for both type 1 [Dose Adjustment for Normal Eating (DAFNE)235] and type 2 (X-PERT Health236and DESMOND237) diabetes patients so that they include relevant content on eye health and visual loss and its relation to general diabetes care; design and deliver media campaigns to ensure that the message is targeted and relevant to people with diabetes; strategies could include components that target the emotional influences on DRS attendance, which we identified as being under-represented in current interventions.

• Team changes: expand the role of retinal screeners so that they provide patients with information about the findings at the time of screening and feed back to patients how the impact of these on vision can be lessened; involve pharmacists to provide information, including a point of contact, to patients when they attend a pharmacy for any diabetes-related medications or supplies; reassess and realise the importance of the GP in terms of information provision and encouraging patients to attend for screening [GPs (and general practice staff) are among the most trusted professions involved in patient care].

• Environment: restructure the screening setting, for example the use of ‘one-stop shops’, so that all appointments related to diabetes occur in one place and during one patient visit; flexible timing of appointments, including appointments outside normal working hours; provide reminders; enable patients to self-book appointments, including the ability to cancel and reschedule; manage and update existing electronic registers in terms of completeness both by fostering better relationships with other health-care providers and through the use of health system intelligence such as electronic care records.

• There were no data on the effectiveness of financial incentives to encourage DRS attendance. The UK stakeholder group proposed this as a possible intervention to improve uptake for those with a history of poor attendance, especially in areas where socioeconomic deprivation is high and for those for whom loss of income while attending appointments is a barrier. An abstract238for a UK-based RCT investigating the use of financial incentives to improve DRS attendance was identified, which did not meet our inclusion criteria. This study has now reported in full. 239

Recommendations for future research

The evidence synthesis identified interventions that could improve the uptake of DRS in different health systems including in the context of a national DRS programme as in the UK. We have identified a number of knowledge gaps that should be addressed by future research. We were unable to identify any RCTs that evaluated the effectiveness of QI interventions for ongoing attendance for DRS. There has been considerable debate on the optimal screening interval for diabetic retinopathy. A recent UK cost-effectiveness study concluded that annual screening was not cost-effective and recommended the introduction of personalised screening intervals. 224Nonetheless, periodic screening is essential to detect sight-threatening retinopathy and future trials should be of a sufficient length to capture multiple screening episodes (e.g. over 5 years). The longer timescale would also allow other important longer-term outcomes to be evaluated, for example the effect of an intervention on the progression of diabetic retinopathy and the impact on visual function.

Although the included studies provided data on a large number of participants, the characteristics of the study participants were often poorly described and data were not provided on the relative effectiveness of interventions in population subgroups. These data are important to illuminate health inequities. It was disappointing that < 5% of the included trials were conducted in the UK.

Through exploratory analyses of evidence from both randomised and non-randomised studies, we identified BCTs that we would regard as ‘best bets’ for evaluation in future RCTs, based on their theoretical coherence with identified barriers and enablers and their clinical effectiveness and cost-effectiveness. However, the complexity of the QI interventions made it difficult to completely disentangle the BCT components to identify the optimal combination of BCT components that would lead to a greater likelihood of DRS attendance. Future research could test particular BCTs and BCT combinations in adequately powered studies using head-to-head comparisons, including those theoretically coherent BCTs identified in our mapping exercise that have not been commonly used in previous RCTs. Particular gaps occur in relation to BCT components that target the emotional barriers to and enablers of screening attendance.

Research priorities

The evidence from this study can be used to inform the development of a future UK-based RCT evaluating QI interventions that use BCTs to improve DRS attendance. The design of the RCT can be framed using thepopulation,intervention,comparator,outcome(s) andtime frame (PICOT) format. Based on our findings we suggest the following design.

• Population: the evidence on cost-effectiveness suggests that the focus should be on groups with a low baseline level of attendance.

• Intervention: given that we identified that general (untargeted) QI interventions were statistically indistinguishable from targeted interventions in terms of improving DRS attendance, both types of intervention could be considered. Ideally, the intervention should be multifaceted and utilise BCTs that target the patient (e.g. ‘feedback on outcomes of behaviour’, ‘social support’ or ‘information about health consequences’) or the HCP (e.g. ‘adding objects to the environment’, such as decision support or monitoring tools such as diabetes passports). A number of additional BCTs that are not included in existing interventions may also be relevant to addressing important theoretical determinants of screening attendance. In particular, there are gaps around the inclusion of intervention components to address the emotional barriers to and enablers of screening attendance (such as anxiety about the experience of the screening test itself or fear of the screening result and sight loss).

• Comparator: the findings of our evidence synthesis were based on comparisons with usual care. However, in included studies, usual care was either poorly described or not described at all. It is therefore important that usual care is specified in sufficient detail so that BCTs present in the control arm can be identified.

• Outcome(s): the primary outcome should be attendance for DRS. Secondary outcomes could include other quality indicators of diabetes care, for example HbA_1clevels, screening for other microvascular complications, ongoing attendance for DRS and the likely cost-effectiveness of the intervention.

• Time frame: the study should be of a sufficient duration to capture ongoing attendance for DRS. The duration of the trial should be a minimum of 24 months (and ideally longer).

Before conducting the proposed RCT we recommend the following programme of research.

• Identify the population with a low rate of DRS attendance: review electronic systems for recording DRS programme attendance, audit attendance records and analyse the characteristics of non-attenders, with special emphasis on those who have missed several appointments as these patients might be at the highest risk for visual loss.

• Conduct a qualitative exploration of the key determinants of attendance in subgroups of low attendees to inform the design of meaningful interventions, with attention to ethical issues around identifying subgroups and targeting interventions to these subgroups.

• Conduct empirical research to evaluate the validity of mapping tools to facilitate the selection of theoretically coherent BCTs (i.e. those that target important barriers and enablers) for future research and intervention design.

Facilitators

• Katherine Cowan (James Lind Alliance).

• Michael Bowen (College of Optometrists).

• Sally Crowe (Crowe Associates Ltd).

Research team members

• John G Lawrenson (City, University of London).

• Jill J Francis (City, University of London City).

• Fabiana Lorencatto (City, University of London).

• Ella Graham-Rowe (City, University of London).

• Tunde Peto (Queen’s University Belfast).

• Stephen Rice (Newcastle University).

• Luke Vale (Newcastle University).

• Patricia Aluko (Newcastle University).

• Jennifer M Burr (University of St Andrews).

Stakeholder members

• Andrew Crowder (Diabetes UK, Wales).

• Gozie Joe Adigwe (RNIB, Scotland).

• Simon O’Neill (Diabetes UK).

• Caroline Styles (NHS Fife, Scotland).

• Helen Lee (RNIB, London).

• David Owens (Diabetes Research Unit Cymru, Wales).

• Phil Gardner (British Association of Retinal Screening).

• Natalie Owen (Department of Health).

• David Keene (NHS City & Hackney Clinical Commissioning Group).

• David Parkins (College of Optometrists).

Patients

• Vivienne Ruddock, Sheila Burston, Melissa Louis and David Beck.

1. Introduction

This event on the 12 January 2017 brought together the research team, people with diabetes and experience of DRS, health professionals and service providers with an interest in DRS to hear about the early results of this project. It was also an opportunity for participants to discuss their experiences of barriers/enablers that affect regular DRS attendance and members of the research team were particularly interested in hearing ideas about how low rates of attendance could be improved.

The research team presented summaries of the methods and results in the morning, generating interesting questions and debate, and three smaller discussion groups were facilitated in the afternoon to reflect more deeply on the project results and capture ideas and feedback from participants. Discussion groups were facilitated by independent (from the research team) consultants, who took notes, and all discussions were recorded for accuracy.

This report describes these conversations and draws together the themes across the discussion groups, using the notes and recordings as primary material. In the final section the implications of the findings for UK screening programmes are summarised and suggestions are made for future research.

2. Discussion in groups

3. Themes across the three discussion groups

4. Summary