Intravitreal ranibizumab versus aflibercept versus bevacizumab for macular oedema due to central retinal vein occlusion: the LEAVO non-inferiority three-arm RCT

Primary objectives

• To determine whether or not bevacizumab is non-inferior to ranibizumab in treating visual loss due to MO secondary to CRVO.

• To determine whether or not aflibercept is non-inferior to ranibizumab in treating visual loss due to MO secondary to CRVO.

Secondary objectives

• To determine the difference between arms in terms of mean change in BCVA at 52 weeks.

• To determine the difference between arms in the proportion of participants with ≥ 15 ETDRS letter improvement (appreciable visual gain), ≥ 10-letter improvement, < 15-letter loss and ≥ 30-letter loss (severe visual loss) at 52 and 100 weeks.

• To determine the difference between arms in the proportion of participants with ≥ 73 ETDRS letters or > 6/12 Snellen equivalent (i.e. approximate driving visual acuity), ≤ 58 ETDRS letters (≤ 6/24) and ≤ 19 letters (≤ 3/60) [Certificate of Vision Impairment (CVI) partial and severe visual impairment] at 52 and 100 weeks.

• To determine the difference between arms in the mean change in OCT central subfield thickness (CST) and macular volume at 52 and 100 weeks.

• To determine the difference between arms in the proportion of participants with an OCT CST of < 320 µm [as measured with the Spectralis^® (Heidelberg Engineering, Inc., Franklin, MA, USA) or equivalent] at 52 and 100 weeks (key guide to subsequent NHS clinical practice).

• To determine the differences between arms in the mean number of intravitreal injections given to each participant at 100 weeks.

• To determine any differences in the relative effectiveness of the investigational treatments and comparator on quality of life and resource use, reported as incremental cost-effectiveness ratios (ICERs), at 52 and 100 weeks.

• To detect any differences in the prevalence of local and systemic side effects at 100 weeks.

• To determine the differences between arms at 100 weeks in the proportion of (1) persistent non-responders who develop a change in retinal non-perfusion, compared with screening, and (2) participants who develop anterior and posterior segment neovascularisation.

• To determine the differences between arms in terms of mean change in BCVA at 100 weeks.

• To determine the differences between arms in changes in area of non-perfusion at 100 weeks and OCT anatomical features from baseline to 100 weeks.

Selection of participants

Recruitment

The trial recruited participants from 44 UK ophthalmology centres over 24 months. Recruitment was competitive; however, each site was allocated a minimum target number of participants to recruit and was encouraged to exceed this if possible. Sites were set up strategically: larger sites with greater capacity were initiated first to maximise early recruitment and to ensure that the recruitment period was fully utilised. Eligible patients were invited to participate via their local clinics, or in an invitation letter. At each site, participants were identified from subspecialty retinal, general and eye casualty clinics. Once identified, potential participants underwent a clinical examination, followed by discussion of the trial with an experienced trial clinician, and were provided with the patient information sheet. 2

Informed consent procedure

The principal investigator or designated subinvestigator was responsible for ensuring that a patient was fully consented after being provided with an adequate explanation of the aims, methods, anticipated benefits and potential hazards of the trial. Patients were advised that any data collected would be held and used in accordance with the Data Protection Act 1998. 64 Patients were given at least 24 hours after receiving the patient information sheet to consider taking part. The principal investigator or designee recorded in the medical notes the date when the patient information sheet was given to the patient and the facts that patients were under no obligation to enter the trial and that they could withdraw at any time without giving a reason. No clinical trial procedures were conducted before consent was taken from a participant; consent denoted enrolment in the trial. A copy of the signed informed consent form was given to the participant. The original signed form was retained at the trial site and a copy was placed in the medical notes. If new safety information resulted in significant changes in the risk/benefit assessment, or if there were significant changes to the protocol or patient information sheet, participants were consented again as appropriate.

Randomisation

Only one eye of each participant was randomised to the trial. In 95% of cases, one eye was affected by CRVO and so was the ‘worse-seeing eye’ and was randomised. On rare occasions, participants had bilateral CRVO that met the eligibility criteria. In these cases, the worse-seeing eye was randomised unless the participant opted for randomisation of the ‘better-seeing eye’. The plan was to recruit 459 adult participants with MO due to CRVO and to randomise them 1 : 1 : 1 at the level of the individual using the method of minimisation incorporating a random element. The three stratifying factors were (1) visual acuity, stratified by screening BCVA letter score [of ≤ 38 (approximate Snellen equivalent: ≤ 6/60), 39–58 (approximate Snellen equivalent: 6/48 to 6/24) or ≥ 59 (approximate Snellen equivalent: ≥ 6/18)]; (2) duration of disease, from date of CRVO diagnosis to commencement of therapy (< 3, 3–6 or > 6 months); and (3) treatment naive versus previous treatment. Each participant was randomised to one of three arms: bevacizumab, aflibercept or ranibizumab. 2

A patient identification number (PIN) was generated by registering a patient on the MACRO eCRF system (InferMed Macro; Elsevier Ltd, Amsterdam, the Netherlands), after consent had been obtained. Randomisation was carried out in a bespoke web-based randomisation system hosted at the King’s Clinical Trials Unit (KCTU). A unique PIN was generated in the MACRO program; this was recorded on all source data worksheets and was used to identify a participant throughout the trial. 2,63 The trial manager allocated all authorised site staff a username and password for the randomisation system. All authorised staff members, who were typically the principal investigator or designee, logged in to the randomisation system and entered a participant’s details, including the unique PIN. Once a participant had been randomised, the system automatically generated e-mails to key staff in the trial. Unmasked e-mails sent to site pharmacies alerted them to a participant’s treatment arm: ranibizumab, aflibercept or bevacizumab. The pharmacy department used the e-mail to cross-check the trial prescription to ensure that the correct medication was dispensed for the correct participant. Additional masked e-mails were generated from the randomisation system and sent to key trial site staff,63 and unmasked e-mails were sent to the emergency unmasking service (ESMS Global Ltd, London, UK) and unmasked trial management staff. 2

Masking of treatment allocation

In randomisation process, only the pharmacy at a local trial site was informed by e-mail of a subject’s treatment allocation; a copy of the e-mail was sent to the emergency unmasking service (ESMS Global Ltd) and to unmasked trial management staff. The trial drug that a participant received was transferred to the dedicated injection room in an opaque masking bag designed to securely and safely transport medication. A unique seal was attached to the bag before it left the pharmacy. The bag had a safe zipped compartment containing a printed form detailing a participant’s unique PIN, their date of birth, the date the drug was dispensed and the injection batch number. Before a participant entered the injection room, the unmasked injector broke the seal and took the drug out of the masking bag. Bevacizumab was provided in a prefilled syringe, but ranibizumab and aflibercept were provided in a vial and drawn into a syringe by the unmasked injector. The syringe was placed on the injection trolley out of view of the participant, who was then invited into the room and asked to lie on the bed, and then received the injection. During the trial the manufacturer of ranibizumab began to provide the drug, in a unique prefilled syringe and vials ceased to be available. In this situation, the unmasked injector took care not to allow the participant to see the syringe either before or after the injection had been given. This was achieved by administering the injection while the participant was lying down and the injection was given via the pars plana in any quadrant of the eye, with the syringe brought to and taken away from the injection site via a participant’s inferotemporal field of vision so that it did not pass across their line of sight. The unmasked injector signed the source notes to the effect that the treatment in the masked bag had been administered to the participant, without specifying the treatment, and also signed the printed form that was in the masking bag. The empty syringe with needle and vial were disposed of in the injection room. The masking bag and completed printed form were returned to the pharmacy. The outer packaging of the drug was disposed of in the injection room. 2

The clinical assessment team, including the site principal investigator, optometrist (i.e. assessor of the primary outcome), site trial co-ordinator, clinical investigator, clinical assessment trial nurse and ophthalmic technician, remained masked throughout the trial, as there was no record of a participant’s treatment arm in the source notes or the case report form (CRF). Similarly, co-ordinators or administrators completing questionnaires with participants in person (or, in extreme circumstances, only over the telephone at specific time points) had details of a participant’s PIN only. If, at any time, information regarding treatment allocation was shared with the outcome assessors, this was recorded in the trial master file, and the person(s) involved met with the site principal investigator to ensure that no repetition occurred and undertook not to convey this information either to the participant or to others involved in the project. Certain secondary outcomes (e.g. interpretation of FFA) occurred at the remote Network of Ophthalmic Reading Centres (NetwORC) UK (Belfast, UK), where the assessors were masked to the treatment allocation. These masking procedures avoided both performance and detection bias. We have described the completeness of outcome data for each outcome, including any unmasking in error, reasons for attrition and exclusions from the analysis. 2 The trial statisticians had access to the accumulating outcome data that were required for reporting to the Data Monitoring and Ethics Committee (DMEC). Both trial statisticians attended both the open and the closed DMEC meetings.

Screening and baseline assessment

A patient had to receive the patient information sheet not later than 24 hours before the screening assessment. The screening and baseline visits could be undertaken on the same day, provided that all test results were available. A patient could return within 10 days of screening for the baseline assessment, at which point the screening procedures were still valid and were not repeated at baseline (see Appendix 3, Table 29).

Milestone and non-milestone visits

Trial milestone assessments, when key research data were collected, occurred at baseline and at weeks 12, 24, 52, 76 and 100. These visits, as well as treatment visits at weeks 4 and 8, were calculated and agreed with a participant prior to randomisation (with flexibility of 0 to 14 days for weeks 4, 8 and 12, and of –14 to 14 days for weeks 24, 52, 76 and 100, from the date of randomisation). It was mandatory for all participants to attend all milestone visits, even if a milestone visit fell < 4 weeks after a treatment visit or if a participant was being followed up every 8 weeks and the next milestone visit fell during the 8-week interval. The intervening trial treatment visits were deliberately flexible to allow normal clinical practice treatment follow-up to be accommodated. All data from the trial milestone visits were entered into the eCRF. For regular treatment visits, only the following information was entered into the eCRF: BCVA; OCT CST; whether or not an injection was given; and, if no injection was given, the reason why. At milestone visits, refracted visual acuity was tested and health economic questionnaires were completed; colour photography was undertaken at baseline and at weeks 52 and 100; and FFA was undertaken at baseline and at week 100, in addition to the clinical examination and OCT tests performed at all other trial visits (see Appendix 3, Table 29).

Participant demographics, medical and ophthalmic history

This information was retrieved from the participant, from hospital medical records or from a general practitioner. Data included age, sex and ethnic background. Data were also collected on clinically relevant medical history and management in the preceding 24 months, and on any ocular history and treatment. 2

Visual acuity tests

Visual acuity tests were performed by a certified optometrist in a certified visual acuity testing lane using validated ETDRS vision charts and standard operating procedures. 65,66 Refracted visual acuity was carried out in both eyes at screening,63 at weeks 12, 24, 52, 76 and 100, and at the point of withdrawal. For all other visits, the visual acuity was tested with the previous most recent protocol refraction. Visual acuity examiners were masked to the treatment. The visual acuity scores were recorded in the eCRF2 (see Appendix 4).

Standard ophthalmic examination

A standard ophthalmic examination using slit-lamp biomicroscopy included an undilated examination for neovascularisation of the iris (NVI), RAPD and tonometry in both eyes at all visits. Dilated fundus examination was performed in both eyes at all milestone visits (i.e. at screening, at baseline, at weeks 12, 24, 52, 76 and 100, and at the point of withdrawal). At all other visits, dilated fundus examination was performed in the trial eye and, at the discretion of the investigator, in the non-trial eye. Gonioscopy, if indicated, was carried out prior to dilatation at any visit. 2

Spectral-domain optical coherence tomography

The CST and total macular volume in both eyes were recorded in the eCRF from the spectral-domain optical coherence tomography (SD-OCT) thickness map at every visit, and, if applicable, at the point of withdrawal. 63 Any SD-OCT machine could be used for the trial, but the same model of SD-OCT machine had to be used for each individual throughout the period of the trial. SD-OCT images at screening and at weeks 52 and 100 only were transferred to and read by masked graders at the independent NetwORC UK. NetwORC UK provided each site with a trial imaging protocol on how to acquire SD-OCT images, colour fundus photographs and fundus fluorescein angiographs and how to transfer these to NetwORC UK to them. Initial grading of all OCT images at baseline and at weeks 52 and 100 was performed by NetwORC UK. The grading took into account intraretinal oedema, classified as diffuse, cystic or mixed; determined subretinal fluid as being present or absent; and determined vitreoretinal interface abnormalities as being present (as either an epiretinal membrane or vitreomacular traction) or absent. Following the contract variation, additional grading parameters were assessed at NetwORC UK in collaboration with specialised retinal graders at Moorfields Eye Hospital, utilising additional definitions and analyses that had been developed while the trial was in progress. 1,67,68 Only images captured using a Spectralis OCT machine had sufficient detail to support the enhanced grading definitions. Retinal morphology was assessed using the Spectralis^® Heidelberg Macular Raster OCT device (Heidelberg Engineering, Inc.) of 31 line scans, 30 × 25 mm in size, at an interscan distance of 240 µm or the equivalent for alternative devices. MO was graded using the entire line-scan series and the central 1500 µm, that is seven scans were employed for vitreomacular interface abnormality and subretinal detachment or equivalent. The remaining parameters were graded using the central 1000 µm only, that is central five-line scans only. A magnification of 300% was used to assess the ellipsoid zone (EZ), disorganisation of the retinal inner layers (DRIL)67,68 and hyperreflective foci (HRF),69,70 with 100% magnification for the remaining parameters. HRF, external limiting membrane (ELM), EZ and cone outer segment tips (COSTs) were graded as positive only if the foveal line showed involvement of the foveal depression such that it was distorted, lessened or absent. 2 For the grading of normal and abnormal individual morphological features, see Appendix 5, Specific grading of individual morphological optical coherence tomography features, and Figures 22 and 23.

Colour fundus photography

Non-stereo, seven-field conventional or wide-angle colour fundus photography (CFP) was performed at screening and at weeks 52 and 100 in the trial eye. CFP confirmed the diagnosis of CRVO and assisted interpretation of features identified by FFA, for example to differentiate between non-perfusion and masking due to haemorrhage. If applicable, CFP was also performed at the point of withdrawal, and at any other trial visit, as per investigator discretion. Colour fundus photographs were transferred to and read by masked graders at the independent NetwORC UK. Either a colour camera capable of taking seven-field colour fundus photographs or a wide-angle system was used, but the same model of camera was used for each individual throughout the trial. The colour photographs were graded by the NetwORC UK. 2

Fundus fluorescein angiography

Non-stereo, seven-field conventional or wide-angle FFA was performed at screening and at week 100 in the trial eye. Any FFA system capable of taking seven-field FFA pictures or a wide-angle system was allowed, but the same system had to be used in the same individual throughout the trial. 2 FFA was used to quantify the degree of retinal ischaemia and for identification of retinal neovascularisation (see Appendix 5, Fundus fluorescein angiography grading). Pseudo-anonymised FFA images were transferred to NetwORC UK, where the standard NetwORC UK 13-sector grid (see Appendix 5, Figure 24) was applied over the wide-angle or montaged seven-field angiography pictures at baseline and at 100 weeks. The first 100 gradings were double-graded. Discrepancies were adjudicated. Subsequently, one in every eight gradings was double-graded. Kappa values for key fields (e.g. detection of new vessels on the disc and new vessels elsewhere) were required to be > 0.8. Any graders who did not achieve this were required to undergo additional training. Each sector in the grid was semiquantified in terms of percentage of non-perfusion (nil, 1–25%, 26–50%, 51–75% and 76–100%), and all available sets of images were analysed to identify how many participants in each arm had experienced a two-step increase (e.g. zero to 26–50%, or 26–50% to 76–100%) in one to five or more sectors (see Appendix 5, Figure 24). This technique was used in preference to the ischaemic index, which estimates the ratio of ischaemic to total retinal area but is very susceptible to image quality and is applicable to wide-angled images only. 22 Therefore, during the trial we used the concentric rings method, which displays superimposed concentric circles, centred on the fovea. 23,71,72 The innermost circle was 1 disc diameter (DD) in size, and is not graded as it represents the foveal avascular zone. The second circle, representing the macular ring (ring M), has a radius of 2.5 DD. Each of the subsequent rings (rings 1, 2, 3 and 4) is placed at increments of 2.5 DD in radius from the foveal centre. Each of these rings is subdivided into 12 equal segments. 23 To calculate the size of the concentric rings required, we assumed that the mean axial length was 24 mm, and excluded 2 mm from this to account for the cornea and part of the anterior chamber. In the model eye, the radius was 11 mm (diameter 22 mm); therefore, the full circumference would have been 69.1 mm (π = 3.142). The wide-angled imaging system (Optos^®; Optos, Inc., Marlborough, MA, USA) was able to image 200 degrees of the retina; we used this to calculate the average diameter of retina obtained in a single central image. This was calculated to be 38.4 mm. Using the DD of 1.8 mm, this meant that the diameter of the image was 21.3 DD. A diameter of 21.3 DD resulted in the need for a macular ring plus three/four further rings. 23 Based on our validation study, we identified that ring 4 was gradable, but the superior and inferior segments of rings 3 and 4 were ungradable because the ultra-wide field image had better clarity in the horizontal meridian. For details of this method, see Appendix 5, Figure 25.

Health economic questionnaires

The following quality-of-life and resource use questionnaires were administered at baseline, at 12, 24, 52, 76 and 100 weeks, and at the point of withdrawal: the National Eye Institute Visual Function Questionnaire-25 items (VFQ-25), EuroQol-5 Dimensions (EQ-5D), EuroQol-5 Dimensions with vision bolt-on (EQ-5D-V), and a bespoke resource use questionnaire [see www.journalslibrary.nihr.ac.uk/programmes/hta/119203/#/documentation (accessed 14 July 2020)].

Treatment allocation guess form

Participants and masked optometrists were asked to complete a treatment allocation guess form at week 100, or at the point of withdrawal, to assess how well participant and assessor masking worked in the trial. 2

Definition of the end of the trial

Participants were enrolled in the trial for approximately 100 weeks from the point of randomisation. The end of the trial was defined as the last participant’s last trial visit.

Treatment schedule

After mandated administration in all three trial arms at baseline and at 4, 8 and 12 weeks, further pro re nata intervention was administered at weeks 16 and 20 if re-treatment criteria were met and if visual acuity was ≤ 83 letters.

Regardless of whether a treatment was given, the participant was reviewed in 4 weeks. From weeks 24 to 96, the interval was initially 4 weeks (with a visit window of –14 to 14 days), with the potential for the interval to increase to 8 weeks (with a visit window of –14 to 14 days) if criteria for ‘stability’ were achieved. ‘Stability’ was defined as three successive visits from week 16 onwards at which treatment criteria were not met, and so the first time at which treatment could be deferred for 8 weeks was week 24.

Similarly, ‘success’ was defined as an ETDRS letter score of > 83 letters, and if this was present at any re-treatment visit from week 16 onwards, treatment was not given at that visit and the participant was reviewed subsequently. The review occurred 4 weeks later if the initial visit was at 16 weeks, 20 weeks or any other visit if treatment had been given at this or the preceding visit. If no treatment had been given at these two visits, the participant was reviewed 8 weeks later. If, at any subsequent visit, re-treatment criteria were met and BCVA was ≤ 83 ETDRS letters, then re-treatment was commenced (Figure 1). At each visit between weeks 24 and 96 inclusive, ‘non-responder treatment suspension’ criteria could be met. If so, the principal investigator, or their designee, at their discretion, could suspend treatment to prevent therapy in a participant who had not responded to at least their last three injections. If the criteria for restarting therapy after ‘non-responder treatment suspension’ were met, then the participant had to resume therapy. If re-treatment criteria were met at one of the visits that took place every 8 weeks or at an unscheduled visit, then visits every 4 weeks were resumed. Treatment could be ‘deferred’ in certain circumstances, but the participant was asked to still attend the milestone visits.

FIGURE 1.

Re-treatment algorithm for weeks 24–96 of the trial.

Re-treatment criteria

Criteria were met if one or more of the following was present:

• a decrease in visual acuity of ≥ 6 letters between the current and most recent visit, attributed to an increase in OCT CST

• an increase in visual acuity of ≥ 6 letters between the current and most recent visit

• OCT CST of > 320 µm (on Spectralis, or of > 300 µm on other machines) because of intraretinal or subretinal fluid

• an OCT CST increase of > 50 µm from the lowest previous measurement.

Comparator: ranibizumab (0.5 mg/0.05 ml)

Ranibizumab is a humanised recombinant monoclonal antibody fragment that binds to VEGF A, preventing receptor interaction and blocking downstream action of VEGF, that is increased vascular permeability, leading to MO in CRVO. It is licensed by the EMA, and NICE has recommended it for use in the treatment of nvAMD, DMO and RVO. NICE TA28313 for MO due to RVO was issued in May 2013. Ranibizumab has been the mainstay of routine clinical care for this condition since the third quarter of 2013 and was the comparator in this trial. It was supplied to each site hospital pharmacy directly from the manufacturer as a part of routine hospital stock. 2

Intervention: aflibercept (2.0 mg/0.05 ml)

Aflibercept is a fusion protein that includes the key binding domains of human VEGF receptors 1 and 2 with human IgG Fc and acts as a dummy receptor for all VEGF isoforms and placental growth factor, preventing increased permeability and MO in CRVO. At initiation of this trial, it was licensed by the EMA, and NICE has recommended it for nvAMD. NICE TA30512 was published in February 2014; NICE recommends this drug as first-line use for CRVO-related MO. Aflibercept was supplied in a glass vial to each site hospital pharmacy directly from the manufacturer as part of routine hospital stock. 2

Intervention: bevacizumab (1.25 mg/0.05 ml)

Bevacizumab is a full-length humanised monoclonal antibody that binds to VEGF A, forming a protein complex incapable of binding to the VEGF receptor, thus blocking downstream VEGF action. In this trial, bevacizumab was supplied in a prefilled plastic syringe in a sealed package to each trial site pharmacy from the Liverpool and Broadgreen Pharmacy Aseptic Unit, Royal Liverpool University Hospital, Liverpool, UK. 2

Site pharmacy storage, ordering and handling procedures of investigational medicinal products

A trial medication dispensing and return log was maintained by the trial site pharmacies. Administration records from these sites were retained by the pharmacy and monitored by the trial manager to ensure that accurate CRF data were recorded. The randomisation system was linked to the investigational medicinal product (IMP) supply. Each site pharmacy was also responsible for appropriate storage, dispensing, disposal, and recall and destruction logs, in accordance with good manufacturing practice73 and good clinical practice,74 and the site hospital pharmacy’s approved policies for IMP accountability and management. Furthermore, each site pharmacy maintained a record of trial drug administration, based on the pre-printed form signed by the unmasked investigator that was returned to the pharmacy at each centre. 2

Investigational medicinal product accountability

Description and justification of route of administration and dosage of investigational medicinal product

The approved route of administration (i.e. by intravitreal injection through the pars plana of the eye) was used in all cases under sterile conditions in a designated treatment area in accordance with the guidelines75 for intravitreal injection of the Royal College of Ophthalmologists and any approved procedures at the individual site hospital. The injection could be performed by the unmasked injector(s) only, who was (were) on the hospital site LEAVO delegation log and was (were) experienced in intravitreal injection procedures. The dosages of ranibizumab (0.5 mg/0.05 ml) and aflibercept (2.0 mg/0.05 ml) used in this trial were approved by the EMA, and NICE recommends these doses of these agents for intraocular use. 12,13 The dosage of bevacizumab (1.25 mg/0.05 ml) was the dosage used in the IVAN clinical trial and the CATT of treating wet age-related macular degeneration (AMD), and the standard dose used in clinical practice. Post-injection checks were conducted in accordance with local hospital policy and included a visual acuity, intraocular pressure or optic nerve head perfusion check, or a combination of these. The interval between two doses of all three drugs was not recommended to be less than 4 weeks. 2

Routine reporting

The MHRA definitions of AEs and SAEs were adopted for this trial. AEs were reported by the site in the AEs log in the eCRF. All SAEs, serious adverse reactions and suspected unexpected serious adverse reactions (SUSARs) were recorded and reported on the SAE form to the chief investigator/delegate within 24 hours of learning of their occurrence. A record of this notification (including date of notification) was clearly documented to provide an audit trail. In the case of incomplete information at the time of initial reporting, a follow-up report was provided as soon as the information became available. The sites responded promptly to any queries raised by the chief investigator/delegate. The principal investigator/delegate, who had to be a clinician at the site, assessed the relationship of the SAE to any of the trial interventions. The chief investigator was responsible for assessing the expected or unexpected nature of all serious adverse reactions. The chief investigator/delegate, with the support of the KCTU, ensured that Moorfields Eye Hospital, as sponsor, was made aware of any SUSARs and serious adverse reactions that occurred. The chief investigator/delegate, in conjunction with the sponsor, was responsible for reporting all SUSARs to the MHRA and relevant ethics committee within the appropriate time frame.

All principal investigators were informed of all SAEs that were assessed as fulfilling the criteria for a SUSAR (i.e. possibly, probably or definitely related to any trial intervention, and unexpected as per the summary of product characteristics or the protocol). 2

Planned ‘hospitalisations’, non-emergency procedures and adverse event reporting

Some AEs met the definition of serious but did not need to be reported on a SAE report form. Common ophthalmology- and non-ophthalmology-related events that resulted in planned, non-emergency hospital admissions for the investigation or treatment of those events and that were not possibly, probably or definitely related to the IMPs did not need to be reported on a SAE report form. These events were recorded on the AE form and the investigation and treatment of ophthalmology-related events were recorded on the ophthalmology-related concomitant procedure forms. All concomitant medications were recorded on the concomitant medication form. These forms were updated following each trial visit to ensure that the independent DMEC received accurate reports of the occurrence and treatment of AEs. 2

Confidentiality

Data were handled, computerised and stored in accordance with the Data Protection Act 1998. 64 Participants were identified via a unique PIN, their date of birth and their initials. Identifiable information was stored in the eCRF and did not leave the site. Any participant contact information was stored in the site on password-protected computers or in secured locations with restricted access.

Data collection tools and source document identification

Written informed consent was obtained before screening and other trial-specific procedures were performed. SAE data were collected on paper SAE report forms and e-mailed or faxed to the KCTU. Summary details of SAEs were transcribed to the AE section of the eCRF. For all other data collected, source data worksheets were used for each patient and data were entered onto the eCRF database. Source data worksheets were reconciled at the end of the trial with a patient’s NHS medical notes in the recruiting site. During the trial, critical clinical information was written in the medical notes to ensure that informed medical decisions could be made in the absence of the trial team. Trial-related clinical letters were copied to the medical notes during the trial. It was the responsibility of the principal investigator and his/her team to ensure that the accuracy of all data entered in the worksheets and the eCRF was in accordance with good clinical practice. The delegation log identified all those personnel with responsibilities for data collection and handling, including those who had access to the trial database. The principal investigator was responsible for ensuring that source data worksheets were filed in a suitably secure location so that source data verification could be undertaken throughout the trial. 2

Data handling and analysis

All trial data and site files were kept on site in a secure location with restricted access.

The trial used an eCRF created using the InferMed MACRO database system. Data were managed using this system. The eCRF was created in collaboration with the trial statistician and the chief investigator and maintained by the KCTU. It was hosted on a dedicated secure server in King’s College London. This system is regulatory compliant; has a full audit trail, data discrepancy functionality and database lock functionality; and supports real-time data cleaning and reporting. The trial manager was responsible for providing usernames and passwords to permitted local trial personnel. Only those authorised by the trial manager were able to use the system. 2,63

Sample size calculation

Bevacizumab and aflibercept were hypothesised to be substantially inferior to ranibizumab if, in each case, the mean of the primary outcome (i.e. change in BCVA ETDRS letter score) was worse by a margin of 5 letters, a previously used non-inferiority margin,26,76 representing the minimum visual acuity change that a patient may distinguish. A similar CRVO population9 reported a standard deviation (SD) of 14.3 letters in the ranibizumab (0.5 mg) arm; the 12-month rate of those lost to follow-up was 8.4% in the ranibizumab arms (0.5 mg and 0.3 mg). In the absence of 24-month data, we assumed a comparable SD of 14.3 letters at 100 weeks, and allowed for 15% dropout. The two null hypotheses, that bevacizumab was substantially inferior to ranibizumab, and that aflibercept was substantially inferior to ranibizumab, were each planned to be rejected if the estimated 95% confidence interval (CI) for the difference in treatment means was wholly above the 5-letter margin in each case. Assuming equal efficacy, there was 80% power to reject each null hypothesis and to declare non-inferiority, with 130 followed-up patients analysed per arm. Allowing for 15% missing data at 100 weeks, 459 patients were planned to be randomised to the three arms (equal allocation ratio of 153 participants per arm). Sample size calculations were performed using nQuery Advisor version 4.0 (Statistical Solutions, Saugus, MA, USA). The primary method of analysis was a linear mixed-effects (LME) model with adjustment for baseline, which was expected, other things being equal, to increase the power to detect non-inferiority. The primary method of analysis included all available refracted data of the primary outcome up to and including 100 weeks, including data from the 15% of participants who we anticipated could miss the 100-week primary outcome end point. 2

Statistical considerations

The trial statisticians were responsible for all statistical aspects of the trial, from design through to analysis and dissemination. 2 A detailed statistical analysis plan was completed before the start of the trial; it was commented on by the DMEC and approved by the Trial Steering Committee (TSC). The plan was accompanied by a health economics analysis plan, and was updated and re-approved by the TSC when the protocol was amended.

Hypotheses

The hypotheses refer to the populations of relevant patients, rather than to trial subjects:

• The working hypothesis – the so-called ‘working hypothesis’ was the hypothesis that motivated the trial, which the trial results may or may not support. It was that the change in BCVA is non-inferior in patients treated with either aflibercept or bevacizumab, compared with patients treated with ranibizumab.

• The statistical null hypothesis 1 – bevacizumab is inferior to ranibizumab in eyes with MO due to CRVO at 100 weeks.

• The statistical null hypothesis 2 – aflibercept is inferior to ranibizumab in eyes with MO due to CRVO at 100 weeks.

• Statistical alternative hypothesis 1 – bevacizumab is non-inferior to ranibizumab in eyes with MO due to CRVO at 100 weeks.

• Statistical alternative hypothesis 2 – aflibercept is non-inferior to ranibizumab in eyes with MO due to CRVO at 100 weeks.

Treatment arms

The trial was randomised with equal allocation of participants in a 1 : 1 : 1 ratio to the three arms (see Chapter 2, Randomisation).

Trial samples

Outcomes

Subgroup variables

Three subgroup variables were considered: (1) baseline visual acuity (low, moderate and high: ≤ 38 letters, 39–58 letters and 59–78 letters, respectively), (2) disease duration (< 3 months or ≥ 3 months) and (3) ischaemic compared with non-ischaemic. These variables were based on the fact that the visual gain in the worse-vision group may be higher than that achieved by the better-vision group, and this effect may differ between arms. The shorter the duration of disease, the better the visual acuity outcomes, and this may have varied between treatment arms.

Outcomes requiring derivation

The VFQ-25 is a validated tool for assessing vision-related quality of life. It consists of a base set of 25 vision-targeted questions, representing 11 vision-related subscales, plus an additional single-item question rating general health. The overall composite score is computed as the simple average of the vision-targeted subscale scores, excluding the general health rating question. The overall score can range from 0 (worst possible score) to 100 (best possible score).

The EQ-5D and the EQ-5D-V

The EQ-5D is a generic instrument for describing and valuing health. It is based on a descriptive system that defines health in terms of five dimensions (mobility, self-care, usual activities, pain/discomfort and anxiety/depression). Each dimension [in the EuroQol-5 Dimensions, five-level version (EQ-5D-5L)] has five response categories corresponding to ‘no problems’, ‘slight problems’, ‘moderate problems’, ‘severe problems’ and ‘unable to/extreme problems’. A preference-based score ranges from states worse than dead (< 0) to 1 (full health), anchoring dead at 0. In addition, the EQ-5D includes a visual analogue scale, which records a respondent’s self-rated health on a vertical scale where the end points are labelled ‘best imaginable health state’ (marked as 100) and ‘worst imaginable health state’ (marked as 0). The EQ-5D-V is similar to the EQ-5D-5L, but with another dimension (vision) added to overcome perceived inadequacies in a particular population.

More information is given in Chapter 4, Health-related quality-of-life measures.

Defining outliers

Outliers are observations that have extreme values relative to other observations under the same conditions. An outlier was defined as a data point at least 4 SDs from the mean of its distribution of values observed across other participants. A ‘bivariate outlier’ for checking was defined as a pair of successive serial data points of the same measure for a participant whose difference was at least 4 SDs from the mean of all participants’ such differences. Simple plots of successive pairs of serial measures were used throughout the 24-month period to help identify outliers for data-checking. 63

Handling outliers

Outliers were identified for further investigation by looking at the distributions of the data using histograms, scatterplots or box plots. Univariate tests for the compatibility of the distribution with a normal distribution were not undertaken because they can be too sensitive to departures that are often not relevant to the comparison of means (central limit theorem).

Once an outlier was found, a masked member of the team with sufficient clinical experience was involved in the decision about whether a data value was impossible or implausible or plausible. If an outlier was impossible, then it was set to missing. If an outlier was clinically plausible, then the outlier remained. If an outlier was clinically implausible (but possible), then it was not ignored or deleted, but was retained for the ITT analysis. If outliers remained in the distribution of a variable, then data transformations or non-parametric methods of analysis were considered. A sensitivity analysis was undertaken to check whether or not the outlier was influential by obtaining results with and then without the inclusion of the outlier. If the conclusions changed, then this was noted. 63

Baseline comparability of randomised groups

Baseline descriptions of participants by treatment and overall were summarised. No significance testing was carried out as any differences found might have been chance-generated and not for hypothesised reasons. Continuous variables, such as OCT CST values and VFQ-25 scores, were summarised using means and SDs and/or medians and interquartile ranges (IQRs) for variables presenting a skewed distribution. Categorical variables, such as the proportion of participants gaining ≥ 15 BCVA letters or participants with OCT CST of < 320 µm, were described using numbers and percentages.

Comparison of rates of adherence and follow-up

High compliance and low attrition rates were anticipated for this trial based on previous clinical trial experience. In CRUISE (a study on CRVO), 91.6% of participants completed the active treatment arms at 12 months, and withdrawals were mainly due to physician and patient decisions. 8 A cumulative dropout of approximately 15% by year 2 was predicted for LEAVO and this was reflected in the sample size calculations. Nevertheless, compliance rates and attrition rates were compared and reported by arm using Fisher’s exact test.

Analysis covariates

The ICH E9 guideline78 recommends that consideration be given to accounting for randomisation stratifiers by adjusting for them as covariates in the linear model. This tends to improve the precision of estimated treatment effects. Therefore, for continuous outcomes, the analysis included adjustment for the randomisation stratifiers of screening BCVA letter score (three levels) and disease duration (two levels). This excluded the third stratifier of previous treatment (eye treatment naive vs. had received previous treatment), because the numbers of participants who had received previous treatment was very small; this was approved in the statistical analysis plan [see www.journalslibrary.nihr.ac.uk/programmes/hta/119203/#/documentation (accessed 14 July 2020)] by the TSC.

Statistical model

The following description of the statistical analysis was applied to obtain results for each of the two investigational treatments, bevacizumab and aflibercept, compared with the standard treatment, ranibizumab.

Primary outcome analysis

Part of this section is reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium. The text includes minor additions and formatting changes to the original.

The primary efficacy measure was the change from baseline in refracted BCVA in the trial eye, using the ETDRS letter score, at 100 weeks. The continuous primary outcome was a participant’s longitudinal change in BCVA from baseline to 100 weeks. As more fully described later, this baseline is adjusted for as a continuous covariate. This analysis approach gives results equivalent to those of an approach in which the primary outcome is instead defined to be the cross-sectional 100-week measurement in the same participants. Of these two equivalent approaches, we chose to analyse BCVA and other continuous outcomes at the cross-sectional measurement point. This is convenient, because it means that, if a baseline measurement is missing in a participant with a 100-week outcome, the end point is not considered to be missing. The primary outcome may, therefore, be referred to later as the 100-week visual acuity, rather than as the change in BCVA from baseline to 100 weeks.

The primary outcome was analysed using a LME model incorporating the five post-baseline measurements of the refracted BCVA outcome (at 12, 24, 52, 76 and 100 weeks). This mixed model was, by definition, a mix of random- and fixed-effect terms. The random effect in the model was the participant, represented as a random intercept at each follow-up time point, with allowance for within-participant correlation in the adjusted post-baseline outcomes. The fixed effects in the model were the main effect terms for arm; the two stratifiers, visual acuity and disease duration; ‘time’; and the baseline of the outcome and its missing indicator required for the missing indicator method. The other fixed effects in the model were the interactions between ‘time’ and each of the other fixed effects in the model. This model allowed the treatment effect to be formally tested at 52 weeks and at the primary time point of 100 weeks, and estimated at 24 and 76 weeks. 63

Secondary outcome analysis

Secondary outcome analyses (see Appendix 3, Table 28) were on an ITT basis only. All tests were two-sided at the 5% significance level and were interpreted cautiously, with a focus on interpreting effect sizes with 95% CIs. Safety outcomes were reported as unadjusted patient proportions and as rates within and between arms, with 95% CIs, using exact methods when appropriate. Significance tests were used sparingly and were restricted, when possible, to addressing stated hypotheses.

Trial Steering Committee

The TSC was the committee responsible for monitoring the overall integrity, conduct and safety of the trial. It monitored trial progress, investigated any SAEs, and took account of regular reports from the DMEC and communication from the Trial Management Group (TMG). Ultimate responsibility for any decision required on the trial’s continuation lay with the TSC. The TSC comprised an independent chairperson, a professor of statistics, an independent ophthalmologist and general physician, a consultant in public health, a senior Department of Health and Social Care policy-maker and two patient representatives. TSC meetings were held at least annually and arranged by the chief investigator and the trial manager in conjunction with the chairperson. For a list of committee members, see Appendix 2. A representative of Moorfields Eye Hospital (the sponsor) was invited to each meeting. 2

Data Monitoring and Ethics Committee

An independent DMEC of three individuals, one professor of statistics and two retina specialists, met regularly to safeguard the interests of trial participants, assess the safety and efficacy of the interventions during the trial, and monitor the overall conduct of the clinical trial (see Appendix 2). Its terms of reference were to receive and review the progress and accruing data of the trial and to provide the TSC with advice and recommendations on trial conduct. The trial would have been discontinued on the basis of new safety information, or for other reasons given by the DMEC and/or TSC, sponsor, regulatory authority or Research Ethics Committee concerned. All data reviewed by the DMEC determined safety issues. All serious adverse reactions were reported to the KCTU within 24 hours of learning of their occurrence. 2

Trial Management Group and site monitoring

The TMG was responsible for monitoring the delivery of the trial on a day-to-day basis, and was supported and managed via the KCTU. The TMG membership consisted of the chief investigator, the co-lead, the trial manager, the data manager, the lead and trial statisticians and senior members of the KCTU. Other members of the wider research team were also invited on a meeting-by-meeting basis, depending on the scope covered. Trial conduct and data collected were monitored by a combination of central review and site monitoring visits to ensure that these were in accordance with good clinical practice. Trial site monitoring was undertaken by the trial manager, the assistant trial manager and an experienced KCTU trial monitor. The main areas of focus were consent, SAEs and essential documents in trial site files.

Site monitoring included:

• reviewing all consent forms in the site file and medical notes

• source data verifying SAEs against medical records

• source data verifying a proportion of the primary outcome measure against medical records

• checking essential documents in the investigator site file and trial files.

Central reviews included:

• ensuring accuracy and completeness of all applications for trial authorisations and submissions of progress/safety reports, prior to submission

• ensuring that all documentation essential for trial initiation was in place prior to site authorisation

• reporting and following up all monitoring findings with the appropriate persons in a timely manner.

The investigators and institutions also permitted trial-related monitoring, audits, Research Ethics Committee review and regulatory inspections, providing direct access to source data/documents. Trial participants were informed of this during the informed consent discussion. Participants consented to provide access to their medical notes.

Overview

NIHR acknowledges the need for experienced trial management and recommends the involvement of a specialised clinical trials unit to conduct trials. We were fortunate to have the multidisciplinary team from the KCTU participate in the trial. As a LEAVO collaborator, the team provided a trial manager, deputy trial manager and experienced monitors, in addition to a senior and a junior statistician, and the expertise of their core team, including the Clinical Trials Unit operations manager, senior data manager and trial methodologist. All these members attended TMG, TSC and DMEC meetings, when appropriate. In addition, the KCTU team members were all available for advice and guidance on a daily basis; working in conjunction with the trial manager, the KCTU was, ultimately, the cornerstone of the trial. 63 It recognised the need to open as many sites as quickly as possible and its senior team spent many hours with the trial manager, ensuring that she was fully familiar with the trial and was able to begin site initiations before recruitment commenced on 1 December 2014. The largest and most experienced sites (e.g. Moorfields and Leeds) were initiated first. Unfortunately, a few weeks before the initiation of the first site, the original trial manager was absent on sick leave and she announced her resignation at the beginning of December 2014. Not unexpectedly, this had a significant impact on site initiation and could have led to very prolonged trial delays. Fortunately, an experienced assistant trial manager had just been appointed and agreed to step up to the trial manager position within a few days of starting. Quite understandably, he took time to familiarise himself with the trial protocol and procedures; therefore, the trial fell significantly behind with site initiations and recruitment. The low point was 39 participants recruited by the end of May 2015, against a predicted target of 76 (51%). However, the new trial manager began to recover the situation in the second quarter of 2015, and the number of site initiations increased: we initiated only eight sites in the first 4 months of recruitment, compared with 13 sites in the succeeding 2 months. As a result, actual recruitment kept pace with predicted recruitment in October, November and December 2015. By November 2015, that is after 12 months of recruitment, we had opened 38 sites, against a target of 40, and recruited 176 participants, against a target of 268 (66%). An additional eight sites were subsequently initiated, to give 46 greenlighted sites open in the first quarter of 2016. By 31 May 2016, when recruitment should have been completed, we had recruited 320 participants, against a target of 459 (70%); by December 2018, we had completed recruitment almost exactly 6 months behind schedule (Figure 3). Table 1 shows the number of participants recruited each month by site, and Table 2 shows the number of participants whom each site recruited per trial arm.

FIGURE 3.

Actual vs. projected recruitment per month.

Barriers to recruitment and corrective strategies

The following barriers to recruitment were identified:

• Availability of trial staff, for example masked injectors and trial co-ordinators. Despite fulfilling our initial trial site requirements, several sites were unable to provide sufficient clinician unmasked injector cover (e.g. Rugby) as a result of limited staff availability, or sufficient research co-ordinator time for the trial (e.g. Addenbrooke’s and Hillingdon), the latter in some cases because NHS support costs attributable to LEAVO were not available to the local trial team. We largely resolved the former issue in a substantial amendment that allowed nurses and optometrists who were certified intravitreal injectors in standard NHS clinics to provide unmasked injector cover for LEAVO. We also approached a number of local ophthalmology CLRNs to provide additional co-ordinator time for the trial based on CLRN support costs, and received very helpful support from Rupert Bourne, CLRN National Lead for Ophthalmology, in this regard.

• Difficulties with the protocol. The following changes were made to the protocol (see Appendix 3, Table 30) –

  • The upper limit of the visual acuity eligibility at baseline was increased from 73 (Snellen equivalent: ≈ 6/12) to 78 letters (Snellen equivalent: ≈ 6/9). Patients in clinical practice with a visual acuity of 6/9 had previously been excluded from the trial as their visual acuity was too good, and they were receiving NHS treatment instead. The change allowed patients with a visual acuity of 6/9 to enrol in the trial.

  • The inclusion criterion for diabetic retinopathy in the trial eye was changed from ‘any previously documented diabetic retinopathy or diabetic macular oedema in the study eye’ to ‘any diabetic retinopathy or diabetic macular oedema at baseline clinical examination of the study eye’. 2 This was to prevent patients being excluded from the trial who presented with a documented history of diabetic retinopathy, which may not have been reliable, rather than clinical evidence based on the trial screening examination.

  • The allowed number of previous anti-VEGF injections was increased from three to six in order to allow patients who had had longer-term treatment for MO due to CRVO (i.e. six injections) to be considered for the trial.

  • Patients who had had recent pan-retinal photocoagulation for NVE, NVD or NVI were considered eligible for the trial within 1 month of treatment rather than within 3 months, as treatment within 1 month would not have had an adverse outcome on anti-VEGF therapy 1 month later.

  • The protocol was altered to change the rescreening interval to 2 weeks, except for visual acuity eligibility, which remained at 4 weeks. Several patients had not enrolled in LEAVO because, for example, they had forgotten to take blood pressure medication, leading to high blood pressure and a screen fail. If they needed to wait 4 weeks before rescreening, as the protocol originally stated, then they typically opted for NHS treatment in the interim; being able to rescreen after 2 weeks prevented them being lost to NHS care.

• Number of sites. Although we planned for 40 sites initially, four withdrew before being initiated, and so we took an early decision to add additional sites. Initially, we planned for a further 12 sites, which would have taken the total to 48 active sites. However, two of these withdrew, and 10 were greenlighted, although one failed to recruit any participants. Nevertheless, these additional sites made a very significant contribution to the last 6 months of recruitment.

• Site equipment. Several sites had issues with equipment, in particular with wide-angled fluorescein angiography imaging devices and information technology support that allowed communication with the KCTU randomisation software and MACRO trial database, and also allowed data export to the reading centre. We worked with the sites and providers of equipment (e.g. Optos wide-angled imaging) to overcome these issues as quickly as possible.

• Although we had held an investigator meeting prior to the trial start, a number of optometrists had not been able to attend this and required certification before a site could be greenlighted to recruit patients. To minimise certification delays, we arranged for prompt visits by either lead trial optometrist to any site to undertake optometry certification.

• Other measures that were used to try to maximise recruitment included the following –

  • A monthly newsletter to every site detailing progress63 and acknowledging each site that had recruited one or more participants in the previous month.

  • An e-mail from the chief investigator to each site team every 2 months encouraging further recruitment.

  • A thank-you e-mail to each site from the chief investigator after each participant was recruited.

  • Reward vouchers each month to the site recruiting the most participants and ‘best site of the month’.

  • Very prompt replies to any site that had queries on any aspect of the trial. We think that this point was critical in keeping sites focused on recruitment and willing to recruit over and above their target, which was something we specifically asked large sites to do.

Withdrawals

Appendix 3, Table 31, shows the number of participants who did not complete the week 100 visit in the three arms, and the week of their last visit. Appendix 3, Table 32, shows the number of weeks all withdrawal participants participated in the trial and the reasons for withdrawal. Withdrawals were balanced across treatment arms; overall, more participants completed their week 100 visit [87.9% (407/463)] than had been predicted in the sample size calculation (85%).

Primary outcome

The mean gain in BCVA letter score was 12.5 with ranibizumab (SD 21.1), 15.1 with aflibercept (SD 18.7) and 9.8 with bevacizumab (SD 21.4) at 100 weeks (Figure 4). First, the primary outcome at 100 weeks was unable to show that bevacizumab was non-inferior in terms of BCVA in both the ITT and per-protocol populations (Table 5). The 95% CI for the adjusted difference between arms at 100 weeks lay below the prespecified acceptable margin of –5 letters (Figure 5). Second, aflibercept was non-inferior, but not superior, to ranibizumab in terms of BCVA in both the ITT and the per-protocol populations (see Table 5 and Figure 5). The 95% CI for the adjusted difference between arms at 100 weeks lay above the prespecified acceptable margin of –5 letters (see Figure 5). The mean BCVA letter score at 24 weeks had decreased by approximately 3 letters across groups following pro re nata injections at weeks 16 and 20, when fewer injections were given (ranibizumab injections, n = 123; aflibercept, n = 76; and bevacizumab, n = 121), but increased gradually thereafter across groups to week 100, during which period participants were seen at least every 8 weeks and received injections promptly if re-treatment criteria were met (see Figure 4). Such peak-and-trough changes in visual acuity were closely mirrored by OCT trough and peak CST results over the 2-year period.

FIGURE 4.

Adjusted mean BCVA letter score across groups to 100 weeks. Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium. This includes minor additions and formatting changes to the original.

Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium.

FIGURE 5.

Forest plot of the primary outcome at 100 weeks. PP, per protocol. Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium. This includes minor additions and formatting changes to the original.

Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium.

The principled sensitivity analysis for missing data supported the primary outcome results (Figures 6 and 7). The sensitivity analysis for outliers was not conducted, as there were no outliers in the ITT and per-protocol populations [see www.journalslibrary.nihr.ac.uk/programmes/hta/119203/#/documentation (accessed 14 July 2020)]. The sensitivity analysis for concomitant treatments taken by one participant in the trial supported the primary outcome results.

FIGURE 6.

Sensitivity analysis for the missing-at-random assumption in the primary outcome analysis assessing non-inferiority of aflibercept. (a) ITT; and (b) per protocol. Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium. This includes minor additions and formatting changes to the original.

Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium.

FIGURE 7.

Sensitivity analysis for the missing-at-random assumption in the primary outcome analysis assessing non-inferiority of bevacizumab. (a) ITT; and (b) per protocol. Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium. This includes minor additions and formatting changes to the original.

Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium.

The sensitivity analysis assessed the potential impact on the treatment effect from including participants with unobserved BCVA 24-month primary outcome data in the adjusted primary outcome model. In this analysis, participants with unobserved data were, on average, specified to be able to have score ranging from –20 to 20 BCVA letters away from the scores of their counterparts who did have outcome data observed. This was applied to participants in three scenarios. Scenario 1 involved applying this to participants in the investigative treatment arm (aflibercept) only. Scenario 2 involved applying this to those in the comparator arm (ranibizumab) only. Scenario 3 involved applying this to participants in both arms equally. The x-axis in Figures 6 and 7 represents the range of –20 to 20 BCVA letter scores that those participants with unobserved data were, on average, specified to be able to have relative to the scores of their counterparts who did have data outcome observed. This analysis follows previously described methods. 81 The treatment effect in the main analysis is shown at zero. Vertical bars are 95% CIs for the treatment effect. The 95% CI bars all lie above the non-inferiority margin of –5 letters, supporting the non-inferiority of aflibercept in both the ITT (see Figure 6a) and the per-protocol (see Figure 6b) populations.

For scenario 3, and within most of the ranges of scenarios 1 and 2, the lower CI limit lay below the non-inferiority margin of –5 letters, supporting the main analysis conclusion that bevacizumab lacked non-inferiority. The difference in the mean between those with unobserved BCVA data and those with observed BCVA data would need to be assumed to be 12 letters higher for bevacizumab than for ranibizumab in scenario 1 (or 12.4 letters higher in scenario 2) in order to change the main analysis conclusion of a lack of non-inferiority in both the ITT (see Figure 7a) and the per-protocol (see Figure 7b) populations.

Secondary visual acuity outcomes

Both aflibercept and bevacizumab were non-inferior to ranibizumab at 52 weeks (Table 6). The 95% CI for the adjusted difference in BCVA between arms lay above the prespecified acceptable non-inferiority margin of –5 letters at 52 weeks for both aflibercept and bevacizumab.

The proportions of participants with a ≥ 15-letter gain were 47%, 52% and 45% (Figure 8) in the ranibizumab, aflibercept and bevacizumab arms, respectively, with 63%, 68% and 63%, respectively, gaining ≥ 10 letters at 100 weeks (Figure 9).

FIGURE 8.

Percentage of participants in each group with BCVA improvement of ≥ 15 ETDRS letters at 52 and 100 weeks. Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium. This includes minor additions and formatting changes to the original.

Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium.

FIGURE 9.

Percentage of participants in each group with improvement of ≥ 10 ETDRS letters at 52 and 100 weeks.

The proportions of participants with a < 15-letter loss were 90%, 93% and 90% in the ranibizumab, aflibercept and bevacizumab arms, respectively (Figure 10), and the proportion of participants with a ≥ 30-letter loss in BCVA was < 6% in each group (Figure 11).

FIGURE 10.

Percentage of participants in each group with loss of < 15 ETDRS letters at 52 and 100 weeks.

FIGURE 11.

Percentage of participants per group with loss of ≥ 30 ETDRS letters at 52 and 100 weeks.

There were no meaningful differences in the proportion of participants in each group who had prespecified categorical outcomes, for example a final visual acuity of < 19 letters (i.e. eligible for blind registration) (Table 7). Furthermore, there were no subgroup differences in the final visual acuity outcome by baseline stratifiers (Tables 8–10).

The were no differences between subgroups in the treatment effects on final visual acuity for any of the three baseline stratifiers.

Optical coherence tomography outcomes

The mean reductions in OCT CST from baseline to 100 weeks were –405 µm for ranibizumab (95% CI –450 µm to 360 µm), –378 µm for aflibercept (95% CI –412 µm to –343 µm), and –334 µm for bevacizumab (95% CI –374 µm to –293 µm). There were no clinically relevant differences across treatment groups for the adjusted difference in CST at 100 weeks: aflibercept versus ranibizumab was –29.3 µm (95% CI –60.9 µm to 2.3 µm); and bevacizumab versus ranibizumab was 21.9 µm (95% CI –9.7 µm to 53.4 µm). The adjusted mean OCT CST across groups increased by approximately 50 µm following pro re nata visits at weeks 16 and 20, closely mirroring the visual acuity data, and decreased gradually thereafter to week 100 (Figure 12). There was no difference in mean macular volume in each trial group at 100 weeks (see Appendix 3, Table 33).

FIGURE 12.

Adjusted mean OCT CST across groups to 100 weeks. Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium. This includes minor additions and formatting changes to the original.

Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium.

The proportion of participants with an OCT CT of < 320 µm at 52 weeks was significantly higher in the aflibercept group (76%) than in the ranibizumab group (63%), a difference of 12.4% (95% CI 1.7% to 23.1%). A similar difference was found at 100 weeks [aflibercept group (81%) and ranibizumab group (66%), a 15.3% difference (95% CI 4.9% to 25.7%)], but a difference between the bevacizumab and ranibizumab groups was found only at week 24 (–18.7%, 95% CI –30.1% to –7.4%) (Figure 13).

FIGURE 13.

Percentage of participants with OCT CST of < 320 µm at 24, 52 and 100 weeks. Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium. This includes minor additions and formatting changes to the original.

Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium.

Injection number

Part of this section is reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium. This text includes minor additions and formatting changes to the original.

By 100 weeks, ranibizumab group participants had received a mean of 11.8 injections, compared with 10.0 injections for the aflibercept group and 11.5 injections for the bevacizumab group. The difference between the aflibercept and ranibizumab groups was meaningful as early as week 24 [mean difference: –0.4 at week 24 (95% CI –0.6 to –0.2); –1.1 at week 52 (95% CI –1.6 to –0.5); and –1.9 at week 100 (95% CI –2.9 to –0.8)] (Figure 14).

FIGURE 14.

Mean number of injections across treatment groups by weeks 24, 52 and 100. Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium. This includes minor additions and formatting changes to the original.

Reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium.

Post hoc bevacizumab versus aflibercept analysis

After being approved by the DMEC, a post hoc analysis was unable to demonstrate that bevacizumab was non-inferior to aflibercept in the ITT analysis at 52 weeks (adjusted mean difference –1.35 letters, 95% CI –5.29 to 2.59 letters) or at 100 weeks (adjusted mean BCVA difference was –3.96 letters, 95% CI –8.34 to 0.42 letters; p = 0.32). The results of the per-protocol analysis were similar. At 100 weeks, there was a significant difference of 1.6 (95% CI 0.5 to 2.7) between the mean number of injections received by participants randomised to bevacizumab and the mean number received by those randomised to aflibercept.

Retinal imaging

Treatment allocation guess form

The optometrists assessing primary outcomes provided a response on the treatment allocation guess form for 409 of their 463 participants: for 356, they said they did not know; for 53, they made a guess, and were correct in 18 instances, which is consistent with chance. Of the 409 participants, 406 provided a response: 386 did not know and 20 made a guess, of whom eight [i.e. 2% (8/406)] guessed correctly, which is consistent with chance.

Safety outcomes

Part of this section is reproduced from Hykin et al. 2 This is an open access article distributed under the terms of the CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium. This text includes minor additions and formatting changes to the original.

There was one case of infectious endophthalmitis in a trial eye, which followed trabeculectomy bleb infection rather than intravitreal injection. The frequencies of all ocular AEs and Antiplatelet Trialists’ Collaboration (APTC)-defined events were similar between trial arms (Table 16). At 52 weeks, the proportions of participants who were persistent non-responders (defined as not more than a 5-letter gain in visual acuity and an OCT CST decrease of < 50 µm after 24 weeks) were 1/139 for ranibizumab, 5/133 for aflibercept and 5/135 for bevacizumab; at 100 weeks, only one participant, in the bevacizumab group, was a non-responder. During the trial, 25 (5.4%) eyes developed an ischaemic CRVO, 13 (2.8%) developed anterior segment neovascularisation and 6 (1.3%) developed retinal neovascularisation, with no difference across arms (see Table 16).

Eight ranibizumab, seven aflibercept and eight bevacizumab arm participants required panretinal photocoagulation. Two pregnancies were reported during the trial: one in a participant and one in the spouse of a participant. Both of these were followed to term with the delivery of normal neonates.

Systemic serious AEs occurred with an expected and similar frequency between groups (see Table 16), and there were no meaningful differences between groups in the frequency of AEs in the same body system (Table 17).

Interventions

A full health economic evaluation was conducted, comparing three interventions for MO due to CRVO using data collected as part of LEAVO. The interventions were as follows:

• Interventions (investigational treatments) –

  • Arm A, treatment: an intravitreal injection of aflibercept (2.0 mg/0.05 ml) administered at baseline and at 4, 8 and 12 weeks as a mandated injection. From week 16 to week 96, treatment was given if one or more re-treatment criteria were met, as specified in the trial protocol. 88 Beyond the 100-week trial period, injections were given based on treatment continuation rules.

  • Arm B, treatment: an intravitreal injection of bevacizumab (1.25 mg/0.05 ml) administered at baseline and at 4, 8 and 12 weeks as a mandated injection. From week 16 to week 96, treatment was given if one or more re-treatment criteria were met. Beyond the 100-week trial period, injections were given based on treatment continuation rules.

• Comparator (standard care) –

  • Arm C, control: an intravitreal injection of ranibizumab (0.5 mg/0.05 ml) administered at baseline and at 4, 8 and 12 weeks as a mandated injection. From week 16 to week 96, treatment was given if one or more re-treatment criteria were met. Beyond the 100-week trial period, injections were given based on treatment continuation rules.

Method of economic evaluation

The economic evaluation comprises two parts: a model-based analysis (the primary analysis) and a within-trial analysis (the secondary analysis). The model-based analysis evaluates the three interventions over participants’ lifetimes, extrapolating clinical outcomes beyond the trial period and relating these to costs and QALYs. The within-trial analysis evaluates the three interventions during the trial period using the individual patient-level cost and HRQoL data collected during the trial. The economic evaluation uses CUA. The methods for the economic evaluation were prespecified in health economic and decision-modelling analysis plan (and associated addendum) documents, prior to database lock. 89

The within-trial analysis provides the short-term cost-effectiveness evidence using individual patient-level data on quality of life and costs; therefore, it avoids extrapolation uncertainty. The model-based analysis provides the long-term cost-effectiveness evidence (extrapolating outcomes and costs beyond the trial period); this is the preferred approach for resource allocation decision-making (in line with NICE’s Guide to Methods of Technology Appraisal 201384). To support the development of the economic model, a systematic literature review was undertaken to identify evidence to inform inputs and assumptions.

Settings

Presentation of results

Quality assurance

The model was developed by two economic modellers. When one economic modeller added coding or inputs to the model, the other modeller checked these to identify and resolve any errors. The model was debugged by following simulated patients throughout the model, and verifying that the model was picking up the correct inputs and that calculations were being performed as intended. Simulated patient histories for a sample of patients were reviewed to ensure face validity. Results were compared with those from previous models and the within-trial analysis to ensure external validity. The within-trial analysis was checked for face validity, and coding was checked for errors by a second health economist.

Objectives

A systematic literature review was undertaken in line with current recommendations. 90,91 The aim of this review was to identify evidence to inform inputs and assumptions for the long-term (> 2 years) economic model of LEAVO. Data requirements for patients with MO secondary to CRVO who were treated with intravitreal injections of ranibizumab (0.5 mg/0.05 ml), aflibercept (2.0 mg/0.05 ml) and bevacizumab (1.25 mg/0.05 ml) comprised:

• relative clinical effectiveness and safety (including withdrawals and mortality)

• HRQoL estimates

• resource use and costs related to treatment, clinic visits, staffing and equipment

• presence of ischaemic CRVO at baseline

• prior treatment for CRVO at baseline

• trial eye OCT CST

• trial eye BCVA

• non-trial eye OCT CST

• non-trial eye BCVA

• new-onset MO

• injection frequency.

Methods

Results

A total of 1338 unique records were retrieved through literature searches and supplementary searching. Of these, three articles24,34,94 provided evidence of limited relevance for informing or validating the LEAVO economic model (see Appendix 6, Figure 26). A summary of included studies is presented in Table 20. For a list of excluded studies with reasons, see Appendix 6, Table 56. 95

Conclusion

Overall, there was limited evidence to adequately compare the long-term clinical effectiveness and cost-effectiveness of anti-VEGFs used in the management of MO secondary to CRVO. Comparative long-term studies of available vascular therapies for patients with MO secondary to CVRO are needed to inform treatment choices. For a detailed report of this systematic review, see Appendix 6, Systematic literature review to support the long-term health economic model.

Model design

A discrete event simulation is used for the health economic model. Discrete event simulations are structured around a set of mutually exclusive events and model the pathway of individual patients through those events according to the time at which each event happens. Each individual patient has specific characteristics that may influence which events happen and when. A patient’s history through the model is recorded and can influence if and when future events happen. Events can occur at any time. Discrete event simulations are so named because they model a discrete sequence of events, but they operate in continuous time (rather than in discrete time intervals).

A discrete event simulation model has five key advantages in this application:

1. Health states are not required – each individual patient’s visual acuity can be tracked over time on a continuous scale.

2. The trial eye and non-trial eye can be modelled separately using data on the change in visual acuity over time.

3. Each patient’s history (previous visits and visual acuity) can be tracked, so the treatment continuation rule (see Chapter 2, Treatment schedule) from LEAVO can be used.

4. The follow-up visit times can be modelled by fixing the time to milestone visits and using the treatment continuation rule from LEAVO to determine other visit times.

5. Individual patients can have different baseline characteristics, to incorporate heterogeneity.

The model diagram is shown in Figure 15. The model was built and all analyses were run in Simul8 Professional Edition (SIMUL8 Corporation, Boston, MA, USA). Once a patient is simulated and has baseline characteristics and an intervention assigned, their times to events are set; these times may be fixed or sampled from a distribution. For times to each event, see Model inputs. The event with the shortest time is the next event that the patient experiences, at which point their characteristics, QALYs and costs are updated. The patient then waits until the next event. The model ends when either the patient has died or the model time horizon is reached. The process is repeated for a large number of patients, and the total costs and QALYs are calculated. The same patients are then simulated through the model again, but with a different intervention. The total costs and QALYs are compared for each intervention to calculate cost-effectiveness results.

FIGURE 15.

Health economic model structure. Reproduced with permission from Pennington et al. 95 This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Reproduced with permission from Pennington et al. 95 This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

At each model event, costs, utilities, total costs and QALYs are updated. Each model event (i.e. visit to ophthalmologist, ocular AE, withdrawal, new-onset MO in the non-trial eye, annual change in visual acuity, death, model end) is explained in more detail in the following subsections.

Model inputs

Health-related quality of life

The model considers patients’ BCVA over their lifetimes. To include patients’ utilities over time, the model predicts utility from BCVA and other demographic variables. This prediction is termed a ‘mapping’ or ‘crosswalk’ and may be used in economic evaluation to convert clinical measures to health utilities when either utility data are not directly available or there is a need to relate clinical outcomes to health utilities in the long term. Developing a mapping requires a data set that contains both the clinical measure and the utility measure. LEAVO provided this data set for BCVA and three measures of utility.

Mapping from best corrected visual acuity to utility

Data from all milestone visits were combined to maximise the number of observations using a complete-case analysis. At each observation, variables were generated for the visual acuity in the better-seeing eye (BSE) and the worse-seeing eye (WSE), according to whether BCVA was greater in the trial eye or the non-trial eye.

Standard statistical models are often a poor fit to the distribution of utility data113 (particularly EQ-5D data), so adjusted limited dependent variable mixture models (ALDVMMs) are used. Mixture models can be used to represent latent classes (discrete variables that are inferred rather than directly observed) within an overall population, or to provide a very flexible semiparametric framework for modelling distributions with unusual shapes. Limited dependent variables are those whose range of possible values are restricted. ALDVMMs, therefore, represent a flexible framework for developing models to reflect the distribution of utility data.

The ALDVMMs were estimated with one to four components (classes). 114 Models were fitted for the three utility measures. The independent variables used to predict utility in the components were age, sex, BSE BCVA and WSE BCVA. The interaction between the BSE and the WSE was considered as a variable, but its inclusion worsened model fit; therefore, it was excluded from the model specification. BSE BCVA and WSE BCVA are used to determine the probability of a patient belonging to the different components. Intervention is not included as an independent variable, as its impact on utility is expected to be through changing BCVA and not through a treatment-specific effect.

To determine the number of components that should be used for each utility measure, model fit was compared using the mean error, mean absolute error (MAE), root-mean-square error (RMSE), AIC, BIC and visual inspection. In each utility measure, the mean error, MAE and RMSE were generally similar for models with two, three and four components for which component membership was predicted by the BCVA score of the BSE and the WSE. The AIC and BIC, which penalise models with more parameters to reduce overfitting, indicated that the best-fitting model for VFQ-UI has three components, whereas the best-fitting models for the EQ-5D and the EQ-5D-V have two components. Appendix 3, Table 34, shows the model parameters.

The utility in each component is calculated as follows:

1. A temporary variable u is calculated by multiplying the within-component coefficients by the individual patient’s characteristics (as per a regression equation).

2. Parameter a is calculated as –(4)a=uu−uσ.

3. Parameter b is calculated as –(5)b=ul−uσ.

4. Parameter c is calculated as –(6)c=φ(a)−φ(b).

5. Parameter d is calculated as –(7)d=Φ(a)−Φ(b).

6. If parameter c is between –0.00000001 and 0.00000001 parameters, c is set to 0 and d is set to 1.

7. The expected utility within the component is calculated as –(8)(1–Φ(a)+(Φ(a)−Φ(b)))×(u+ (σ×cd))+(ul×Φ(b)),

where u_u is the highest feasible utility next to 1, l is the lowest feasible utility, σ is the variance of the component, φ is the probability density function for the normal distribution with mean 0 (SD 1), and Φ is the cumulative distribution function for the normal distribution with mean 0 (SD 1).

The probability of belonging to each component is calculated by the exponentiation of the product of the between-components by an individual patient’s characteristics. For the last component, this will equal 1. The probabilities are then normalised by dividing by the sum of all probabilities.

The expected utility within each component is multiplied by the probability of belonging to each component. The sum of these gives a patient’s utility. The relationship between visual acuity and utility for the BSE and the WSE is provided in Figure 16. A Microsoft Excel^® (Microsoft Corporation, Redmond, WA, USA) tool that calculates a patient’s utility score (VFQ-UI, EQ-5D, EQ-5D-V) based on our mappings is provided. 115

FIGURE 16.

The VFQ-UI mapping comparison of observed and predicted scores: (a) VFQ-UI score vs. BSE BCVA score; (b) VFQ-UI score vs. WSE BCVA score; (c) VFQ-UI score vs. BSE ETDRS score; and (d) VFQ-UI score vs. WSE ETDRS score.

Resource use and costs

Costs were calculated using Great British pounds for 2017/18; when costs were not available for this year, they were inflated using the Hospital and Community Health Service (HCHS) index. 116

Addressing uncertainty

The base-case analysis uses the VFQ-UI, and scenario analyses consider the EQ-5D and EQ-5D-V. Scenario analyses consider shorter time horizons and a cost of £243 for bevacizumab.

Patient Access Schemes are in place for ranibizumab and aflibercept, offering a discount on the list price. 13,51 However, the level of discount is confidential and so is unknown. Therefore, we consider threshold analyses to determine the level of discount that would be needed to change the decision about the most cost-effective intervention.

Results are presented for the base-case, EQ-5D, EQ-5D-V and 100-week scenarios using probabilistic sensitivity analysis to incorporate parameter uncertainty. Whereas deterministic analysis (see Appendix 6, Model-based analysis results: additional data) uses point estimate (mean) inputs, probabilistic sensitivity analysis simultaneously samples all uncertain inputs from their associated distributions. Microsoft Excel was used to sample uncertain parameters from their distributions and to maintain relationships between related parameters. Mean total costs and QALYs are calculated for the modelled patient cohort for each simulation. 95% CIs around the mean and total costs and QALYs are presented using the SE to reflect the uncertainty around the mean. The mean of the mean total costs and QALYs for each intervention are calculated from all of the simulations and used to calculate mean probabilistic ICERs. The uncertainty around the mean probabilistic ICER is calculated using the incremental net monetary benefit (INMB) approach to avoid the mathematical limitations of interpreting uncertainty around a ratio. 124

Running probabilistic sensitivity analysis on a discrete event simulation model is computationally expensive, but it is vital that a sufficient number of simulations are performed so that the model results converge. The number of simulations required for the results to converge can be calculated by comparing the upper and lower bounds of the INMB with zero for a defined cost-per-QALY threshold. 124 Using the tutorial provided by Hatswell et al. 124 and a threshold of £30,000 per QALY, very few probabilistic simulations are required for the analyses. This is because the ICERs are so far away from the threshold of £30,000 per QALY that there is very little uncertainty associated with the decision about which intervention is most cost-effective (the INMB CIs exclude zero). Probabilistic sensitivity analysis is presented using 500 simulations for all scenarios. This is sufficient to ensure that the INMBs have converged for each comparison of two interventions.

Method of economic evaluation

The methods for the within-trial analysis were prespecified in a health economic analysis plan prior to database lock. 89 The primary outcome of the within-trial health economic analysis was to establish the short-term cost-effectiveness of:

• aflibercept compared with ranibizumab

• bevacizumab compared with ranibizumab

• aflibercept compared with bevacizumab.

A fully incremental analysis (ranking the alternative treatment options by total costs and ruling out dominated and extendedly dominated options from the comparison) was also performed. The economic analysis used individual patient-level data collected as part of LEAVO. The total costs and QALYs over the 100-week follow-up period of the trial were used to calculate the incremental cost per QALY gained.

An ITT population was used, including all of the participants randomised to each treatment group. Analyses were conducted using Stata^® version 15 (StataCorp LP, College Station, TX, USA) and R (The R Foundation for Statistical Computing, Vienna, Austria).

Health-related quality of life

The individual patient-level QALYs were calculated from the utility scores for each HRQoL questionnaire at baseline and at subsequent follow-up time points using linear interpolation.

Resource use

The costing approach included identification of resource use, measurement and valuation. 125 The resource use associated with delivery of the intervention, hospital admissions, health-care contacts, continuous care and support and concomitant medications and procedures, and the costs associated with blindness, were measured.

Analytical methods

The base-case CUA was based on multiple imputation using chained equations to account for missing data. The VFQ-UI was used to calculate QALYs. The ICER was estimated comparing bevacizumab and aflibercept with ranibizumab, and aflibercept with bevacizumab. If applicable, the ICER was then compared with the NICE cost-effectiveness threshold range of £20,000–30,000 per QALY gained.

Addressing uncertainty

A parametric approach was used to address the uncertainty around the CUA estimates using the following key parameters estimated from the SUR output:

• difference in mean QALYs

• SE of mean differential QALYs

• difference in mean total costs

• SE of mean differential total costs

• covariance between total costs and QALYs.

To illustrate uncertainty, cost-effectiveness confidence ellipses and net monetary benefit (NMB) lines with CIs were produced for each pairwise comparison of treatments. In addition, a cost-effectiveness acceptability curve (CEAC) was constructed, illustrating the probability that each treatment was the most cost-effective compared with all alternative treatments at a range of threshold values that varied from £0 to £400,000. To calculate the probability that a treatment was the most cost-effective option, costs and QALYs for each treatment were sampled using a bivariate normal distribution.

Scenario analyses were calculated using SUR output as in the base-case analyses. The scenario analyses were as follows:

1. QALYs estimated using the EQ-5D, using imputed data

2. QALYs estimated using the EQ-5D-V, using imputed data

3. drug price discounts – the CUA carried out using imputed data and applying a 30% and 50% discount to the drug prices of ranibizumab and aflibercept, reflecting possible confidential PASs offered by pharmaceutical companies to the NHS

4. list price for bevacizumab – the CUA carried out using the list price for bevacizumab taken from the British National Formulary (£243)

5. complete-case analysis – the CUA carried out using complete-case data from LEAVO only

6. 52-week analysis – the CUA carried out using imputed data from LEAVO up to the week 52 milestone visit.

Base-case analysis

The results are presented in Table 21. In the base-case analysis, bevacizumab generates the most QALYs, followed by ranibizumab and aflibercept. Aflibercept generates the highest costs, followed by ranibizumab and bevacizumab. The CIs for the incremental costs and QALYS do not contain zero, demonstrating that there is a difference in both costs and effects for the three interventions (however, for QALYs this is numerically small). Bevacizumab dominates (i.e. it is more effective and less costly than) both ranibizumab and aflibercept. The 95% CIs for the NMB at £30,000 per QALY do not contain zero. At a threshold of £20,000–30,000 per QALY, bevacizumab is the most cost-effective intervention (ranibizumab dominates aflibercept).

The cost-effectiveness scatterplots (Figure 17) display the variation in the incremental costs and QALYs in the probabilistic samples. These are akin to presenting the SD; although they display the dispersion of the set of values, they do not present the uncertainty around the mean. The 95% CIs using the SE present the uncertainty around the mean, and show a difference in incremental QALYs for the three comparisons.

FIGURE 17.

Model-based analysis: cost-effectiveness scatterplots. (a) Ranibizumab vs. bevacizumab; (b) aflibercept vs. ranibizumab; and (c) aflibercept vs bevacizumab.

The CEAC (Figure 18) shows that, at £20,000 and at £30,000 per QALY, bevacizumab has the highest probability of being cost-effective (99.6% and 98.4%, respectively). Even at a threshold of £100,000 per QALY, bevacizumab has the highest probability of being cost-effective (92.8%). The probabilistic results demonstrate that bevacizumab is the most cost-effective intervention.

FIGURE 18.

Model-based analysis: CEAC.

The difference in QALYs is due to the difference in the effectiveness of the three interventions (see Table 5) and the relationship between visual acuity and utility. The difference in costs is due to the difference in the cost of the intravitreal anti-VEGF injections (intervention costs), as demonstrated by the cost breakdown in Table 22. The trial eye CST and visit costs are higher for bevacizumab than for aflibercept and ranibizumab because patients require more injections of bevacizumab. However, the drug costs are lower for bevacizumab because the cost of the injection is much lower.

Scenario analyses

For the results of scenario analysis, see Table 21. In the scenarios using the EQ-5D and EQ-5D-V, the costs are unchanged from the base case using VFQ-UI, and the total QALYs for the three interventions are similar. Using the EQ-5D, aflibercept generates the most QALYs, followed by ranibizumab. This is different from the findings for the VFQ-UI base case because the relationship between visual acuity and utility differs for the three utility measures. In these scenarios, although ranibizumab and aflibercept are slightly more effective than bevacizumab, they are not cost-effective because they are much more expensive. The ICER for ranibizumab versus bevacizumab is £908,532 (95% CI £659,881 to £1,476,254) and for aflibercept versus ranibizumab is £128,513 (95% CI £110,116 to £152,663). Using the EQ-5D-V, the results indicate the same trends, but the CI for the incremental effectiveness of ranibizumab compared with bevacizumab contains zero, indicating that the difference is not statistically significant. The CI around the INMB is presented for this comparison, as the ICER may contain dominated results when ranibizumab is less effective than bevacizumab. Aflibercept is more effective than bevacizumab and ranibizumab, but it is not cost-effective.

Using a 100-week time horizon, as in LEAVO, bevacizumab is slightly less effective than ranibizumab and aflibercept, but the ICERs for ranibizumab versus bevacizumab and for aflibercept versus bevacizumab are £34,067,841 and £2,610,554 per QALY, respectively (see Appendix 6, Table 62). However, in this analysis, the 95% CIs for the incremental QALYs for ranibizumab versus bevacizumab contain zero, demonstrating that ranibizumab is not statistically significantly better. Bevacizumab remains the most cost-effective intervention in this scenario.

In the scenario using the list price of £243 per vial of bevacizumab, the costs for ranibizumab and aflibercept and the QALYS for the three interventions are unchanged from the base-case analysis, but the cost of bevacizumab has increased (see Appendix 6, Table 63). However, bevacizumab remains significantly cheaper than ranibizumab and aflibercept, demonstrated by the fact that the CIs for the incremental costs do not contain zero. Bevacizumab continues to dominate ranibizumab and aflibercept.

In deterministic analysis using a 5- and 10-year time horizon, bevacizumab remains the most cost-effective intervention at £20,000–30,000 per QALY (see Appendix 6, Table 58).

In deterministic analysis, to have comparable costs to bevacizumab, at £28 per injection, the PAS discounts for aflibercept and ranibizumab would need to be at least 95%.

Data completeness

Over the 100-week data collection period, data were missing for some participants for baseline utility, QALY parameters for the three quality-of-life measures and total costs. For the extent of these missing data, see Appendix 6, Table 64. The highest proportion of missing data (56%) was recorded for total costs. There were only small differences in the proportion of missing data between the treatment arms. Appendix 6, Figures 29–32, explores the patterns of missing data for cost and HRQoL outcomes. The plots suggest that costs and utilities can be combined at each time period without any major loss of information. The plots suggested that the data were non-monotonic and missing at random; therefore, it was appropriate to use multiple imputation. 126

Utilities

Figure 19 summarises the mean VFQ-UI utility score at each milestone visit, with 95% CIs. There is overlap between the intervals at each of the time points, suggesting no statistical differences between the three arms at all follow-up time points. The mean utility scores from the EQ-5D and the EQ-5D-V are provided (see Appendix 6, Figures 34 and 35, and Table 65).

FIGURE 19.

Within-trial analysis: mean utility scores calculated using the VFQ-UI over 100 weeks. Table shows number of observations at each time point. Reproduced with permission from Pennington et al. 95 This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Reproduced with permission from Pennington et al. 95 This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure.

Costs

Table 23 gives a breakdown of the total costs for each of the three treatment arms. Complete-case data were used for estimating the mean costs for each item, and the mean total costs were calculated from imputed data. The mean total costs in each arm are driven by the intervention costs, accounting for 84%, 87% and 76% of the total costs for ranibizumab, aflibercept and bevacizumab, respectively.

Appendix 6, Figure 33, shows histograms of the distribution of total costs by treatment arm. In each arm, few patients incurred extreme high costs, resulting in little skewedness in the cost data.

Base-case analysis

In the base-case analysis (Table 24), the difference in mean total costs between aflibercept and ranibizumab was £1245 (95% CI £421 to £2070); between bevacizumab and ranibizumab arms, the difference was –£6760 (95% CI –£7546 to –£5973); and between aflibercept and bevacizumab, it was £7984 (95% CI £7209 to £8759). Bevacizumab dominated (less costly and with no difference in benefit) ranibizumab, with a probability of cost-effectiveness of 1.00 at the threshold of £20,000 per QALY.

Aflibercept was more costly, with a mean QALY difference of 0.004 (95% CI –0.0430 to 0.0518), than ranibizumab, with an ICER of £283,595 per QALY gained and a probability of cost-effectiveness of 0.04 at the threshold of £20,000 per QALY. Aflibercept was dominated by bevacizumab [more costly, with a mean QALY difference of –0.015 (95% CI –0.0618 to 0.0322)], with a probability of cost-effectiveness of 0.00 at the thresholds of £20,000 and £30,000 per QALY.

Uncertainty analysis

The CEAC generated from the parametric analysis, in the base-case analysis, is presented in Figure 20. The CEAC illustrates the probability that each treatment is the most cost-effective, compared with alternative treatments, at a range of threshold values. Bevacizumab has the highest probability of being the most cost-effective of the three treatments at all thresholds considered.

FIGURE 20.

Within-trial analysis: CEAC.

The confidence ellipses graphs (see Appendix 3, Figure 21a–c) represent the point estimate of the ICER in the cost-effectiveness plane, with 50%, 75% and 95% CIs around the point estimate. The ICER for bevacizumab compared with ranibizumab falls in the south-east quadrant of the cost-effectiveness plane, with the 95% confidence ellipse wholly under the horizontal axis but spanning the vertical axis, suggesting certainty around the difference in costs but uncertainty around the difference in QALYs between the two interventions. The ICER for aflibercept compared with ranibizumab falls in the north-east quadrant, again with the 95% confidence ellipse wholly above the horizontal axis but spanning the vertical axis, suggesting certainty in the difference in costs but uncertainty in the difference in QALYs. Uncertainty is also illustrated (see Appendix 6, Figure 36) using the INMB.

Scenario analysis

The results from secondary analyses using the EQ-5D and the EQ-5D-V to estimate QALYs are summarised in Table 25. Although the three HRQoL measures (VFQ, EQ-5D and EQ-5D-V) generated slightly different results, the differences between the three interventions in terms of QALYs were small and uncertain in each analysis. The overall conclusion regarding the most cost-effective treatment is unchanged. Bevacizumab consistently dominates ranibizumab, and, although aflibercept might be slightly more clinically effective than bevacizumab and ranibizumab, it is more costly, resulting in a low probability of cost-effectiveness for both treatments at the £20,000 per QALY threshold.

The results from the fully incremental analysis show that bevacizumab dominates all alternative treatment options (i.e. it is less costly and more effective); therefore, ranibizumab and aflibercept are ruled out by dominance (see Appendix 6, Table 66).

The results from scenario analyses when discount rates of 30% and 50% are applied to the drug costs of ranibizumab and aflibercept are shown in Table 26. These findings suggest that the within-trial cost–utility base-case analysis results are not sensitive to these discount rates. Although the probability of aflibercept being cost-effective, compared with ranibizumab, increased to 11% and 24% at the £20,000 per QALY threshold for the 30% and 50% discounts, respectively, this was still a low probability. Bevacizumab was still cheaper and more effective than ranibizumab, and aflibercept was more costly and less effective than bevacizumab.

Results from scenario analyses using the list price of £243 for bevacizumab, complete-case data and a 52-week time horizon are summarised in Appendix 6, Tables 67–69. The same conclusions can be drawn from these analyses as from the base-case analysis, with bevacizumab remaining the most cost-effective treatment option.

Main findings from the model-based analysis

The model-based analysis found that bevacizumab is consistently the most cost-effective intervention at a threshold of £20,000–30,000 per QALY. Bevacizumab, aflibercept and ranibizumab generate very similar QALYs, but bevacizumab leads to substantial cost savings, even when it is assumed that bevacizumab vials cannot be split, which would incur a higher cost per injection. The cost savings associated with bevacizumab are due to the much lower drug cost. For aflibercept and ranibizumab to have comparable costs to bevacizumab, and therefore have a chance of being cost-effective, the PAS discounts on the two licensed products would need to be at least 95%.

The findings were robust to sensitivity analyses, but the use of different utility measures led to differences in the absolute QALYs and ordering of each intervention. This indicates that the estimates of the differences in HRQoL are uncertain, but were consistently small across instruments.

Main findings from the within-trial analysis

The within-trial health economic analysis found strong evidence that bevacizumab is considerably cheaper than ranibizumab and aflibercept, even when potential discount rates are applied to the two licensed products. There was no evidence to suggest a difference in HRQoL between the three treatments regardless of the HRQoL questionnaire used to measure utility; however, the estimates of QALY difference are uncertain. Bevacizumab is more cost-effective option than ranibizumab and aflibercept. Aflibercept is highly unlikely to be cost-effective in the short term (100 weeks), compared with ranibizumab or bevacizumab, using the commonly used cost-effectiveness threshold of £20,000–30,000 per QALY. The cost-effectiveness results are driven by the higher intervention cost for aflibercept, with no additional benefit in terms of QALYs.

Comparison of model-based and within-trial findings

The model-based and within-trial analyses both concluded that bevacizumab is the most cost-effective intervention for treating MO due to CRVO. Both analyses found small differences in QALYs between the three treatments, and substantial cost savings for bevacizumab. Despite the different approaches used for estimating utilities in the model- and trial-based analyses, the cost-effectiveness conclusion remained the same, indicating the robustness of the economic evaluation results.

The total QALYs for each intervention were similar for bevacizumab (model, 1.641; trial, 1.666), aflibercept (model, 1.646; trial, 1.651) and ranibizumab (model, 1.641; trial, 1.627). The total costs for each intervention were also similar for bevacizumab (model, £6349; trial, £6292), aflibercept (model, £18,844; trial, £14,328) and ranibizumab (model, £15,254; trial, £13,014). The similarities between the model- and trial-based costs and QALYs can be viewed as validation of the model-based analysis.

However, there are some differences between the model- and trial-based results. The model-based analysis leads to higher costs for each intervention, despite the exclusion of concomitant medications and procedures (although these make up < £250 per intervention in the trial-based analysis). The differences in costs may be explained by higher intervention drug and administration costs in the model. The within-trial analysis uses information recorded in LEAVO on whether or not a participant had an injection at each visit, whereas the model uses data from LEAVO in combination with the LEAVO re-treatment criteria to allow extrapolation beyond the trial period. The difference between the analyses indicates that some modelled patients receive the intervention when they did not in LEAVO.

The model results follow the same trend as the trial, in that the number of injections was smaller for aflibercept than for bevacizumab or ranibizumab, but the absolute number of injections in each arm is larger. The re-treatment criteria in the model dictate that patients will be re-treated if their CST is > 320 µm, and the CST data used in the model suggest that, on average, bevacizumab and aflibercept patients have a CST above this threshold throughout the trial. Variation between individual patients may have meant that a greater proportion of patients in the trial than in the model had a CST below the threshold. Alternatively, the difference may have arisen because the re-treatment criteria in the trial stipulated that patients should have a CST of > 320 µm due to intraretinal or subretinal fluid, and the model does not consider the reason for CST values. There may have been patients in the trial who had a CST of > 320 µm for other reasons who were not treated, but who would be assumed to be treated in the model. In addition, participants in LEAVO might have missed appointments, which would lead to reduced injection frequency.

There are also differences between the QALYs in the model-based and within-trial analyses. The model- and trial-based analyses both find no significant difference between bevacizumab and ranibizumab, but the model finds that aflibercept generates significantly more QALYs than the other two interventions. This is because the model-based analysis uses BCVA in both eyes (as well as age and sex) to predict utility (and utility is higher for patients with better visual acuity), but the within-trial analysis uses utility data directly. The trial utility data will capture other factors relating to patients’ utility that may not relate to their BCVA, thus adding noise to the data. The relationship between visual acuity and utility is complex, non-linear and, in the observed LEAVO data for WSE, non-monotonic at times (see Figure 16). The mapping used ALDVMMs to try to capture the complex relationship and the distribution of utility data, but found some unusual features: typically ALDVMMs for EQ-5D contain at least three components, with one component representing patients with utility at or below zero. However, in this case, BCVA in the BSE or the WSE did not correlate with EQ-5D scores below zero, and so the model does not contain these separate components, as the covariates cannot predict membership of it.

Some of the QALY differences may also be due to differences in mortality. The within-trial analysis uses mortality data directly, and so includes the deaths of three participants in the ranibizumab group, six in the aflibercept group and four in the bevacizumab group. The model instead links mortality to baseline age, sex and the presence of CRVO; because these are the same for the modelled patients on each treatment, there is no mortality difference between the treatments.

The model-based analysis does not include blind registration and low-vision aid costs for patients who are partially sighted, consistent with previous analyses. 13,51 The within-trial analysis captures these costs and includes blind registration and low-vision aid costs using the same estimates as for patients with severe sight impairment. As cost of blindness was a small proportion of the total costs in both the within-trial and model-based analyses, this difference does not influence the results.

Clinical effectiveness and side-effect profile

The results of this prospective multicentre Phase III RCT demonstrate that repeated intravitreal injections of the three anti-VEGF agents markedly improves BCVA in patients with MO secondary to CRVO over 100 weeks. Aflibercept was non-inferior to ranibizumab in the management of CRVO-related MO at 100 weeks, but it was not superior. The trial was unable to demonstrate that bevacizumab was non-inferior to ranibizumab, as the lower 95% CI extended beyond the non-inferiority margin of –5 letters. The results were consistent in that both the ITT and the per-protocol analyses gave similar results for both comparisons. Furthermore, subsequent sensitivity analyses supported the reliability of the two non-inferiority comparisons. Although post hoc analyses should be interpreted with caution, a comparison of bevacizumab with aflibercept could not demonstrate that bevacizumab was non-inferior to aflibercept.

In clinical terms, the result confirms aflibercept, as well as ranibizumab, use for MO due to CRVO, which was important to demonstrate as both are used widely in UK clinical practice but previously had not been compared directly for this condition. Bevacizumab, on the other hand, could be worse than ranibizumab and aflibercept, or it could be no worse. Practically, this means that, if a patient was being advised on treatments for MO due to CRVO with anti-VEGF therapy, the three agents could not be presented to the patient as being completely equivalent. Clinicians would have a low level of confidence in recommending that a patient who was receiving ranibizumab or aflibercept switch to bevacizumab therapy.

Other visual outcome results across the three groups were similar, with no meaningful differences between ranibizumab, aflibercept and bevacizumab in the number of participants in each group achieving key secondary end points, such as a gain of ≥ 15 BCVA letters or remaining stable (i.e. a < 15-letter loss of visual acuity). The former means that, for patients commencing therapy, there is a 45–50% chance of achieving a three-line improvement in visual acuity. Patients can easily comprehend this by reference to a visual acuity chart when discussing the likely benefits of therapy with their clinician. All patients can be advised that, with regular attendance and adherence to treatment recommendations, there is at least a 90% chance that visual acuity will not deteriorate further. It is reassuring, from a patient perspective, to note that < 4% of participants in the bevacizumab arm experienced a significant loss of vision of ≥ 30 letters, in keeping with data pertaining to ranibizumab and aflibercept in this and in previous studies. 29–31,33

As anticipated, visual acuity improved rapidly during the initial monthly mandated injection phase, but a small mean decrease in visual acuity occurred across all three arms of the trial between weeks 16 and 24, which coincided with the pro re nata injection phase at week 16. Previous studies employed a protocol of six mandated monthly injections from week 0 to week 24. 9,21,27,29,31 During the trial design, we reviewed the available data and concluded that four mandated injections would be sufficient because in CRUISE the increase in visual acuity had reached a plateau by 4 months. 9 This may have been because the study enrolled a carefully selected population of people with non-ischaemic CRVO, who were likely to respond well to therapy. However, we now recognise that subsequent studies27,29,31,58 that introduced broader and more generalisable eligibility criteria indicate that the initial gain in visual acuity takes longer to maximise. Thus, our findings suggest that the loading phase should be extended to 6 months. Had we employed the longer loading phase, it is possible that the gain in visual acuity achieved by the LEAVO participants at week 24 could have been some ≥ 3 letters higher, and more in keeping with gains at 6 months in other studies. 27,29,31

It is also worth noting that SCORE227 and other studies, for example COPERNICUS,29,31 did not maintain such early gains through 1 and 2 years, most likely because follow-up in year 2 was too infrequent to identify and treat those patients who needed regular medication. Notably, the final gain in visual acuity at week 100, compared with baseline, is higher in LEAVO than in any other previously reported study on this condition, and could possibly have been even higher.

We believe that this reflects the importance of timely monitoring in the second year of the study, which should initially be every 4 or 8 weeks, in keeping with the LEAVO protocol. Longer intervals of follow-up in other studies probably led to loss of initial visual gains. 31,34,38 It is possible that follow-ups every 4 or 8 weeks could be extended for selected patients, but this approach was not tested in LEAVO. The adjusted mean visual acuity gains at each time point after baseline had a consistent hierarchy throughout the trial, in that aflibercept group values were higher than ranibizumab group values, which, in turn, were higher than bevacizumab gains. Even at week 76, when the differences between the groups were small, this hierarchy was maintained.

As expected, the three anti-VEGF agents caused a significant and immediate reduction in adjusted OCT CST during mandated injection phase of baseline to 12 weeks. However, the CST increased by approximately 50 µm over the next three visits, as the number of injections administered reduced markedly. This was because intense treatment during the mandated phase meant that re-treatment criteria were frequently not met at the visits at 16 and 20 weeks, leading to a rebound increase in CST by week 24, which closely mirrored the decrease in visual acuity. However, as participants entered the remaining 18 months of the trial, their visits were regularly scheduled every 4 or 8 weeks, resulting in patients who met criteria for re-treatment being treated promptly. This meant that OCT values gradually decreased through to week 100, mirrored by a gradual increase in visual acuity during the same time period, in contrast to other studies in which OCT data did not closely reflect visual acuity changes. 27 A previously unreported finding was that a significantly greater percentage of participants in the aflibercept arm than in the ranibizumab arm had an OCT CST of < 320 µm at weeks 52 and 100. This suggests that aflibercept is more effective than ranibizumab at resolving MO in the longer term, a finding previously reported in exudative AMD and DMO. 43,54 Interestingly, bevacizumab was no less effective than ranibizumab in this regard, unlike in other retinal disorders. 54

Fewer injections were required for aflibercept than for ranibizumab over 100 weeks, a difference that has been reported previously only in a treat-and-extend protocol. 57 The difference was significant as early as 24 weeks, and gradually increased by approximately 0.5 of an injection every 6 months. The post hoc analysis also found that fewer aflibercept than bevacizumab injections were required. This probably reflects the higher binding affinity of aflibercept to the VEGF molecule and a prolonged duration of action. This, coupled with a greater visual acuity gain and more patients achieving a normal thickness OCT CST at 2 years, would be a potential advantage of aflibercept over ranibizumab for MO due to CRVO.

No meaningful differences were seen between groups in OCT morphological grading at baseline or at 100 weeks.

Fundus fluorescein angiography did not detect differences across groups at baseline or exit, but, when the whole cohort is considered, there was overall change in distribution of non-perfusion at week 100, which we are further investigating.

No new safety concerns were identified in LEAVO to suggest any discrepancies in the overall safety profile of the three anti-VEGF agents, which is in keeping with previous reports. The chance of severe visual loss while undergoing anti-VEGF therapy remains low (i.e. in the order of 5% over 2 years) and has been noted in all previous studies. 9,27,29–31,57 This is typically due to development of an ischaemic CRVO, that is an increase in severity of the original occlusion to a point at which retinal blood inflow leads to compromised macular perfusion and possible neovascular complications. Patients were promptly treated with panretinal photocoagulation in such cases, and anti-VEGF therapy for MO may have coincidentally limited the risk of neovascularisation.

When this trial was conceived, it was thought that small amounts of anti-VEGF agents were absorbed into the systemic circulation from an intraocular injection, resulting in a reduction in circulating VEGF concentrations and, possibly, in an increased risk of APTC events, although this cause–effect relationship has not been established. Hence, we planned, in the grant application, to perform a meta-analysis of all comparative anti-VEGF safety data from CRVO studies that we anticipated would be carried out simultaneously with LEAVO. In practice, only the comparative US SCORE2 study27 and a small aflibercept versus bevacizumab trial have been conducted. 59 In addition, the SCORE2 investigators re-randomised their patients at 6 months, depending on whether or not the patients met predefined criteria of being good or poor responders. 58 Thus, a comparison was not possible beyond 6 months, and the comparative prevalence of AEs of anti-VEGF agents used in the two studies up to 6 months showed no difference. No study to date in multiple conditions, including nvAMD, DMO,61 branch and central retinal condition and less common conditions such as pathological myopia, has shown an increased risk of APTC events with bevacizumab, and we do not believe that this issue would be a barrier to the use of this drug in the NHS. The recent judicial review by Mrs Justice Whipple53 emphasised this point and she commented that ensuring that enough compounding pharmacies were available to ensure the large-scale safe production of significant amounts of the drug remained a key issue.

After the trial results were made available, we formulated a questionnaire to gather patient feedback, and received responses from members of the LEAVO CRVO users group that was formed prior to trial initiation, additional patients with RVO, the Barts Health/QMUL lay panel and Barts Heath diabetic patients who had a history of eye disease. We found that two-thirds of patients would consider bevacizumab treatment if the outcome could be worse than licensed alternatives but with such a small difference that it would be very unlikely to prevent them from carrying out their regular daily activities. All said that they would be more likely to agree to this if a licensed alternative was available should they not respond as expected to bevacizumab; provision would probably need to be made for this.

Limitations

The interpretation of the results should be considered in the context of patient eligibility and the trial treatment protocol. It is possible that the trial enrolled people whose eyes had limited potential for visual improvement because of a severe CRVO and compromised retinal perfusion and eyes with good visual acuity that had limited potential to improve because of a ceiling effect. Findings from secondary analyses were supportive but should be interpreted with caution as there was no adjustment for multiple testing. Aflibercept was considered an investigative agent because it was unlicensed when the trial commenced; therefore, all comparisons with bevacizumab were post hoc.

Comparative clinical data

The only previous well-powered comparison of anti-VEGF drugs for MO secondary to CRVO prior to LEAVO was SCORE2,27 which randomised 361 patients to aflibercept or bevacizumab, and treated them monthly from baseline to month 5 (six injections). The primary outcome was at 6 months. This differed from LEAVO, in which participants received monthly injections from baseline to month 3 (four injections) followed by pro re nata treatment at mandated visits at weeks 16 and 20, with milestone visual acuity assessments at 6 months. Greater mean BCVA letter gains were achieved in the first 6 months of SCORE2 than in LEAVO [aflibercept: SCORE2, mean 18.9 letters; LEAVO, mean 13.4 letters (SD 16.4 letters)]. This may be explained by the longer initial period of mandated monthly injections in SCORE227 or by differences in eligibility criteria. The baseline BCVA and case mix were dissimilar in these trials, with SCORE2 including patients with hemiretinal vein occlusion and LEAVO including patients with a baseline upper BCVA letter score of 78 (6/9), compared with 74 (6/12) in SCORE2. It is unknown whether or not the initial BCVA gains in SCORE2 could have been maintained through 2 years as the initial patient cohort was re-randomised at 6 months, depending on good and poor response to initial therapy. 58 The CRYSTAL study was a prospective single-arm study of ranibizumab therapy in CRVO with MO that followed up patients for 2 years, with a review at least every 8 weeks in year 2. Although this was a non-comparative study, the follow-up regimen was effective in maintaining first-year visual acuity gains in the second year, even though the number of injections in year 2 averaged only 3.3. This suggests that regular follow-up, with the targeting of patients in need of treatment, is of key importance. 56 Therefore, LEAVO is, to our knowledge, the only large clinical trial of MO due to CRVO to report comparative three-drug outcome data beyond 6 months with sustained visual acuity gains through 100 weeks across treatment arms.

Health economics analysis

The cost-effectiveness analysis found that bevacizumab was the most cost-effective intervention when compared with licensed agents (ranibizumab and aflibercept). In the treatment of MO due to CRVO, the model-based and within-trial analyses found small differences between the QALYs generated by aflibercept, ranibizumab and bevacizumab, but found that bevacizumab led to substantially lower costs. The finding that bevacizumab was the most cost-effective intervention was robust to scenario analyses that varied assumptions and data inputs. If bevacizumab was the standard of care and aflibercept or ranibizumab were new interventions being appraised by NICE, it is highly unlikely that they would be recommended as a cost-effective use of NHS resources.

Treatment with bevacizumab saves £5560 per year when compared with aflibercept, or £4546 per year when compared with ranibizumab (see Appendix 6, Table 69). If the estimated 5700 people diagnosed with MO due to CRVO each year in England and Wales8 were treated with bevacizumab instead of aflibercept, the NHS would save approximately £32M in 1 year (approximately £26M if the patients treated with bevacizumab instead of ranibizumab). Because the cost savings are due to a difference in intervention costs, this result would hold across other health-care systems, as long as the cost per injection for bevacizumab is lower than for aflibercept and ranibizumab.

This trial provides evidence of the cost-effectiveness of anti-VEGF treatment in MO due to CRVO, for which evidence is currently limited. A recent systematic review of the three interventions across retinal conditions did not identify any cost-effectiveness evidence in RVO. 58 This review identified two large US trials that provided evidence that ranibizumab and aflibercept are not cost-effective compared with bevacizumab in other retinal conditions (nvAMD and DMO). The cost-effectiveness findings for MO in the LEAVO trial are consistent with these findings.

The analyses adhered to good practice guidelines,84,86,87,125,126 and had the strengths of being based on data from a well-conducted multicentre randomised trial and having good retention rates over a 100-week follow-up. A key strength of the economic evaluation is the use of three different HRQoL outcome measures, including both disease-specific (VFQ-UI and EQ-5D-V) and generic (EQ-5D) measures. A range of scenario analyses have also been performed providing evidence based on a range of discounted prices for the alternative medications. In the health economics literature, there is always a debate over the relative merits of condition-specific versus generic preference-based measures (in this case VFQ-UI vs. EQ-5D). The argument is that generic measures are not sensitive to particular disease-specific improvements; therefore, the VFQ-UI was seen as a better alternative for the LEAVO population. In addition, bolt-ons to generic measures, such as the EQ-5D-V, were proposed as an alternative approach to retain comparability across different disease areas while improving sensitivity. In this trial, we used the three alternative approaches; we found that the VFQ-UI generated more QALYs for each of the three interventions. However, the incremental QALYs were similar across the three quality-of-life measures.

The strengths of the model-based analysis lie in the model design and the data inputs. A discrete event simulation facilitates the use of a continuous BCVA scale, and avoids arbitrarily grouping patients. This enables the detection of small differences in visual acuity, which are linked to utility and costs, to ensure that the differences between the three treatments are reflected. The model structure further enables consideration of both eyes, and their relationship to utility. The utility mappings follow best practice guidelines113 and up-to-date statistical methods to capture the distributions of utility. The inclusion of age and sex as variables in the utility mappings improved the model fit. In this trial population, quality of life is more likely to be affected by BCVA in both eyes (WSE and BSE). Therefore, our mappings were used to predict utility for each modelled patient using three quality-of-life measures (VFQ-UI, EQ-5D and EQ-5D-V) as a function of age, sex and BCVA in both eyes. Analysing resource use data from LEAVO allowed this to be linked to visual acuity to reflect the changing resource use associated with improvements or deterioration, which has not previously been captured in economic models for MO. 12,13 The use of growth models fitted to longitudinal BCVA and CST data allowed the extrapolation of these inputs over time, and avoided the need to make assumptions regarding effectiveness and injection frequency beyond the trial, as in previous models. 12,13

There are large numbers of missing data in the health economic analysis, but the multiple imputation model for the trial-based analysis suggests that the results are robust. Resource use questionnaires are vulnerable to recall bias. However, in LEAVO, the resource use questionnaire was designed especially for the trial. Resource use is also a small proportion of the overall total cost in each arm, so any changes are unlikely to change the health economic conclusions. Furthermore, results from the complete-case analysis provided similar conclusions, and bevacizumab remained the most cost-effective option.

The primary outcome in LEAVO concerned visual acuity in the trial eye. The model-based analysis considered both the trial and the non-trial eyes and their relationship to utility. However, consideration should be given to the relationship between these outcomes and the reality for patients; although clinical measures assess visual acuity in each eye separately, patients’ overall sight is determined by their visual acuity in both eyes together. Patients’ day-to-day functioning and quality of life may therefore not relate closely to the assessment of visual acuity in the trial eye, and this may explain why the differences in the utilities and QALYs between arms were not significant in the economic evaluation. The mapping from BCVA to utility used a robust estimator of the variance used in the statistical model. A limitation of this was the inclusion of repeated observations of the same patients to increase the number of observations available. A cluster-robust estimator of the SEs could have been used that is robust in the presence of correlation between observations for each individual. This does not change the estimated coefficients from the ALDVMM; it affects only the SEs used in the probabilistic sensitivity analyses.