Prognostic factors for the development and progression of proliferative diabetic retinopathy in people with diabetic retinopathy

1Diabetes mellitus (DM) is a chronic metabolic disease characterised by elevated blood glucose levels which, over time, lead to multiorgan dysfunction. In 2021, the International Diabetes Federation estimated that 537 million adults globally were living with diabetes (IDF 2021). It estimates this figure will rise to 643 million people by 2030, due to population expansion and ageing, urbanisation, increasing levels of obesity, inadequate nutrition, and sedentary lifestyles (IDF 2021; Saeedi 2019). Diabetic retinopathy (DR) occurs because of neurovascular degeneration triggered by hyperglycaemia, and is the most common microvascular complication of diabetes. Worldwide prevalence of retinopathy related to diabetes, including diabetic macular oedema, was recently determined to be 27% in the period 2015 to 2019 (Thomas 2019). However, in their review, Thomas and colleagues report limitations in determining a more precise estimate due to differences in study populations and methodology.

2DR is a progressive condition with advancing levels of severity. A classification in stages based on DR microvascular features, as observed on fundus photographs, was proposed by the Early Treatment Diabetic Retinopathy Study (ETDRS) Group. As a result, DR is categorised into two main stages: non‐proliferative (NPDR), and the more serious, sight‐threatening, proliferative stage (PDR) (ETDRS 1991a). The earliest visible clinical signs of NPDR are microhaemorrhages and microaneurysms which represent damage to retinal capillaries. Mild NPDR is defined by the presence of at least one retinal microaneurysm or microhaemorrhage. As disease severity progresses to moderate and severe NPDR, the number of microaneurysms and haemorrhages increase, and hard exudates, cotton‐wool spots, venous beading, and intraretinal microvascular abnormalities (IRMA) develop, signifying increasing capillary loss, hyperpermeability and non‐perfusion. Severe NPDR is defined by the '4:2:1: rule', which is the presence of retinal haemorrhages in all four quadrants, venous beading in at least two quadrants, or IRMA in at least one quadrant.

3Retinal ischaemia (also referred to as retinal capillary non‐perfusion) is considered to be the main catalyst for the occurrence of PDR. PDR is characterised by the development of abnormal new blood vessels (so‐called 'new vessels'), with or without accompanying fibrous tissue (i.e. fibrovascular membranes), at the optic disc (new vessels in the disc (NVD)) or elsewhere in the retina (new vessels elsewhere (NVE)). The ischaemic retina triggers the release of growth factors, including vascular endothelial growth factor (VEGF), which promote the growth of these new vessels in a futile attempt to restore vascular supply to the retina. However, new vessels are fragile and often rupture, leading to haemorrhages inside the eye (so‐called vitreous haemorrhages or pre‐retinal haemorrhages). PDR can progress in severity from mild to high‐risk characteristics (HRC‐PDR). The latter is defined by the presence of NVD of more than one‐fourth to one‐third disc area in size, or NVD or NVD/NVE of any size associated with bleeding, in the form of vitreous or pre‐retinal haemorrhages (Diabetic Retinopathy Study Research Group 1991). In severe cases, PDR can lead to complete visual loss resulting from proliferation of fibrovascular membranes and retinal detachment.

4Almost all, if not all, individuals with DM will develop DR if they live for a sufficient period of time. During the first two decades of disease, nearly all people with type 1 diabetes (T1D) and 60% of those with type 2 diabetes (T2D) develop DR (Fong 2003). A pooled analysis to determine the global prevalence of DR found that over one‐third of individuals with DM had DR; of these, approximately 7%, equating to 17 million individuals, will develop PDR (Yau 2012). A more recent pooled analysis estimated the global prevalence of PDR to be 1.4% for the period of 2015 to 2019 (Thomas 2019). However, the authors acknowledge significant heterogeneity in study populations and methodology as limiting factors in accurately deriving the global prevalence of DR and PDR (Thomas 2019).

5The International Diabetes Federation advises regular eye examinations every one to two years for people with diabetes and no retinopathy (IDF 2021). Once DR develops, the frequency of assessments should be increased depending on the severity of the retinopathy and level of control of systemic factors (Fred Hollows Foundation 2015). Currently, treatment options for NPDR are scarce (Royle 2015); treatment is most often only given when PDR or diabetic macular oedema (DMO) have ensued.

6The Diabetic Retinopathy Study (DRS) demonstrated that risk of severe visual loss in people with HRC‐PDR was reduced by 50% at two and five years with laser panretinal photocoagulation (PRP) treatment (Diabetic Retinopathy Study Research Group 1987). A Cochrane intervention review also verified that PRP is beneficial in reducing vision loss and progression in PDR (Evans 2014). PRP involves burning the retina, avoiding the macula (the area responsible for the central sight), with spots of laser, leading to regression of new vessels following treatment. The exact mechanism of action of PRP remains unclear, but it is presumably due to the reduced oxygen requirement of the less extensive viable retina post‐treatment, and diminished growth factor production resulting from ablation of the ischaemic retina (Doft 1984). PRP generally preserves rather than improves vision and may be associated with adverse side effects, such as diminished peripheral vision, night vision, or both, and exacerbation of DMO.

7The advent of intravitreal anti‐VEGF injections has become a pharmacological alternative to PRP (Cheung 2010). A 2014 Cochrane intervention review determined that evidence from randomised controlled trials (RCTs) for the efficacy and safety of anti‐VEGF drugs in the treatment of PDR was of low quality, but did find a reduction in the risk of intraocular bleeding (Martinez‐Zapata 2014). Recent trials have shown that anti‐VEGFs are non‐inferior to PRP in the treatment of PDR (Gross 2015; Sivaprasad 2017). However, the great majority of participants included in these RCTs did not have HRC‐PDR, where laser PRP has been shown to be most beneficial. Furthermore, anti‐VEGFs appear not to have any beneficial effect on retinal ischaemia, which seems to continue to progress despite this treatment (Chatziralli 2022; Zhu 2021). Recent studies have shown that people with PDR who are treated with anti‐VEGF therapy alone and become temporarily lost to follow‐up are more susceptible to developing irreversible blindness when compared with those treated with laser PRP (Obeid 2018; Wubben 2019). Furthermore, anti‐VEGFs do not appear to be cost‐effective unless they are used to treat people with concomitant DMO and PDR (Hutton 2017). Given that several long‐term studies have verified that the beneficial effects of PRP generally last indefinitely (Chew 2003; Dogru 1999), PRP remains the mainstay therapy for PDR. Even with treatment, however, progression of PDR and the development of further complications may still occur in severe cases.

8The moment of prognostication is any time after an individual has been diagnosed as having diabetes and DR, and prior to the occurrence of PDR.

9Although many people develop DR, few will progress to the stage of PDR. However, all individuals with DR require lifelong follow‐up, and diabetic eye screening services and eye health services are currently finding it very challenging to contend with the demand (Foot 2017). A concerning report revealed that lack of capacity within hospital eye services resulted in permanent sight loss in people of all ages, due to delayed appointments, including in people with DR (Foot 2017). The Liverpool Risk Calculation Engine study group determined that implementing individualised screening intervals based on standard clinical data would facilitate more effective management of resources into targeting high‐risk groups (Eleuteri 2017). Thus, identifying prognostic factors signalling risk of visual loss would be extremely beneficial in the enhancement and development of predictive models to optimise resources.

10This systematic review focused on identifying prognostic factors for progression from DR to PDR and to HRC‐PDR. We outline some of the risk factors below.

11Diabetes duration appears to be a key predictor of the development and progression of DR, independent of glycaemic control (Fong 2003). For example, in individuals with T1D, PDR is not usually observed for the first 10 years of disease, but there is a rapid increase in incidence, to approximately 60%, by 20 years of disease duration (Klein 2008).

12The Diabetes Control and Complications Trial (DCCT) provided evidence that rigorous glycaemic control delays development and progression of DR in T1D (Diabetes Control and Complications Group 1998). Similarly, the UK Prospective Diabetes Study (UKPDS) was pivotal in establishing the beneficial effect of regulating glycaemic levels in people with T2D (Turner 1998). A meta‐analysis of 16 RCTs found that the risk of retinopathy progression was lower after two years of intensive glucose control (Wang 1993). However, it concluded that progression to and within NPDR is clinically different from progression to PDR, but not all studies separate these stages. In those that did, long‐term intensive glucose control significantly reduced retinopathy progression to PDR (odds ratio (OR) 0.44, 95% confidence interval (CI) 0.22 to 0.87; P = 0.018; test for heterogeneity, P = 0.991) (Wang 1993).

13A Cochrane Review assessed the effects of intensive versus conventional glycaemic control on long‐term diabetic complications in people with T1D, and aimed to determine whether near normoglycaemic values are beneficial. The review confirmed that tight blood sugar control significantly reduced the risk of developing retinopathy (23/371 (6.2%) versus 92/397 (23.2%); risk ratio (RR) 0.27, 95% CI 0.18 to 0.42; P < 0.001; 2 studies, 768 participants; high‐quality evidence). However, the beneficial effect of tight blood sugar control seemed to become weaker once retinopathy was present (Fullerton 2014). A recent review, consisting of five RCTs with large sample sizes and long‐term follow‐up, found that in people with worse‐than‐moderate NPDR, intensive glycaemic control may not confer any benefits in terms of progression (Liu 2020).

14International evidence‐based clinical practice guidelines recognise the benefit of glycaemic control (Fred Hollows Foundation 2015). However, current management approaches do not fully prevent progression to PDR, and there is no glycaemic threshold below which protection is certain (Diabetes Control and Complications Group 1993).

15Prior to the undertaking of this systematic review, the current evidence on the effect of hypertension on progression to and within PDR seemed unclear. Although the Wisconsin Epidemiological Study of Diabetic Retinopathy determined hypertension to be associated with progression to PDR in people with T1D (Klein 1998), and the UKPDS identified a corresponding relationship in those with T2D (Turner 1998), other studies failed to corroborate these findings (Chew 2014; Harris 2013; Jin 2015). In the Action to Control Cardiovascular Risk in Diabetes (ACCORD) Eye study, intensive blood pressure control did not have a significant effect on retinopathy progression (Chew 2014). A Cochrane Review of 15 RCTs, including participants with T1D and T2D conducted mainly in North America and Europe, determined an association between reduced blood pressure and prevention of DR for up to four to five years (Do 2015). However, the review concluded that the available evidence did not support a benefit of intervention on blood pressure on progression to PDR or moderate/severe visual loss after five years of follow‐up. Similarly, a recent meta‐analysis concluded that intensive blood pressure control reduced relative risk of incidence of DR by 17% in T2D (Zhou 2018a). However, the available data were insufficient to confirm a relative risk reduction for DR progression or incidence of PDR (Zhou 2018a).

16The effect of cholesterol on the progression of DR also remains uncertain. The Collaborative Atorvastatin Diabetes Study found no difference in the progression of DR between participants randomised to receive a daily dose of atorvastatin and those randomised to placebo (Colhoun 2004). Investigation of fibrates in the Fenofibrate Intervention and Event Lowering in Diabetes (FIELD) study found a significant relative reduction in the need for PRP in people with T2D treated with a fibrate; a reduction in DR progression was observed only in those with retinopathy at baseline (Keech 2007). However, it is acknowledged that fenofibrate may benefit the retina independently of its lipid‐lowering effects (reviewed by Stewart and Lois) (Stewart 2018). An ongoing Cochrane Review with a published protocol will evaluate the evidence in this regard (Inoue 2019). A recent systematic review and meta‐analysis of observational studies exploring associations between serum lipids and the occurrence of DR found a slightly higher low‐density lipoprotein (LDL) cholesterol in cases with DR (Zhou 2018b). The review identified that in a large, population‐based, longitudinal, observational study of people with pre‐existing DR at baseline, poor control of total cholesterol was associated with a higher incidence of sight‐threatening retinopathy after adjusting for potential confounders. Poor control of triglycerides was also associated with progression to PDR, and this was greater when all lipid types were abnormal (Srinivasan 2017). There is currently no Cochrane Review evaluating the relationship between cholesterol and DR. Although definitive evidence is lacking regarding the effect of optimal control of blood lipids on reducing the incidence and progression of DR, it is advisable in terms of benefits to overall health.

17Diabetes duration, hyperglycaemia, hypertension, and hyperlipidaemia, whilst likely relevant for determining the risk of DR development (i.e. from no DR to presence of DR), may not fully explain the highly variable progression of NPDR to PDR, as also pointed out in a recent review by Sivaprasad and colleagues (Sivaprasad 2019). Many studies have assessed generalised DR progression using data from screening programmes where the majority of people included had no DR or only mild NPDR. To our knowledge, there are currently no systematic reviews on prognostic factors for the development of PDR and its progression.

18This review aimed to identify factors conferring increased risk of PDR and HRC‐PDR in people with diabetes once retinopathy is present.

19This review considered the prognostic factors associated with the development of PDR and progression from less than HRC‐PDR to HRC‐PDR. Thus, we investigated the health outcomes PDR and HRC‐PDR.

20As stated above, PDR is diagnosed by the presence of: NVD, defined as new vessels on or within one disc diameter of the disc; or NVE, defined as new vessels at any other locations in the retina. HRC‐PDR is defined according to the ETDRS as NVD of more than one‐fourth to one‐third disc area, or NVD or NVE of any size if associated with the presence of vitreous haemorrhage or pre‐retinal haemorrhage.

21Alarmingly, many people with diabetes can progress to the sight‐threatening stage of PDR without developing any obvious prior warning symptoms. The DRS found that approximately 50% of people with PDR who do not receive timely treatment will become legally blind within five years (Diabetic Retinopathy Study Research Group 1981a). The ETDRS was important in establishing that PRP treatment can be deferred in people with NPDR or PDR until high‐risk characteristics develop (Diabetic Retinopathy Study Research Group 1991). The study also identified that only 50% of eyes assigned to deferral of treatment (until HRC‐PDR ensued) progressed to HRC‐PDR after seven years of follow‐up (Diabetic Retinopathy Study Research Group 1991).

22A large cohort study ‐ of 7.7 million people who contributed data to the Clinical Practice Research Datalink ‐ evaluated population trends in the 10‐year incidence and prevalence of DR in the UK from 2004 to 2014 (Mathur 2017). The study considered trends by diabetes type, age, sex, ethnicity, deprivation, region, and calendar year (Mathur 2017). It found that the age‐standardised prevalence of DR decreased over time from 2.6% to 2.2%, whilst the age‐standardised prevalence of severe DR remained stable at 0.1%. The incidence also remained stable at one event per 10,000 person‐years (Mathur 2017). This suggests that despite improved medical management of DM, the threat of PDR and its complications remain a significant problem.

23The time horizon for the evaluation of health outcomes in this review was three years (± two years), eight years (± two years), or lifelong, if available.

24We undertook this review to gather evidence on prognostic factors for the development and progression of PDR. This information is essential for ophthalmologists and other healthcare professionals for the counselling and management of people with diabetes and thus for people with diabetes and their families. Our findings will help clinicians to provide advice to their patients regarding modifiable risk factors, to determine in a more personalised manner the interval required for the purpose of monitoring their disease, and to consider early intervention in high‐risk groups. Due to the increasing prevalence of diabetes and the limited resources of healthcare systems, tailoring health care in an individualised manner seems essential, avoiding the need to review patients in low‐risk groups too often and guaranteeing prompt and close evaluation and treatment, if required, of those who are at high risk.

25This prognosis review may help to identify targets for new interventions that aim to modify the course of the disease. Furthermore, the findings may guide the design and analysis of future interventional clinical trials, and highlight areas where further research is required.

26To our knowledge, there are currently no systematic reviews on prognostic factors specifically for the development of PDR and its progression to high‐risk PDR. A systematic review on prognostic prediction models for DR progression was published recently (Haider 2019). This review aimed to summarise the performance of existing models in predicting progression of retinopathy and the models' applicability for higher‐risk DR patients under hospital care to predict the need for treatment or loss of vision. Based on their findings, the authors identified the need for an accurate model that can determine patients’ individual risk of progression to a treatment stage or loss of vision. They determined that this knowledge will allow for a more appropriate use of resources and further optimisation of services, especially for individuals with a higher risk of progression (Haider 2019). This current Cochrane Review provides evidence‐based information on risk factors for the development and progression of PDR that can be used for the development of future prognostic models.

27To assess prognostic factors for predicting the occurrence of PDR in individuals with diabetic retinopathy.

28To assess prognostic risk factors for predicting the progression of PDR from less than HRC‐PDR to HRC‐PDR.

29Eligible study designs included prospective or retrospective cohort and case‐control longitudinal studies including participants who have not had previous treatment for DR. Although we initially planned to include randomised controlled trials (RCTs) evaluating therapeutic interventions to prevent progression of DR where there was a control, untreated arm, ultimately we decided not to include these (see Differences between protocol and review). We also included studies based on longitudinal registry data. It was a mandatory requirement for inclusion in the review that studies had to evaluate prognostic factors specifically for the development and progression of PDR, as opposed to generalised progression of DR.

30Studies investigating general microvascular complications of diabetes but including a subset of data related to factors involved in the development of PDR were eligible for inclusion if specific information on this group (PDR) was given.

31We excluded case reports, as they would have introduced selection bias, and editorials and letters to editors not containing primary data. We did not include cross‐sectional studies, as this type of study design is less appropriate for the evaluation of prognostic factors for the development or progression of disease.

32The target population consisted of adults (≥ 18 years of age) of any gender with NPDR or PDR with less than HRC‐PDR, diagnosed as per standard clinical practice. Studies including participants of all ethnicities, geographical locations, and socioeconomic status were eligible for inclusion. Any appropriate studies including a subset of relevant participants were considered as potentially eligible if data from this subset were given separately.

33This review considered and included prognostic factor studies only. Specific prognostic factors of interest included, but were not restricted to:

34routinely collected patient demographics and information, such as age, gender, ethnicity, socioeconomic status, and smoking habits;frequently obtained standard clinical data, such as comorbidities (presence/absence of cardiovascular disease, cerebrovascular disease, nephropathy and specifically chronic kidney failure (defined as estimated GFR of < 60 mL/min/1.73 m2), peripheral neuropathy and specifically foot ulcers, amputation), BMI, neck/waist circumference, glycated haemoglobin, blood pressure, low‐density lipoprotein, high‐density lipoprotein, triglycerides; andfunctional and structural retinal biomarkers in the prognostic context of the development and progression of PDR.

35routinely collected patient demographics and information, such as age, gender, ethnicity, socioeconomic status, and smoking habits;

36frequently obtained standard clinical data, such as comorbidities (presence/absence of cardiovascular disease, cerebrovascular disease, nephropathy and specifically chronic kidney failure (defined as estimated GFR of < 60 mL/min/1.73 m2), peripheral neuropathy and specifically foot ulcers, amputation), BMI, neck/waist circumference, glycated haemoglobin, blood pressure, low‐density lipoprotein, high‐density lipoprotein, triglycerides; and

37functional and structural retinal biomarkers in the prognostic context of the development and progression of PDR.

38We excluded studies evaluating prognostic factors involving invasive procedures that cannot be practically undertaken in a clinical setting (such as aqueous/vitreous sampling) and are thus unlikely to be translatable to routine clinical practice.

39We expected that prognostic factors would generally have been measured at the time participants entered the studies, and indeed after diagnosis of DR or PDR. If measures of prognostic factors were available at other time points, and these coincided in more than one study, we planned to consider investigating them at these other time points.

40The development of PDR was determined by the presence of NVD or NVE, as diagnosed based on fundus examination, fundus photography, or fundus fluorescein angiography. We considered participants requiring laser treatment for PDR specifically to have progressed to the outcome of PDR.

41Progression from less than HRC‐PDR to HRC‐PDR. HRC‐PDR was defined according to the ETDRS as: i) NVD 0.5 disc area plus vitreous haemorrhage or pre‐retinal haemorrhage; ii) vitreous haemorrhage or pre‐retinal haemorrhage obscuring more than one disc area (Diabetic Retinopathy Study Research Group 1991). These features could have been determined by clinical examination or by the grading of ophthalmic images, both fundus photography and fundus fluorescein angiograms. Participants requiring laser treatment for HRC‐PDR specifically were considered as having progressed to the outcome of HRC‐PDR.

42The time horizon for the evaluation of health outcomes in this review was three years (± two years), eight years (± two years), or lifelong, if available. If not, we accepted and presented other time points.

43A medical librarian specialist from Queen’s University Belfast, and the Cochrane Eyes and Vision Information Specialist searched the following electronic databases. There were no restrictions on language or year of publication. The date of the search was 27 May 2022. The search was developed around the following components: “prognostic factors”, “proliferative diabetic retinopathy”, and “development and progression”.

44Cochrane Central Register of Controlled Trials (CENTRAL; 2022, Issue 5) (which contains the Cochrane Eyes and Vision Trials Register) in the Cochrane Library (searched 27 May 2022) (Appendix 1).MEDLINE Ovid (1946 to 27 May 2022) (Appendix 2).Embase Ovid (1980 to 27 May 2022) (Appendix 3).

45Cochrane Central Register of Controlled Trials (CENTRAL; 2022, Issue 5) (which contains the Cochrane Eyes and Vision Trials Register) in the Cochrane Library (searched 27 May 2022) (Appendix 1).

46MEDLINE Ovid (1946 to 27 May 2022) (Appendix 2).

47Embase Ovid (1980 to 27 May 2022) (Appendix 3).

48We supplemented the above searches by screening reference lists of all eligible articles. We did not include grey literature sources in the review, as we did not expect these to be sufficiently informative to justify the extra resources required to conduct these searches.

49Two review authors (amongst JP, NL, JE, RH, JL), independently and masked to each other’s initial decisions, reviewed titles and abstracts of studies identified by the electronic searches and classified them as potentially eligible or ineligible. We used an online review management software for this purpose (Covidence). Discrepancies were resolved by discussion. We obtained full‐text articles of potentially eligible studies. Two review authors (amongst JP, NL, JE, RH, JL) independently classified them as included or excluded. Discrepancies were resolved by discussion. We recorded the study selection process in a PRISMA flow diagram, specifying reasons for exclusion of studies excluded after full‐text review. One review authors (JP) scrutinised reference lists of included studies; two independent reviewers (amongst JP, NL, JE, RH, JL) then classified studies as potentially eligible or ineligible. Two independent review authors ( amongst JP, NL, JE, RH, JL) then retrieved full‐text articles of potentially eligible studies for review and classified them as eligible or ineligible. As above, discrepancies were resolved by discussion.

50To account for heterogeneity amongst studies, data extraction involved two stages. The first stage consisted of a mapping exercise to categorise eligible studies according to their design, prognostic factors evaluated, time points of prognostic factor measurements and outcomes, and type of analysis/effect estimates. One review author (JP) undertook this stage. Information was then entered into a pilot‐tested spreadsheet specifically designed for this purpose and reviewed by the review team.

51In the second stage, data were extracted, firstly during a pilot stage, and then in full, for all eligible studies. Two review authors (amongst JP, JC, EL, NL) independently undertook data extraction. Disagreements were resolved by discussion or with the involvement of a third review author. We used the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS‐PF) to guide data extraction (Appendix 4).

52We extracted and entered the following data, if available, according to the following categories.

53StudyTitleAuthors’ contact detailsSources of fundingDatesStudy designProspective or retrospective cohort or case‐control studies and longitudinal registry dataParticipantsEligibility criteria and recruitment methodParticipant descriptionDetails of treatments received, if relevantOutcomes to be predictedDefinition and method of measurement of outcomesTypes of outcomes: 1) developing PDR; 2) progressing from less than HRC‐PDR to HRC‐PDRTime of outcome occurrencePrognostic factorsNumber and type of prognostic factorsDefinition and method for measurementTiming of prognostic factor measurementSample sizeSample size calculationNumber of participants and number of outcomesOutcomes per variableMissing dataAnalysisModelling methodResultsUnadjusted and adjusted prognostic effect estimates (e.g. risk ratio, odds ratio, hazard ratio, or mean difference) for each prognostic factor of interest and corresponding measure of uncertainty (e.g. standard errors, variances, or confidence intervals)For each extracted adjusted prognostic effect estimate of interest, the set of adjusted factors

54Study

55Title

56Authors’ contact details

57Sources of funding

58Dates

59Study design

60Prospective or retrospective cohort or case‐control studies and longitudinal registry data

61Participants

62Eligibility criteria and recruitment method

63Participant description

64Details of treatments received, if relevant

65Outcomes to be predicted

66Definition and method of measurement of outcomes

67Types of outcomes: 1) developing PDR; 2) progressing from less than HRC‐PDR to HRC‐PDR

68Time of outcome occurrence

69Prognostic factors

70Number and type of prognostic factors

71Definition and method for measurement

72Timing of prognostic factor measurement

73Sample size

74Sample size calculation

75Number of participants and number of outcomes

76Outcomes per variable

77Missing data

78Analysis

79Modelling method

80Results

81Unadjusted and adjusted prognostic effect estimates (e.g. risk ratio, odds ratio, hazard ratio, or mean difference) for each prognostic factor of interest and corresponding measure of uncertainty (e.g. standard errors, variances, or confidence intervals)

82For each extracted adjusted prognostic effect estimate of interest, the set of adjusted factors

83We used the Quality in Prognosis Studies (QUIPS) tool to assess risk of bias of the included studies (Appendix 5) (Hayden 2013). We considered six risk of bias domains: study participation, study attrition, prognostic factor measurement, outcome measurement, adjustment for other prognostic factors, and statistical analysis and reporting.

84The study participation domain consisted of six items: adequate participation in study by eligible individuals (i.e. sampling frame and recruitment adequately described, including methods to identify the sample); description of target population (i.e. source population for cohort with DR is clearly described); description of baseline study sample (i.e. number of people with DR at baseline is given); adequate description of recruitment process (i.e. way of establishing the sample population, selection criteria, and key characteristics of the population clearly described); adequate description of period and place of recruitment (time period and place of recruitment for both baseline and follow‐up examinations are clearly described); and adequate description of inclusion/exclusion criteria.

85The study attrition domain consisted of five items: adequate response rate for study participants of at least 80%; description of process for collecting information on participants who dropped out (i.e. attempts to collect information on participants who dropped out are described); reasons for loss to follow‐up provided; adequate description of participants lost to follow‐up; and no important differences between participants who completed the study and those who dropped out.

86The prognostic factor domain consisted of six items: clear definition of prognostic factor provided; method of prognostic factor measurement is adequately valid and reliable; continuous variables are reported (i.e. standard categories for prognostic factors/cut‐offs); method and setting of measurement of prognostic factor is identical for all participants; adequate proportion of study sample has complete data for prognostic factor; and appropriate methods of imputation used for missing prognostic factor data.

87The outcome measurement domain consisted of three items: clear definition of outcome provided; method of outcome measurement is adequately valid and reliable (measurement of PDR/HRC as part of a diagnostic assessment); and method and setting of outcome measurement is identical for all participants.

88The adjustment for other prognostic factors domain consisted of seven domains: all other important prognostic factors measured (i.e. HbA1c and duration of DM as a minimum); clear definitions of important prognostic factors measured provided; measurement of all important prognostic factors adequately valid and reliable; measurement and setting of prognostic factor measurement identical for all participants; appropriate methods are used to deal with missing values of prognostic factors (i.e. strategy to impute missing confounder data is described); important prognostic factors accounted for in study design; and important prognostic factors accounted for in analysis (i.e. important confounders are accounted for in multivariable logistic regression and Cox proportional hazards models).

89The statistical analysis and reporting domain consisted of four items: sufficient presentation of data to assess adequacy of analytic strategy (i.e. mean or median values, including confidence intervals or standard errors or standard deviations provided); strategy for model‐building appropriate and based on a conceptual framework or model; selected statistical model adequate for design of study (mainly incidence rates, uni‐ and multivariate logistic regression, Cox proportional hazard models); and no selective reporting of results.

90Two review authors (amongst JP, JC, EL, NL) independently assessed risk of bias. We assessed each risk of bias domain as low, moderate, or high risk, and detailed the reasoning for such assessments (see Appendix 6 for signalling questions).

91For each factor of interest, we extracted estimates of prognostic effect, such as hazard ratios (HR), risk ratios (RR), odds ratios (OR), or mean differences (MD) with a measure of their uncertainty (standard errors (SE), variances, or confidence intervals (CIs)). We collected adjusted prognostic effect estimates preferentially and documented the set of adjustment factors used.

92We contacted study authors when we required further information or clarification. When time‐to‐event analyses were performed, and adjusted hazard ratio estimates and their uncertainty were unavailable, we planned to derive unadjusted estimates and their standard errors, following guidance described by Tierney and colleagues (Tierney 2007), if the summary statistics reported permitted it.

93Between‐study heterogeneity related to two key areas:

94clinical heterogeneity, including the effect of different comorbidities, medications, and interventions in study cohorts, and differences in how outcomes were measured, such as diagnoses of PDR (clinical examination versus supported by imaging/imaging technologies used) and how progression was defined;methodological heterogeneity generated from different study designs, and how robustly studies were conducted with regard to risk of bias and approach to analysis.

95clinical heterogeneity, including the effect of different comorbidities, medications, and interventions in study cohorts, and differences in how outcomes were measured, such as diagnoses of PDR (clinical examination versus supported by imaging/imaging technologies used) and how progression was defined;

96methodological heterogeneity generated from different study designs, and how robustly studies were conducted with regard to risk of bias and approach to analysis.

97We explored the effects of these aspects of heterogeneity on the meta‐analyses we undertook.

98Since the I2 statistic can be problematic in certain situations (Rücker 2008), we planned to quantify heterogeneity using Tau2. Where there was an appropriate number of studies included in a meta‐analysis, we also planned to present 95% prediction intervals.

99We planned to assess small‐study effects using contour‐enhanced funnel plots when 10 or more studies were included in a meta‐analysis. We anticipated variation in effect measures, length of follow‐up, and other characteristics, and therefore expected to include few studies in each meta‐analysis. Consequently, we did not plan to perform funnel plot asymmetry tests given the low power of such tests when studies are few (Debray 2018).

100We conducted meta‐analysis (i.e. report a weighted average of the individual study measures of association) in clinically relevant groups using a random‐effects approach. We stratified by different time points of outcomes and meta‐analysed HR, OR, and RR separately for each prognostic factor and outcome available for meta‐analysis. Similarly, we reported unadjusted and adjusted associations separately. Our primary analyses focused on adjusted estimates, but we could only do two of this type of meta‐analysis due to insufficient data available. We present most data below in a narrative or tabulated summary because we did not identify enough studies of sufficient homogeneity to permit meta‐analysis. We used 95% confidence intervals throughout.

101We planned to perform sensitivity analyses to explore the impact of the following factors (when applicable) on effect sizes by excluding:

102studies at high risk of bias in one or more domains;retrospective studies.

103studies at high risk of bias in one or more domains;

104retrospective studies.

105However, due to the very limited data available, we were unable to perform these analyses.

106We prepared a summary of findings table assessing the certainty of the evidence using GRADE modified for prognostic factor studies (Foroutan 2020; Higgins 2022). In this table, we included all prognostic factors investigated in eligible studies for their potential role in the development of PDR using adjusted analysis in multivariable regression models. The certainty of evidence was based on grading the following domains: risk of bias; inconsistency; imprecision; and indirectness. We rated evidence down for risk of bias if more than 80% of studies included in the multivariable regression analyses had unclear or high risk of bias. We rated evidence down for inconsistency when the direction of effect estimates differed amongst studies, and for imprecision when confidence intervals were wide. If populations and outcomes evaluated did not correspond to the populations and outcomes of interest in the review, we rated the certainty of the evidence down for indirectness.

107Two independent review authors (NL, JP) conducted the GRADE assessments, and resolved any discrepancies through discussion. We used the summary of findings table to clearly identify factors that influenced the development of PDR and our confidence in the estimates of effects observed. We planned to also include HRC‐PDR in this table, but there were no data available.

108The electronic searches yielded a total of 8007 records. After removing 1776 duplicates, a total of 6231 records remained. We also identified an additional 160 records by screening the reference lists of eligible studies, and undertaking additional searches for relevant studies to supplement the Background and Discussion sections of the manuscript. In total, we screened 6391 records at the title and abstract screening stage, and determined that 6087 records were irrelevant to the review. We classified the remaining 304 records as 'potentially eligible', and retrieved full‐text articles to screen them for eligibility. We contacted the corresponding authors of three studies to request additional information necessary to determine eligibility for inclusion in our review (Leese 2004; Mathur 2016; Takaike 2018). We did not establish contact with one author (Takaike 2018); the other two authors provided further information, which allowed us to confirm that these studies were not eligible for inclusion in our review. Thus, we identified a total of 87 reports of 59 studies as eligible for inclusion in this review (see Figure 1 for study flow diagram).

109Of the included studies, several articles referred to the same original study cohorts. Thus, the WESDR study reported its findings on various prognostic factors associated with PDR in T1D and T2D at different time points, ranging from four to 25 years in 23 separate publications. The population was also divided into a “younger‐onset” group (diabetes diagnosis at < 30 years) and “older‐onset” group (diabetes diagnosis at ≥ 30 years), and further subdivided into participants “taking insulin” or “not taking insulin” (WESDR). Similarly, the studies by Grauslund 2009, Hsieh 2018, Lloyd 1995, Mathiesen 1990, Nielsen 1984, and Roy 2006 reported findings on different prognostic factors and time points in their cohorts in separate reports.

110Of the 59 unique studies included in this review, 25 were conducted in Europe (Bojestig 1998; Burditt 1968; Grauslund 2009; Gurreri 2019; Hovind 2003; Janghorbani 2000; Jones 2012; Keen 2001; Kofoed‐Enevoldsen 1987; Kullberg 1993; Lee 2017; Lestradet 1981; Mathiesen 1990; McCance 1989; Nielsen 1984; Nordwall 2015; Pirart 1977; Porta 2001; Simonsen 1980; Skrivarhaug 2006; Styles 2000; Teuscher 1988; Vesteinsdottir 2010; Voigt 2018; Zavrelova 2011), 18 in North America (Arfken 1998; Ballard 1986; Dwyer 1985; Gange 2021; Hardin 1956; Harris 2013; Klein 1984; Lee 1992; Lloyd 1995; Nelson 1989; Pambianco 2006; Rodriguez‐Villalobos 2005; Roy 2006; Rudnisky 2017; Silva 2015; Valone 1981; Varma 2010; WESDR), 13 in Asia (Chawla 2021; Chen 1995; Cho 2019; Gui 2013; Hsieh 2018; Jeng 2016; Kalter‐Leibovici 1991; Kim 1998; Kim 2014; Lee 2021; Miki 1969; Okudaira 2000; Yokoyama 1994), and one each in Africa (Burgess 2015), South America (Rodriguez‐Villalobos 2005), and Australia (McCarty 2003).

111Fifty‐seven studies were prospective cohort (n = 35) (Bojestig 1998; Burgess 2015; Chen 1995; Grauslund 2009; Hardin 1956; Hovind 2003; Hsieh 2018; Kalter‐Leibovici 1991; Keen 2001; Kim 1998; Kullberg 1993; Lee 1992; Lestradet 1981; Lloyd 1995; Mathiesen 1990; McCance 1989; McCarty 2003; Miki 1969; Nelson 1989; Nielsen 1984; Nordwall 2015; Okudaira 2000; Pambianco 2006; Pirart 1977; Porta 2001; Rodriguez‐Villalobos 2005; Roy 2006; Silva 2015; Simonsen 1980; Skrivarhaug 2006; Teuscher 1988; Valone 1981; Varma 2010) or retrospective cohort (n = 22) (Arfken 1998; Ballard 1986; Burditt 1968; Chawla 2021; Cho 2019; Gui 2013; Gurreri 2019; Janghorbani 2000; Kim 2014; Lee 2021; Rudnisky 2017; Verdaguer 2009; Vesteinsdottir 2010; Voigt 2018; WESDR; Yokoyama 1994; Zavrelova 2011) studies, with six of these based on data from electronic registers only (Dwyer 1985; Gange 2021; Harris 2013; Jeng 2016; Jones 2012; Lee 2017). The two remaining studies were retrospective case‐control studies (Kofoed‐Enevoldsen 1987; Styles 2000).

112Twenty‐three studies evaluated participants with T1D (Arfken 1998; Bojestig 1998; Chawla 2021; Grauslund 2009; Hardin 1956; Hovind 2003; Kalter‐Leibovici 1991; Klein 1984; Kofoed‐Enevoldsen 1987; Kullberg 1993; Lestradet 1981; Lloyd 1995; Mathiesen 1990; McCance 1989; Nordwall 2015; Pambianco 2006; Porta 2001; Roy 2006; Simonsen 1980; Skrivarhaug 2006; Styles 2000; Verdaguer 2009; Yokoyama 1994), 19 with T2D (Ballard 1986; Chen 1995; Cho 2019; Gange 2021; Gui 2013; Gurreri 2019; Hsieh 2018; Jeng 2016; Kim 1998; Kim 2014; Lee 1992; Lee 2021; Nelson 1989; Okudaira 2000; Rodriguez‐Villalobos 2005; Rudnisky 2017; Valone 1981; Voigt 2018; Zavrelova 2011), and 17 included mixed populations (T1D and T2D) (Burditt 1968; Burgess 2015; Dwyer 1985; Harris 2013; Janghorbani 2000; Jones 2012; Keen 2001; Lee 2017; McCarty 2003; Miki 1969; Nielsen 1984; Pirart 1977; Silva 2015; Teuscher 1988; Varma 2010; Vesteinsdottir 2010; WESDR). Of the latter, three included participants with T1D and T2D but reported outcomes for the subgroups separately (Janghorbani 2000; Nielsen 1984; WESDR). In one study, the type of diabetes (T1 or T2) was not specified (Jeng 2016).

113Studies on T1D included from 39 to 3250 participants at baseline, followed up for one to 45 years. Studies on T2D included from 100 to 71,817 participants at baseline, followed up for one to 20 years. The studies on mixed populations of T1D and T2D ranged from 76 to 32,553 participants at baseline, followed up for four to 25 years.

114We attempted to contact corresponding authors of six studies included in the review to request clarification on methodology (Chawla 2021; Gange 2021; Harris 2013; Jeng 2016) or results (Lee 2017; Lee 2021). We did not establish contact with three of these authors (Chawla 2021; Jeng 2016; Lee 2021). The others provided further information which we used in this review.

115We present detailed descriptions of the included studies in supplementary files, which can be viewed here: osf.io/sjfy5/?view_only=23c87cd105bb49639d88b90cee1e68d1.

116We excluded 217 reports of 140 studies. See Characteristics of excluded studies for details.

117We provide a summary of the risk of bias results for each of the domains (study participation, study attrition, prognostic factor measurement, outcome measurement, adjustment for other prognostic factors, and statistical analysis and reporting) in Figure 2 and Figure 3.

118Only two studies reported adequately on this domain (Janghorbani 2000; WESDR). In the remaining studies, reporting was unclear in 53 (Arfken 1998; Ballard 1986; Bojestig 1998; Burditt 1968; Burgess 2015; Chawla 2021; Chen 1995; Cho 2019; Dwyer 1985; Gange 2021; Grauslund 2009; Gui 2013; Gurreri 2019; Hardin 1956; Harris 2013; Hovind 2003; Hsieh 2018; Jones 2012; Kalter‐Leibovici 1991; Keen 2001; Kim 1998; Kim 2014; Klein 1984; Kofoed‐Enevoldsen 1987; Kullberg 1993; Lee 1992; Lee 2017; Lee 2021; Lestradet 1981; Lloyd 1995; Mathiesen 1990; McCance 1989; Miki 1969; Nelson 1989; Nielsen 1984; Nordwall 2015; Okudaira 2000; Pambianco 2006; Pirart 1977; Porta 2001; Roy 2006; Rudnisky 2017; Silva 2015; Simonsen 1980; Skrivarhaug 2006; Styles 2000; Teuscher 1988; Valone 1981; Varma 2010; Vesteinsdottir 2010; Voigt 2018; Yokoyama 1994; Zavrelova 2011), and at high risk of bias in four (Jeng 2016; McCarty 2003; Rodriguez‐Villalobos 2005; Verdaguer 2009). This was due to inadequate participation in the study by eligible individuals (n = 47/59, 80%), description of the target population (n = 45/59, 76%), description of baseline study sample (n = 34/59, 57%), description of the recruitment process (n = 44/59, 75%), the period and place of recruitment (n = 25/59, 42%), and description of the inclusion/exclusion criteria (n = 34/59, 58%).

119The risk of bias for study attrition was unclear in 25 studies (Bojestig 1998; Chen 1995; Grauslund 2009; Gui 2013; Hardin 1956; Hovind 2003; Hsieh 2018; Kim 1998; Kim 2014; Klein 1984; Lee 1992; Lloyd 1995; McCarty 2003; Nielsen 1984; Nordwall 2015; Okudaira 2000; Pambianco 2006; Porta 2001; Roy 2006; Rudnisky 2017; Skrivarhaug 2006; Teuscher 1988; Valone 1981; Varma 2010; WESDR) and high in 34 (Arfken 1998; Ballard 1986; Burditt 1968; Burgess 2015; Chawla 2021; Cho 2019; Dwyer 1985; Gange 2021; Gurreri 2019; Harris 2013; Janghorbani 2000; Jeng 2016; Jones 2012; Kalter‐Leibovici 1991; Keen 2001; Kofoed‐Enevoldsen 1987; Kullberg 1993; Lee 2017; Lee 2021; Lestradet 1981; Mathiesen 1990; McCance 1989; Miki 1969; Nelson 1989; Pirart 1977; Rodriguez‐Villalobos 2005; Silva 2015; Simonsen 1980; Styles 2000; Verdaguer 2009; Vesteinsdottir 2010; Voigt 2018; Yokoyama 1994; Zavrelova 2011). This was due to inadequate response rate for study participants (n = 48/59, 81%), inadequate description of the process for collecting information on participants who dropped out (n = 54/59, 92%), reasons for loss to follow‐up not being provided (n = 37/59, 63%), inadequate description of participants lost to follow‐up (46/59, 78%), and important differences between participants who completed the study and those who did not (n = 55/59, 93%).

120Only two studies were at low risk of bias for this domain (McCarty 2003; Porta 2001). In the remaining studies, reporting was unclear in 55 (Arfken 1998; Ballard 1986; Bojestig 1998; Burditt 1968; Burgess 2015; Chawla 2021; Chen 1995; Cho 2019; Dwyer 1985; Gange 2021; Grauslund 2009; Gui 2013; Gurreri 2019; Hardin 1956; Harris 2013; Hovind 2003; Hsieh 2018; Janghorbani 2000; Jeng 2016; Jones 2012; Kalter‐Leibovici 1991; Keen 2001; Kim 1998; Kim 2014; Klein 1984; Kofoed‐Enevoldsen 1987; Kullberg 1993; Lee 1992; Lee 2017; Lee 2021; Lestradet 1981; Lloyd 1995; Mathiesen 1990; McCance 1989; Miki 1969; Nelson 1989; Nielsen 1984; Nordwall 2015; Okudaira 2000; Pambianco 2006; Pirart 1977; Roy 2006; Rudnisky 2017; Silva 2015; Simonsen 1980; Skrivarhaug 2006; Styles 2000; Teuscher 1988; Valone 1981; Varma 2010; Vesteinsdottir 2010; Voigt 2018; WESDR; Yokoyama 1994; Zavrelova 2011), and at high risk of bias in two (Rodriguez‐Villalobos 2005; Verdaguer 2009). This was due to unclear definition of prognostic factor (n = 13/59, 22%), inadequate method of prognostic factor measurement (n = 25/59, 40%), inadequate reporting of continuous variables (n = 41/59, 69%), differences in how prognostic factors were measured (n = 33/59, 56%), inadequate proportion of study sample having complete data for prognostic factor (n = 53/59, 90%), and inappropriate methods of imputation for missing prognostic factor data (n = 57/59, 97%).

121This domain had the highest number of studies at low risk of bias (18) (Arfken 1998; Burgess 2015; Chen 1995; Grauslund 2009; Gurreri 2019; Hsieh 2018; Klein 1984; Kullberg 1993; Lee 1992; Lloyd 1995; McCance 1989; McCarty 2003; Porta 2001; Roy 2006; Silva 2015; Teuscher 1988; Varma 2010; WESDR). In the remaining studies, reporting was unclear in 39 (Ballard 1986; Bojestig 1998; Burditt 1968; Chawla 2021; Cho 2019; Dwyer 1985; Gange 2021; Gui 2013; Hardin 1956; Harris 2013; Hovind 2003; Janghorbani 2000; Jones 2012; Kalter‐Leibovici 1991; Keen 2001; Kim 1998; Kim 2014; Klein 1984; Kofoed‐Enevoldsen 1987; Lee 2017; Lee 2021; Lestradet 1981; Miki 1969; Nelson 1989; Nielsen 1984; Nordwall 2015; Okudaira 2000; Pambianco 2006; Pirart 1977; Rodriguez‐Villalobos 2005; Rudnisky 2017; Simonsen 1980; Skrivarhaug 2006; Styles 2000; Valone 1981; Verdaguer 2009; Vesteinsdottir 2010; Voigt 2018; Yokoyama 1994; Zavrelova 2011), and at high risk of bias in two (Jeng 2016; Mathiesen 1990). This was due to unclear definition of outcome (n = 44/59, 25%), unreliable method of outcome measurement (n = 28/59, 47%), and differences in method and setting of outcome measure (n = 44/59, 58%).

122We assessed only one study as being at low risk of bias in this domain (Porta 2001). In the remaining studies, reporting was unclear in 25 studies (Arfken 1998; Bojestig 1998; Cho 2019; Gange 2021; Grauslund 2009; Gui 2013; Harris 2013; Hsieh 2018; Janghorbani 2000; Kalter‐Leibovici 1991; Kim 1998; Kim 2014; Lee 1992; Lloyd 1995; Mathiesen 1990; Nelson 1989; Nordwall 2015; Okudaira 2000; Roy 2006; Silva 2015; Skrivarhaug 2006; Styles 2000; Voigt 2018; WESDR), and at high risk of bias in 33 (Ballard 1986; Burditt 1968; Burgess 2015; Chawla 2021; Chen 1995; Dwyer 1985; Gurreri 2019; Hardin 1956; Hovind 2003; Jeng 2016; Jones 2012; Keen 2001; Klein 1984; Kofoed‐Enevoldsen 1987; Lee 2017; Lee 2021; Lestradet 1981; McCance 1989; McCarty 2003; Miki 1969; Nielsen 1984; Pambianco 2006; Pirart 1977; Rodriguez‐Villalobos 2005; Rudnisky 2017; Simonsen 1980; Teuscher 1988; Valone 1981; Varma 2010; Verdaguer 2009; Vesteinsdottir 2010; Yokoyama 1994; Zavrelova 2011). In 63% of studies, HbA1c and duration of DM were not controlled for when assessing the effect of other prognostic factors.

123We determined that fifteen studies were at low risk of bias (25%) in the statistical analysis and reporting domain (Cho 2019; Gui 2013; Hsieh 2018; Janghorbani 2000; Jeng 2016; Kalter‐Leibovici 1991; Kim 1998; Kim 2014; Lee 2017; Lloyd 1995; Okudaira 2000; Porta 2001; Roy 2006; Silva 2015; WESDR). In the remaining studies, reporting of risk of bias was unclear in 39 studies (Arfken 1998; Ballard 1986; Bojestig 1998; Burditt 1968; Burgess 2015; Chawla 2021; Chen 1995; Dwyer 1985; Gange 2021; Grauslund 2009; Gurreri 2019; Hardin 1956; Harris 2013; Hovind 2003; Jones 2012; Keen 2001; Klein 1984; Kofoed‐Enevoldsen 1987; Kullberg 1993; Lee 1992; Lee 2021; Lestradet 1981; Mathiesen 1990; McCance 1989; Nelson 1989; Nielsen 1984; Nordwall 2015; Pambianco 2006; Rodriguez‐Villalobos 2005; Rudnisky 2017; Skrivarhaug 2006; Styles 2000; Teuscher 1988; Valone 1981; Varma 2010; Vesteinsdottir 2010; Voigt 2018; Yokoyama 1994; Zavrelova 2011), and high in five studies (McCarty 2003; Miki 1969; Pirart 1977; Simonsen 1980; Verdaguer 2009). This was due to insufficient presentation of data to assess adequacy of analytic strategy (n = 20/59, 34%), inadequate statistical model for study design (n = 41/59, 69%), and potentially selective reporting of results (n = 10/59, 17%)

124Twenty‐five studies investigated gender: 10 studies in people with T1D (Arfken 1998; Janghorbani 2000; Lestradet 1981; Lloyd 1995; Porta 2001; Roy 2006; Skrivarhaug 2006; Verdaguer 2009; WESDR: Klein 1989b, 1994b, 1998, 2008; Yokoyama 1994); 10 studies in people with T2D (Ballard 1986; Gange 2021; Gui 2013; Janghorbani 2000; Kalter‐Leibovici 1991; Lee 1992; Lee 2021; Nelson 1989; Rudnisky 2017; WESDR: Klein 1989a, 1994b); and five studies in mixed populations of T1D and T2D (Harris 2013; Janghorbani 2000; Jeng 2016; Keen 2001; Lee 2017).

125Six studies undertook multivariable regression analyses (Gange 2021; Harris 2013; Jeng 2016; Lee 2017; Lee 2021; Nelson 1989) (Table 2). Gender was not found to be an independent predictor of PDR in any of these studies (Table 1).

126It was only possible to undertake meta‐analysis of unadjusted effect estimates from two studies ‐ one with a mixed population of T1D and T2D (Harris 2013) and one with an unspecified DM diagnosis population (Jeng 2016) ‐ which we determined were sufficiently homogeneous with respect to study duration, type of analyses, and effect estimate provided. The pooled HR was 1.09 (95% CI 0.96 to 1.23) (Figure 4), which was consistent with the findings from multivariable analyses, in that gender was not likely to increase risk of developing PDR, with a moderate certainty of evidence.

127Other studies compared incidence of PDR based on gender in T1D (Arfken 1998; Janghorbani 2000; Lestradet 1981; Lloyd 1995; Porta 2001; Roy 2006; Skrivarhaug 2006; WESDR: Klein 1989b, 1994b, 1998 and 2008); Yokoyama 1994) and T2D (Ballard 1986; Gange 2021; Gui 2013; Janghorbani 2000; Kalter‐Leibovici 1991; Lee 1992; Rudnisky 2017; WESDR: Klein 1989a and 1994b). All, with one exception (Yokoyama 1994), similarly concurred that gender was not associated with development of PDR. The study by Yokoyama and colleagues, which included 373 participants and was conducted over the longest follow‐up period of 35 years, found a cumulative incidence of PDR of 65% in males and 81% in females. Female participants also developed PDR significantly faster than males (P < 0.002) (Yokoyama 1994).

128One study in people with T1D (Arfken 1998), three in people with T2D (Gange 2021; Kalter‐Leibovici 1991; Lee 2017), and one in a mixed population of people with T1D and T2D (Harris 2013) investigated ethnicity as a risk factor for PDR. Meta‐analysis was not possible given the heterogeneity of the studies.

129Four studies undertook multivariable regression (Arfken 1998; Harris 2013; Kalter‐Leibovici 1991; Lee 2017) (Table 3). In Arfken’s study, African American ethnicity was not statistically significantly associated with progression to PDR when compared to Caucasian (understood to be White) ethnicity (Arfken 1998). The Lee 2017 study included people of non‐Caucasian (understood to be non‐White) ethnicity (no further details provided; it states that 63.97% of the baseline population was White) and ethnicity was not found to have an effect on risk of PDR (Lee 2017). One study found non‐Ashkenazi Jews were at increased risk of developing PDR when compared to Ashkenazi Jews (Kalter‐Leibovici 1991) (Table 1). The certainty of evidence was, however, very low due to risk of bias, inconsistency, and imprecision.

130The Harris 2013 study compared proportion of participants progressing and not progressing to PDR in Whites, Blacks, Latinos, and Asians, and found no statistically significant differences. In further analyses, Black ethnicity was "statistically significantly" associated with progression to PDR in univariable regression analysis (HR 1.41, 95% CI 1.01 to 1.96; P < 0.05), as quoted in the publication, but not in multivariable regression analysis (Harris 2013).

131Gange and colleagues undertook univariable analysis only and found people of Hispanic ethnicity (P = 0.003) to be at increased risk of developing PDR, whilst those of White ethnicity had reduced risk (P = 0.005) (Gange 2021).

132Fourteen studies ‐ eight in people with T1D (Kalter‐Leibovici 1991; Nielsen 1984 (Report I); McCance 1989; Porta 2001; Styles 2000; Verdaguer 2009; WESDR: Klein 1998, 2008); Yokoyama 1994), four in people with T2D (Ballard 1986; Gange 2021; Lee 1992; Okudaira 2000), and two in a mixed population of people with TID and T2D (Janghorbani 2000; Pirart 1977) ‐ evaluated age at diagnosis of DM as a risk factor for PDR. Meta‐analysis was not possible due to heterogeneity among studies.

133Two studies ‐ Porta 2001 and Gange 2021 ‐ undertook multivariable regression analyses (Table 4). Porta and colleagues found age of diagnosis of DM at less than 12 years to be an independent risk factor for progression to PDR in people with T1D (Porta 2001). Gange and colleagues evaluated the development of PDR in the five years following diagnosis of T2D, and found that those diagnosed between 65 and 74 years of age were at higher risk compared to those diagnosed at aged 18 to 64 years and at age 75 years or higher (Gange 2021) (Table 1). However, we rated the certainty of the evidence as low due to a moderate to high risk of bias in the studies and a lack of studies preventing the grading of consistency.

134The other studies on T1D (Kalter‐Leibovici 1991; McCance 1989; Styles 2000; Verdaguer 2009; WESDR) and T2D (Ballard 1986; Lee 1992; Okudaira 2000) undertook univariate analyses only, and found age at diagnosis of DM not to be associated with progression to PDR. The exception to this was the Yokoyama 1994 study which found that participants diagnosed with DM at zero to eight years of age have a "statistically significantly" reduced risk of development of PDR than those diagnosed at nine to 17 years of age (P < 0.001) and 18 to 29 years of age (P < 0.001). However, only descriptive statistics were undertaken, and the study included fewer participants (n = 373) than Porta 2001 (n = 2013), but was conducted over a longer period (35 years) (Yokoyama 1994).

135In a mixed cohort of 3482 participants with T1D and T2D followed for five years, Janghorbani 2000 found, using crude Cox regression coefficients, that being diagnosed at 30 years old or older compared to under 30 years of age was a risk factor for developing PDR (RR 1.24, 95% CI 1.09 to 1.41; P < 0.001). However, using an analysis of variance approach, they found participants who developed PDR were diagnosed with DM at a mean age of 38.2 years (standard deviation (SD) 18.4) when compared with those that did not (mean age of 42.5 years; SD 19.7). According to the publication, the mean difference was "statistically significant" (P < 0.001) (Janghorbani 2000).

136Duration of diabetes was the most frequently evaluated prognostic factor for PDR (n = 25 studies): 15 studies in people with T1D (Grauslund 2009; Janghorbani 2000; Kalter‐Leibovici 1991; Kullberg 1993; Lloyd 1995; McCance 1989; Nielsen 1984; Pambianco 2006; Porta 2001; Roy 2006; Skrivarhaug 2006; Styles 2000; Verdaguer 2009; WESDR; Yokoyama 1994); 11 studies in people with T2D (Chen 1995; Cho 2019; Gui 2013; Janghorbani 2000; Kim 1998; Kim 2014; Lee 1992; Nelson 1989; Nielsen 1984; WESDR; Zavrelova 2011); and three in mixed populations of people with T1D and T2D (Keen 2001; Janghorbani 2000; Varma 2010). Janghorbani 2000 reported findings separately for T1D and T2D, as well as for the combined cohort, but Keen 2001 and Varma 2010 only reported on the combined population (T1D and T2D). The WESDR study reported findings on the significance of duration of diabetes in five separate publications with outcomes at four (WESDR: Klein 1989b), 10 (WESDR: Klein 1994b), 14 (WESDR: Klein 1998) and 25 (WESDR: Klein 2008) years in participants with T1D, and at four (WESDR: Klein 1989a) and 10 (WESDR: Klein 1994b) years in those with T2D. Due to heterogeneity amongst studies, meta‐analysis was not possible.

137Ten studies undertook multivariable regression analyses (Grauslund 2009; Gui 2013; Janghorbani 2000; Kalter‐Leibovici 1991; Keen 2001; Kim 1998; Kim 2014; Lee 1992; Lloyd 1995; Porta 2001) (Table 5). In the studies on T1D (Grauslund 2009; Janghorbani 2000; Kalter‐Leibovici 1991; Lloyd 1995; Porta 2001), duration of DM was only found to be an independent predictor of the development of PDR when DR severity at baseline was not included as a covariate in the models (Janghorbani 2000; Kalter‐Leibovici 1991; Porta 2001). When models were adjusted for DR severity at baseline, the effect of diabetes duration did not remain statistically significant (Grauslund 2009; Lloyd 1995; Porta 2001).

138In the studies on T2D, an association between longer duration of diabetes and increased risk of PDR was found only in studies which did not correct for HbA1c at baseline (Gui 2013; Lee 1992). None of the studies included DR severity at baseline in their models. The studies by Kim 1998 and Kim 2014 did not find duration of diabetes to be an independent predictor of development of PDR. Janghorbani and colleagues found increased risk of progression to PDR in insulin‐treated participants with 12 or more years' duration of diabetes (Janghorbani 2000).

139In mixed populations of people with T1D and T2D, Janghorbani 2000 found participants with eight or more years' duration of diabetes were at increased risk of developing PDR. Similarly, Keen 2001 found duration of diabetes (per five years) to be an independent predictor of PDR. Neither of these studies adjusted for DR severity at baseline (Janghorbani 2000; Keen 2001) (Table 1).

140We downgraded the certainty of evidence for effect of duration of DM to very low due to risk of bias in the included studies, inconsistencies in effect estimates, and imprecision.

141The other studies on T1D which undertook univariable regression analyses generally concluded that diabetes duration had an impact on the development of PDR; however, this was not necessarily linear (Janghorbani 2000; Kalter‐Leibovici 1991; McCance 1989; Nielsen 1984 (Report I); Pambianco 2006; Roy 2006; Verdaguer 2009; WESDR; Yokoyama 1994). Participants were most likely to develop PDR between 13 and 19 years of diabetes duration, with a decline in risk thereafter (Kalter‐Leibovici 1991; Roy 2006; Skrivarhaug 2006; WESDR: Klein 1989b; 1994b); Yokoyama 1994). However, in the longest‐term, 25‐year follow‐up of the WESDR study, Klein and colleagues did not find duration of DM to be associated with progression to PDR, with the cumulative incidence of 42% (95% CI 39 to 46) remaining relatively constant with duration of diabetes, whilst accounting for competing risk of death (WESDR: Klein 2008).

142Studies on T2D which undertook univariable analyses only generally concurred that increased duration of DM increased the risk of development of PDR (Chen 1995; Gui 2013; Lee 1992; Nelson 1989; Nielsen 1984), with the exception of Cho 2019, Kim 1998, and Kim 2014. Kim 1998 had a very small sample size (n = 56). Cho 2019 included 405 participants followed for a mean of four years (SD 2.0); it was unclear how many progressed to PDR (OR 1.00, 95% CI 1.00 to 1.03; P value not provided). Kim 2014 included 231 participants followed for six years (HR 1.26, 95% CI 1.00 to 1.65; P = 0.07).

143Two of the WESDR trial reports (904 participants) evaluated linear trends in the effect of diabetes duration on incidence of PDR in people with older‐onset T2D, which authors also subgrouped as requiring or not requiring insulin (WESDR: Klein 1989a and 1994b). Progression to PDR was only statistically significantly associated with duration of DM (P < 0.005) in the older‐onset group using insulin (WESDR: Klein 1989a and 1994b). Varma 2010 included 324 participants with T1D and T2D; only 17 developed PDR during the four‐year follow‐up. Duration of diabetes was not found to be statistically significantly associated with development of PDR (P = 0.08) (Varma 2010).

144Only two studies compared development of PDR in participants with T1D and T2D (Janghorbani 2000; Keen 2001). Both undertook multivariable regression analyses (Table 33). Janghorbani 2000 found T1D to be associated with a decreased risk of developing PDR. Similarly, Keen 2001 found T1D to be associated with decreased risk of progression to PDR, but only when fasting plasma glucose was included in the model. However, we downgraded the certainty of the evidence to very low due to risk of bias in included studies, inconsistencies in effect estimates, and imprecision.

145Socioeconmic status as a risk factor for PDR was investigated in two studies in people with T1D (Roy 2006; WESDR: Klein 1994a), four in people with T2D (Chen 1995; Gange 2021; Kalter‐Leibovici 1991; WESDR: Klein 1994a), and one in a mixed population of people with T1D and T2D (Harris 2013). The studies used different methods to assess socioeconomic status and, thus, meta‐analysis was not possible. With the exception of the Kalter‐Leibovici 1991 study, socioeconomic status was not found to be a risk factor for PDR (Chen 1995; Gange 2021; Harris 2013; Roy 2006; WESDR: Klein 1994a).

146Only the WESDR study undertook multivariable regression analyses (WESDR: Klein 1994a) (Table 6). Socioeconomic status was determined using the Duncan Socioeconomic Index, which assigns a score according to occupation or the spouse’s occupation, if married but not working, with higher scores indicating higher socioeconomic status (Stevens 1981). Socioeconomic status was not associated with incidence of PDR in males or females in T1D or T2D (WESDR: Klein 1994a). We downgraded the certainty of the evidence to very low due to risk of bias in included studies and imprecision. It was not possible to grade consistency due to only one study having undertaken multivariable regression analysis (Table 1).

147The study by Kalter‐Leibovici and colleagues, which included 330 Jewish participants with T1D at baseline followed for 10 years, was the only study finding a negative correlation between socioeconomic status and progression to PDR. Using an analysis of variance, an increased percentage of participants who progressed to PDR had a family income of less than the national average (P < 0.001) (Kalter‐Leibovici 1991).

148Only two studies evaluated the effect of education, as a stand‐alone variable, on the development of PDR (Gange 2021; WESDR: Klein 1994a). Meta‐analysis was not possible due to heterogeneity between the studies.

149Only the WESDR study undertook multivariable regression analysis at the four‐year follow‐up (WESDR: Klein 1994a) (Table 7). Different levels of education were considered: no high school degree, high school degree, some college, and college graduate; males and females were analysed separately. In females in the younger‐onset group, for every five or more years of education, there was a statistically significantly decreased probability of developing PDR, but this was not the case in males. In the older‐onset group, education was not associated with incidence of PDR in males or females (WESDR: Klein 1994a). We downgraded the certainty of the evidence to very low due to risk of bias in the included studies and imprecision. It was not possible to grade consistency due to only one study having undertaken multivariable regression analysis (Table 1).

150In the WESDR report on outcomes at 25 years, males and females were considered collectively (n = 481) and only univariate analysis was undertaken. Education level (per four years) was not associated with incidence of PDR (HR 1.05, 95% CI 0.94 to 1.19; P = 0.38) (WESDR: Klein 2008).

151Gange and colleagues conducted a study involving 71,817 participants with T2D, followed for five years. Only univariate analysis was undertaken and, although there was a trend indicating that participants with PDR had lower levels of education, differences were not statistically significant (P > 0.05) (Gange 2021).

152The relationship between HbA1c level and the development of PDR was assessed in 24 studies: 12 studies in people with T1D (Arfken 1998; Janghorbani 2000; Kullberg 1993; Lloyd 1995; McCance 1989; Nordwall 2015; Porta 2001; Roy 2006; Skrivarhaug 2006; Styles 2000; Verdaguer 2009; WESDR: Klein 1988, 1994, 1998, 2008); 10 studies in people with T2D (Chen 1995; Cho 2019; Gange 2021; Gui 2013; Kalter‐Leibovici 1991; Kim 2014; Lee 2021; Okudaira 2000; WESDR: Klein 1988, 1994); Zavrelova 2011); and two studies in a mixed population of T1D and T2D (Harris 2013; Janghorbani 2000).

153Eight studies on T1D undertook multivariable logistic regression analyses (Arfken 1998; Janghorbani 2000; Kullberg 1993; Lloyd 1995; Porta 2001; Roy 2006; Skrivarhaug 2006; WESDR: Klein 1988a, 1994c, 1998, 2008). All studies found that increased HbA1c levels were associated with progression to PDR. DR severity at baseline was included as a co‐variate in most of the studies (Arfken 1998; Lloyd 1995; Porta 2001; Skrivarhaug 2006; WESDR: Klein 1988, 1994, 1998) (Table 9).

154Eight studies on T2D undertook multivariable logistic regression analyses (Cho 2019; Gange 2021; Kalter‐Leibovici 1991; Kim 1998; Kim 2014; Lee 2021; Okudaira 2000; WESDR: Klein 1988, 1994c). Similarly to findings in T1D, these studies generally found a positive correlation between increased HbA1c and incidence of PDR. The exceptions were the studies by Cho 2019 and Lee 2021, which found HbA1c level to be statistically significantly associated with progression to PDR in univariable, but not in multivariable, regression analyses. Only two studies corrected for DR severity at baseline (Lee 2021; WESDR: Klein 1988, 1994c) (Table 9).

155It was only possible to undertake meta‐analysis of unadjusted effect estimates combining three studies including a total of 2955 participants with T2D, followed for six to seven years (Kim 2014; Lee 2021; Okudaira 2000). None of these studies included DR severity at baseline in their models. The pooled HR was 1.40 (95% CI 1.20 to 1.63) and the 95% prediction interval ranged from 0.31 to 6.36 (Figure 4).

156Two studies on mixed populations of people with T1D and T2D undertook multivariable regression analyses (Harris 2013; Janghorbani 2000). Harris and colleagues found that for every one per cent point increase in HbA1c, the risk of developing PDR was increased by 14% (Harris 2013). Janghorbani and colleagues found that HbA1c of 11% or higher was an independent predictor of PDR. Additionally, the mean difference in HbA1c values between non‐progressors and progressors to PDR was also statistically significant (MD 0.8%, 95% CI 0.42 to 1.18; P < 0.001), as quoted in the publication (Janghorbani 2000). Neither of these studies corrected for DR severity at baseline (Table 8). We rated the evidence for HbA1c as of moderate certainty because of moderate to high risk of bias in the studies included in the analyses.

157In general, studies using univariable analyses supported the above findings (McCance 1989; Nordwall 2015; Verdaguer 2009). The exception to this was the Styles 2000 study, which had the longest follow‐up (45 years) and did not find a significant mean difference in HbA1c values between progressors (mean 8.3%; range 5.8 to 12.3) and non‐progressors (mean 8.7%; range 6.7 to 11.9) to PDR (P = 0.16) (Styles 2000).

158Some studies attempted to establish whether there was an HbA1c threshold below which progression to PDR would not occur, but this threshold was not found (Porta 2001; Roy 2006; WESDR: Klein 1988, 1994c). However, a longer‐term study by Nordwall and colleagues recommended HbA1c to be below 7.6% to prevent PDR, as none of the 451 participants in their study with levels below 7.6% developed PDR during the study period of 22 years (Nordwall 2015).

159The WESDR study evaluated whether there was a relationship between change in glycaemic control and the risk of progression to PDR, whilst controlling for DR severity and HbA1c at baseline, and hypertension. The OR for a one percentage point increase in HbA1c from baseline to four‐year follow‐up was 1.33 for progression to PDR, equivalent to a 25% increase in the 14‐year incidence of PDR (WESDR: Klein 1998). In the longest‐term report of 25 years, a one percentage point decrease in HbA1c from baseline to four‐year follow‐up was associated with an 18% decrease in the 21‐year rate of progression to PDR (WESDR: Klein 2008). The data from the WESDR reports also suggested that reducing levels of HbA1c, even later during DM, or even when moderate NPDR is present, may reduce the risk of progression to PDR already conferred by higher HbA1c in previous years (WESDR: Klein 1994c, 1998).

160The Gui 2013 study included 190 participants with T2D followed for two years and found higher mean HbA1c in progressors (mean 11.43%; SE 3.09) compared to non‐progressors (mean 7.43%; SE 3.14) to PDR, but the difference was not significant (P = 0.1) (Gui 2013). Zavrelova and colleagues found mean HbA1c values for progressors (9.1%) to be higher than that of non‐progressors (8.4%) (Zavrelova 2011).

161Seven studies evaluated fasting plasma glucose as a prognostic factor for the development of PDR: six in participants with T2D (Ballard 1986; Chen 1995; Cho 2019; Kim 1998; Lee 1992; Lee 2021), and one in a mixed population of people with T1D and T2D (Keen 2001).

162Three studies undertook multivariable regression analyses (Keen 2001; Lee 1992; Lee 2021) (Table 10). In T2D, Lee and colleagues did not find fasting plasma glucose to be an independent predictor for the incidence of PDR (Lee 2021). Conversely, in the Lee 1992 study, fasting plasma glucose was predictive of the development of PDR (Lee 1992). However, HbA1c and DR severity at baseline were not adjusted for in the latter study (Lee 1992). In a mixed population of T1D and T2D, Keen and colleagues also found fasting plasma glucose to be significantly associated with development of PDR, but again, HbA1c and DR severity at baseline were not included as covariates (Keen 2001) (Table 10). Overall, evidence from multivariable regression analysis is very uncertain about the effect of fasting plasma glucose on the risk of developing PDR (Table 8).

163It was only possible to undertake meta‐analysis of unadjusted effect estimates from two studies of T2D (Kim 2014; Lee 2021), which we determined to be sufficiently homogeneous with respect to study duration, type of analyses, and effect estimate provided. The pooled HR was 1.03 (95% CI 0.79 to 1.33) (Figure 4).

164Some studies undertook univariable regression analyses. Ballard 1986 and Chen 1995 evaluated cumulative incidence with fasting plasma glucose and determined a positive correlation between increasing levels of fasting plasma glucose and development of PDR. In univariate analysis, Cho 2019 and Lee 2021 did not identify fasting plasma glucose as being associated with progression to PDR. The Lee 1992 study compared the mean level of fasting plasma glucose in progressors (12.5 mmol/L) versus non‐progressors (9.6 mmol/L) to PDR, and identified a difference in values (P < 0.001) (Lee 1992). Conversely, the Kim 1998 study did not detect any difference in mean fasting plasma glucose between the group that developed PDR (12.6 mmol/L; SD 5.7) and the group that remained stable (11.5 mmol/L; SD 4.5) (Kim 1998).

165Twelve studies evaluated the relationship between DBP and the incidence of PDR: four in people with T1D (Grauslund 2009; Porta 2001; Roy 2006; WESDR); seven in people with T2D (Kim 1998; Kim 2014; Lee 1992; Lee 2021; Okudaira 2000; WESDR; Zavrelova 2011); and two in a mixed population of participants with T1D and T2D (Janghorbani 2000; Keen 2001).

166Seven studies undertook multivariable regression (Grauslund 2009; Keen 2001; Lee 2021; Okudaira 2000; Porta 2001; Roy 2006; WESDR: Klein 1989c) (Table 11).

167In T1D, DBP was only found to be an independent predictor of development of PDR when DR severity at baseline was not included in the models (Porta 2001; Roy 2006). The WESDR study and the Grauslund 2009 study included DR severity at baseline and DBP was not found to be statistically significantly associated with progression to PDR (Grauslund 2009; WESDR: Klein 1989c).

168In T2D, only two studies undertook multivariable regression analyses (Lee 2021; Okudaira 2000), and found that DBP was predictive of the development of PDR. However, only the Lee 2021 study corrected for DR severity at baseline. Evidence from multivariable regression analyses suggesting that DBP is associated with progression to PDR is very uncertain (Table 8).

169Two studies on T2D were appropriate for meta‐analysis with respect to study duration, type of analyses, and effect estimate provided (Hsieh 2018; Lee 2021). The pooled HR of adjusted HR estimates (HbA1c was the only common adjustment factor) was 1.07 (95% CI 0.96 to 1.18) (Figure 4).

170In T1D and T2D, Keen 2001 found no association between DBP (per 5 mmHg increase) and PDR in multivariable regression analysis.

171Other studies undertook univariable regression analyses only. The WESDR study also evaluated linear trends and found that increasing DBP was positively correlated with increasing incidence of PDR in T1D at four (P < 0.001), 10 (P < 0.001), and 14 (P < 0.001) years of follow‐up (WESDR: Klein 1989c, 1995a, 1998). In the 25‐year report, DBP (per 10 mmHg increase) was found to be associated with incidence of PDR (HR 1.3, 95% CI 1.16 to 1.46; P < 0.001) at the univariable level (not considered in multivariable analysis in this report) (WESDR: Klein 2008).

172In participants with T2D, at four years' follow‐up, the relationship between increasing DBP and incidence of PDR was not significant, as evaluated by linear trends, even after taking into account the use of antihypertensive medications (WESDR: Klein 1989c). At ten years, higher levels of DBP were associated with increased risk of developing PDR (P < 0.05) in the older‐onset group taking insulin. In other studies on T2D which undertook univariate analyses only (Kim 1998; Kim 2014; Lee 1992; Zavrelova 2011), DBP was not found to be a risk factor for PDR.

173The Janghorbani 2000 study (n = 3482 T1D and T2D; follow‐up of five years) found DBP of between 85 mmHg and 94 mmHg was associated with an increased risk of progression to PDR (P < 0.05), but levels higher than 95 mmHg were not. Using analysis of variance, they found participants who progressed to PDR had lower levels of DBP (75.1 mmHg; SD 28.4) compared to those who did not progress to PDR (78.3 mmHg; SD 23.4) (P < 0.05) (Janghorbani 2000).

174Thirteen studies in participants with T1D (Arfken 1998; Grauslund 2009; Janghorbani 2000; Porta 2001; Roy 2006; WESDR: Klein 1989c, 1998, 2008), T2D (Kim 1998; Kim 2014; Lee 1992; Lee 2021; Okudaira 2000; WESDR: Klein 1989c; Zavrelova 2011), and mixed populations of both T1D and T2D (Janghorbani 2000; Keen 2001) evaluated SBP as a risk factor for PDR.

175Eight studies undertook multivariable regression (Grauslund 2009; Janghorbani 2000; Keen 2001; Lee 1992; Lee 2021; WESDR: Klein 1989, 1995a, 2008) (Table 12).

176In T1D, SBP was only found to be an independent predictor of development of PDR when DR severity at baseline was not included in the models (Janghorbani 2000; WESDR: Klein 2008). In the studies by Grauslund and WESDR, which included DR severity at baseline, SBP was not significantly associated with progression to PDR (Grauslund 2009; WESDR: Klein 1989c, 1995a).

177In the studies on T2D, SBP was not found to be an independent predictor of development of PDR (Lee 1992; Lee 2021; WESDR: Klein 1989c). Three studies on T2D were also appropriate for meta‐analysis with respect to study duration, type of analyses, and effect estimate provided (Kim 2014; Lee 2021; Okudaira 2000). The pooled HR of unadjusted HR estimates was 1.02 (95% CI 0.91 to 1.14). The 95% prediction interval ranged from 0.32 to 3.21 (Figure 4).

178Janghorbani and colleagues additionally reported findings for the entire cohort of participants with T1D and T2D: SBP above 160 mmHg was significantly associated with increased progression to PDR (Janghorbani 2000). In contrast, the Keen 2001 study, which also included people with T1D or T2D, did not find SBP (increments of 10 mmHg) to be an independent predictor for the development of PDR (Keen 2001). Neither study corrected for DR severity at baseline and Keen 2001 did not adjust for HbA1c. Overall, evidence from multivariable regression analyses suggesting that DBP is associated with progression to PDR is very uncertain (Table 8).

179In univariable analyses, the WESDR study additionally reported a statistically significant linear trend in the incidence of PDR in participants with T1D, with increasing SBP at the four‐, 10‐, and 14‐year periods (P < 0.001) (WESDR: Klein 1989c, 1995a, Report 1998) (linear trend analysis not reported in the WESDR Klein 2008 25‐year outcomes). Other potential confounders were not considered in these analyses.

180Other studies undertook univariable regression analyses only (Arfken 1998; Porta 2001; Roy 2006). Roy 2006 found that participants with T1D in the upper quartile of SBP (≥ 135 mmHg) at baseline had increased risk of progression to PDR during the six‐year follow‐up than those in the lowest quartile (≤ 110 mmHg) (OR 3.09, 95% CI 1.37 to 7.00; P = 0.05) (Roy 2006). Arfken 1998 followed 312 participants for six years and found an increase in SBP in participants progressing to PDR (mean 120 mmHg, SD 21) compared to those that did not (mean 111, SD 16; P = 0.01) (Arfken 1998). Conversely, when comparing participants who progressed to PDR (median: 119 mmHg; 25thand 75th percentiles: 108 and 134, respectively) with those that did not (median: 117 mmHg; 25thand 75th percentiles: 108 and 128, respectively) during a seven‐year observation period, Porta 2001 did not find statistically significant differences in SBP between groups (P = 0.10).

181In the WESDR study, in analysis of linear trends at four years, SBP was not associated with incidence of PDR in either the insulin‐taking or non‐insulin‐taking groups (WESDR: Klein 1989c). At 10 years, in the older‐onset group taking insulin, higher SBP was significantly associated with development of PDR (P < 0.05), but not in the older‐onset group not taking insulin (P = 0.20) (WESDR: Klein 95a).

182Two studies considered the effect of mean arterial blood pressure on the development of PDR (Chen 1995; Roy 2006).

183In multivariable analysis, including HbA1c, age, sex, socioeconomic status, BMI, central retinal artery equivalent (CRAE), ocular perfusion pressure, and refractive error as covariates, Roy 2006 determined that mean arterial blood pressure (per 10 mmHg change) was associated with incidence of PDR during the six‐year study period (OR 1.35, 95% CI 0.91 to 2.00; P < 0.001). However, although the P value was statistically significant, the 95% CIs ranged from being protective to determining increased risk (Table 8). In the Chen 1995 study (follow‐up period not reported), the cumulative incidence of PDR was not associated with increasing quartiles of mean arterial blood pressure (P = 0.13), but other variables were not corrected.

184Four studies considered the effect of dyslipidaemia on progression to PDR (Gange 2021; Harris 2013; Jeng 2016; Verdaguer 2009). The Gange 2021 and Verdaguer 2009 studies included participants with T1D and T2D, respectively, and the Harris 2013 study a mixed cohort of participants with T1D and T2D; Jeng 2016 did not provide information on type of diabetes. None found an association between dyslipidaemia and incidence of PDR (Gange 2021; Harris 2013; Jeng 2016; Verdaguer 2009).

185We undertook multivariable regression analysis in only two of these studies, which were sufficiently homogeneous (Harris 2013; Jeng 2016) (Table 13) (Table 8). The pooled HR of the unadjusted effect estimates was 1.21 (95% CI 0.62 to 2.37), consistent with the negative correlation between dyslipidaemia and development of PDR found in the other studies (Figure 4).

186The remaining two studies undertook univariable analyses only (Gange 2021; Verdaguer 2009). Verdaguer 2009 included only 39 participants with T1D followed for 18 years. The study compared dyslipidaemia in the group with NPDR to the group with PDR and the difference was not significant (P = 0.133) (Verdaguer 2009). Gange 2021 included 71,817 participants with T2D. The percentage of participants with dyslipidaemia who progressed to PDR (25%) in the five‐year period subsequently to being diagnosed with DM was not significantly different to that of those who did not progress (25.6%, P = 0.61) (Gange 2021).

187Eleven studies evaluated total cholesterol as a risk factor for PDR: two studies in people with T1D (Porta 2001; Roy 2006); eight studies in people with T2D (Gui 2013; Kim 1998; Kim 2014; Lee 1992; Lee 2021; Nelson 1989; Okudaira 2000; Zavrelova 2011); and one study in a mixed group of people with T1D and T2D (Keen 2001).

188Five studies undertook multivariable regression analyses (Keen 2001; Lee 1992; Lee 2021; Nelson 1989; Porta 2001) (Table 14). Only Keen 2001 and Nelson 1989 found an association between increased total cholesterol and development of PDR. However, HbA1c and DR severity at baseline were not included in their models (Keen 2001; Nelson 1989) (Table 14).

189The Lee 1992, Lee 2021, and Porta 2001 studies did not find total cholesterol to be an independent predictor of PDR. Porta 2001 found that mean total cholesterol level was elevated in participants who progressed to PDR (mean 5.6; SE 0.1) when compared to those that did not (mean 5.1; SE 0.03; P < 0.001). However, when adjustment for HbA1c and duration of diabetes was undertaken, the association was no longer "statistically significant" (reported descriptively) (Porta 2001). Overall evidence from multivariable regression analyses suggesting that total cholesterol is associated with progression to PDR is very uncertain (Table 8).

190Additionally, meta‐analysis was possible in three other studies on T2D which we considered to be homogeneous with regard to study duration, type of analyses, and effect estimate provided (Kim 2014; Lee 2021; Okudaira 2000). The pooled HR of the unadjusted effect estimates was 1.00 (95% CI 1.00 to 1.01). The 95% prediction interval ranged from 0.97 to 1.03 (Figure 4).

191Other studies included total cholesterol in univariable analyses only, and all but Roy 2006 did not find total cholesterol to be associated with PDR (Gui 2013; Kim 1998; Kim 2014; Okudaira 2000; Roy 2006; Zavrelova 2011).

192Ten studies investigated the impact of triglyceride level on development of PDR: three studies in people with T1D (Lloyd 1995; Porta 2001; Skrivarhaug 2006), and seven studies in people with T2D (Gui 2013; Kim 1998; Kim 2014; Lee 1992; Lee 2021; Okudaira 2000; Zavrelova 2011).

193Three studies undertook multivariable regression (Lee 2021; Porta 2001; Skrivarhaug 2006) (Table 15). In T1D, triglycerides appeared to be an independent predictor of PDR (Porta 2001; Skrivarhaug 2006), but not in the study on T2D (Lee 2021) (Table 8). However, the certainty of evidence was low due to risk of bias in the included studies and imprecision in some studies (wide CIs).

194Additionally, we undertook meta‐analysis combining three studies on T2D which were homogeneous with regard to study duration, analyses type, and effect estimates provided (Kim 2014; Lee 2021; Okudaira 2000). The pooled HR of the unadjusted effect estimates was 11.01 (95% CI 0.95 to 1.07), which was consistent with the non‐significant finding in the multivariable regression analysis (Lee 2021). The 95% prediction interval ranged from 0.59 to 1.73 (Figure 4).

195Some studies undertook univariable analyses. In T1D, Lloyd 1995 found serum triglyceride levels (it is not reported whether participants were fasting) were elevated in the group that progressed to PDR (mean 2.0 mg/dL; SD 0.3) compared to the group that did not (mean 1.9 mg/dL; SD 0.2; P < 0.05) whilst adjusting for duration of diabetes (Lloyd 1995). Porta 2001 found triglycerides (fasting and non‐fasting levels) were increased in participants with T1D progressing to PDR (fasting and non‐fasting triglycerides: progressors 1.15 mmol/L versus non‐progressors 0.92 mmol/L; P < 0.001; fasting triglycerides: progressors 1.11 mmol/L versus non‐progressors 0.88 mmol/L; P < 0.001) (Porta 2001).

196The Lee 1992 study grouped participants with T2D as those with or without plasma triglyceride levels of 22.6 mg/dL or higher, and found the former had an increased risk of PDR (RR 1.7, 95% CI 1.1 to 2.62; P = 0.015). When stratified by duration of diabetes, triglyceride level remained significantly associated with development of PDR (P < 0.05) (Lee 1992). However, also in participants with T2D, Okudaira 2000 (HR 1, 95% CI 1 to 1; P = 0.90) and Kim 2014 (HR 0.51, 95% CI 0.34 to 2.32; P = 0.254) did not find triglyceride levels to be associated with incidence of PDR (Kim 2014; Lee 1992; Okudaira 2000). Similarly, studies which compared mean triglyceride values at baseline in progressors to PDR compared to non‐progressors did not find a statistically significant difference. Other potential risk factors were not corrected (Gui 2013; Kim 1998; Kim 2014; Zavrelova 2011).

197Seven studies evaluated the effect of LDL on the development of PDR: three in people with T1D (Lloyd 1995; Porta 2001; Roy 2006), and four in people with T2D (Gui 2013; Kim 2014; Lee 2021; Zavrelova 2011).

198Only two studies undertook multivariable regression analyses (Lee 2021; Porta 2001). They did not find LDL to be an independent predictor for the development of PDR, but overall evidence was very uncertain about the effect (Lee 2021; Porta 2001) (Table 16) (Table 8).

199Additionally, we undertook meta‐analysis combining two studies in people with T2D which were sufficiently homogeneous with respect to study duration, analyses type, and effect estimates provided (Kim 2014; Lee 2021). The pooled HR of the unadjusted effect estimates was 0.98 (95% CI 0.86 to 1.12), which was consistent with the multivariable regression analysis by Lee 2021 which did not find LDL to be a prognostic factor for PDR (Figure 4).

200Some studies undertook univariable analyses. In people with T1D, Roy 2006 found that higher LDL levels were associated with progression to PDR. Participants with LDL levels at baseline in the upper quartile had approximately three times the rate of progression to PDR than those in the lowest quartile (P = 0.02) (Roy 2006). Also in people with T1D, Lloyd 1995 determined that progressors to PDR had a higher LDL (mean value 121.3 mg/dL, SD 31.1) than non‐progressors (mean value 106.1 mg/dL, SD 28.6; P < 0.05) when adjusting for duration of diabetes only.

201In the studies in people with T2D, there was no significant difference in mean values of LDL in progressors to PDR compared to non‐progressors (Gui 2013; Kim 1998; Zavrelova 2011).

202Five studies evaluated the impact of HDL on the development of PDR: one study in people with T1D (Porta 2001), and four studies in people with T2D (Kim 1998; Kim 2014; Lee 2021; Zavrelova 2011).

203Only two studies undertook multivariable regression analyses and neither found HDL to be an independent predictor of PDR, but overall evidence was very uncertain about the effect (Lee 2021; Porta 2001) (Table 17) (Table 8).

204Additionally, we undertook meta‐analysis combining two studies in people with T2D which were sufficiently homogeneous with respect to study duration, analyses type, and effect estimates provided (Kim 2014; Lee 2021). The pooled HR of the unadjusted effect estimates was 0.84 (95% CI 0.73 to 0.96), suggesting a reduced risk of PDR with higher HDL values (Figure 4).

205The remaining two studies undertook univariable analyses only, and only in people with T2D. These studies found reduced mean HDL levels in progressors to PDR versus non‐progressors (Kim 1998; Zavrelova 2011).

206Only two studies including people with T1D explored the effect of fibrinogen on progression to PDR (Lloyd 1995; Porta 2001). Both compared fibrinogen levels in progressors with those in non‐progressors to PDR. Lloyd 1995 identified an increased fibrinogen level in progressors (mean value 2.5, SD 0.1) compared to non‐progressors (mean value 2.4, SD 0.1; P < 0.01). In contrast, Porta 2001 found no statistically significant difference in fibrinogen levels between progressors (mean value 3.22, SE 0.1) and non‐progressors (mean value 3.17, SE 0.03; P = 0.6)). The effect of other potential risk factors was not considered in these analyses, except for duration of DM which was included in Lloyd 1995.

207The included studies that looked at this risk factor adopted different ways of assessing the effect of kidney function on the development of PDF. We describe these below.

208Five studies considered the effect of nephropathy or renal disease on the incidence of PDR: two in people with T2D (Gange 2021; Lee 1992), and three in mixed populations of people with T1D or T2D (Harris 2013; Jeng 2016; Keen 2001).

209Four studies undertook multivariable regression analyses (Gange 2021; Harris 2013; Jeng 2016; Keen 2001) (Table 18). In the study on T2D, renal disease was found to be an independent risk factor for PDR (Gange 2021). In the studies on mixed populations, the results were variable. Jeng 2016 and Keen 2001 identified a positive correlation between diabetic nephropathy and incidence of PDR. However, their definitions of nephropathy differed. Jeng 2016 defined nephropathy as “persistent albuminuria, progressive decline of GFR [glomerular filtration rate], and elevation of BP [blood pressure]”, whereas Keen 2001 assessed the intensity of protein precipitation in urine as an indicator of renal disease severity. Harris 2013 did not find a statistically significant association between renal disease and development of PDR. We assessed the available evidence based on multivariable regression analyses as very uncertain about the effect of nephropathy on the risk of developing PDR, due to risk of bias, inconsistency, and imprecision in studies (Table 8).

210The Harris 2013 and Jeng 2016 studies were sufficiently homogeneous to meta‐analyse. The pooled HR of the unadjusted estimates was 4.18 (95% CI 0.62 to 28.43), suggesting that the effect of renal disease on progression was not significant (Figure 4).

211Other studies undertook univariable regression analyses in T2D. Whereas Gange 2021 and Harris 2013 found an increased percentage of progressors to PDR in people with renal disease (Gange 2021: 6.7% versus 1.6%; P < 0.001; Harris 2013: 39.1% versus 26%; P < 0.05), Lee 1992 did not identify a relationship between renal disease (defined as creatinine > 133 micrometre [µM] or proteinuria) and the development of PDR (RR 1.19, 95% CI 0.72 to 1.97; P = 0.501).

212Seven studies studied the impact of proteinuria on the incidence of PDR: four studies in people with T1D (Grauslund 2009; Kofoed‐Enevoldsen 1987; Roy 2006; WESDR: Klein 1993, 1998, 2008), reporting outcomes at four, 14, and 25 years' follow‐up, in separate publications); three studies in people with T2D (Chen 1995; Nelson 1989; WESDR); and one study in people with T1D or T2D (Janghorbani 2000). Meta‐analysis was not possible due to study heterogeneity.

213Five studies undertook multivariable regression (Grauslund 2009; Janghorbani 2000; Nelson 1989; Roy 2006; WESDR: Klein 1993, Reports XVII and XXII)) (Table 19). All but Grauslund 2009 and WESDR (Klein 1993) found proteinuria to be an independent predictor of PDR. Again, we deemed the certainty of evidence based on multivariable regression analyses to be very uncertain due to risk of bias, inconsistency, and imprecision in studies (Table 8).

214In the WESDR study reporting outcomes in people with T1D at four years, the presence of gross proteinuria (defined as urine protein concentration of 0.30 g/L or greater as measured by a reagent strip) at baseline was associated with increased risk of PDR. However, this association was only of borderline significance in people with no or mild NPDR at baseline (WESDR: Klein 1993). At 14 years (WESDR: Klein 1998) and 25 years (WESDR: Klein 2008), gross proteinuria at baseline remained a statistically significant risk factor for the development of PDR. Roy 2006 determined that participants with overt proteinuria (albumin excretion rate (AER) > 200 µg/min) at baseline had four times the risk of developing PDR.

215In people with T2D, Nelson 1989 found proteinuria (protein‐to‐creatinine ratio of ≥ 113 mg/mmol) to be statistically significantly associated with development of PDR. However, the WESDR study found that, in the older‐onset group taking insulin with moderate or severe NPDR, gross proteinuria was not correlated with incidence of PDR (WESDR: Klein 1993).

216Five studies investigated the effect of AER on progression to PDR: four studies in people with T1D (Lloyd 1995; Mathiesen 1990; Porta 2001; Roy 2006), and one study in people with T2D (Kim 1998). Meta‐analysis was not possible due to the heterogeneity of these studies.

217Three studies undertook multivariable regression analyses (Kim 1998; Porta 2001; Roy 2006) (Table 20). Both studies including people with T1D found higher AER to be an independent predictor of PDR (Porta 2001; Roy 2006). In the study on T2D (Kim 1998), baseline AER was not associated with development of PDR.(Table 8).

218Other studies undertook univariable analyses. In people with T1D, the Lloyd 1995 study compared mean AER in progressors to that in non‐progressors to PDR, whilst adjusting for duration of DM. Those who developed PDR had a statistically significantly increased mean AER (progressors versus non‐progressors: mean 1.7 LOG, SD 1.0 versus 1.2 LOG, SD 0.7; P < 0.001) (Lloyd 1995). Porta 2001 also found a significantly increased mean AER in progressors compared to non‐progressors (progressors versus non‐progressors: median 29 µg/min versus 12 µg/min; P < 0.001) (Porta 2001). Similarly, in people with T2D, the Kim 1998 study found that participants who progressed to PDR had increased AER at baseline compared to the non‐progressors (progressors versus non‐progressors: mean 67 µg/min, SD 61 versus 23 µg/min, SD 29; P < 0.05) (Kim 1998).

219Only three studies (Hsieh 2018; Kim 2014; Zavrelova 2011), all in people with T2D, evaluated the effect of values of albumin creatinine ratio on incidence of PDR. Meta‐analysis was not possible due to the studies' heterogeneity.

220Only two studies undertook multivariable regression analyses (Hsieh 2018; Kim 2014). Both studies found increasing albumin creatinine ratio to be an independent predictor of PDR (Hsieh 2018; Kim 2014) (Table 21). The certainty of the evidence is moderate, suggesting albumin creatinine ratio is likely associated with increased risk of progression to PDR in T2D (Table 8).

221Two studies (Cho 2019; Hsieh 2018), both in people with T2D, evaluated the effect of eGFR on development of PDR and both undertook multivariable regression analyses (Table 22) (Table 8). Meta‐analysis was not possible due to their heterogeneity. Cho 2019 found a reduction in eGFR of more than 20% (but not in mean eGFR) to be associated with development of PDR. Hsieh 2018 found that participants with decreased mean follow‐up eGFRs were at increased risk of incident PDR.

222Six studies considered the effect of creatinine on incidence of PDR: one studies in people with T1D (Styles 2000); four studies in people with T2D (Hsieh 2018; Lee 2021; Nelson 1989; Zavrelova 2011); and one in a mixed population of people with T1D and T2D. Three studies undertook multivariable regression analyses, all in T2D (Hsieh 2018; Lee 2021; Nelson 1989) (Table 23). Hsieh 2018 and Nelson 1989 found elevated creatinine levels to be an independent predictor of PDR, in contrast to the Lee 2021 study. However, we deemed the certainty of evidence as very low, due to risk of bias in included studies, inconsistency, and imprecision (Table 8). The Hsieh 2018 and Lee 2021 studies were appropriate for meta‐analysis with respect to study duration, type of analyses, and effect estimate provided, with HbA1c, age, gender, and BMI being common adjustment factors. The pooled HR of adjusted HR estimates was 1.61 (95% CI 0.77 to 3.36), although there appeared to be heterogeneity between the studies (Figure 5).

223Other studies undertook univariable regression analysis only. The Styles 2000 study included 52 participants with T1D followed for up to 45 years. According to the study publication, they did not find a "statistically significant" difference between median creatinine values in progressors (median 89; range 63 to 432) and non‐progressors to PDR (median value 78; range 56 to 330; P = 0.15) (Styles 2000).

224In the only study on a mixed population of people with T1D and T2D, in univariable regression analysis and analysis of variance, the risk ratio for creatinine values of 133 mmol/L or higher (RR 1.04, 95% CI 0.88 to 1.24), and the mean difference between progressors and non‐progressors to PDR (mean difference 1.2, 95% CI ‐4.16 to 1.76; P = non‐significant) were not significant (Janghorbani 2000).

225Twelve studies evaluated the effect of DR severity at baseline on progression to PDR (Arfken 1998; Burgess 2015; Klein 1984; Lee 2017; Lee 2021; Lloyd 1995; McCarty 2003; Nielsen 1984; Porta 2001; Roy 2006; Silva 2015; WESDR: Klein 1989a, Klein 1989b, 1994b, 1998, 2008). It was not possible to conduct a meta‐analysis, however, due to their heterogeneity. Most graded DR status based on the modified Airlie House classification system (Diabetic Retinopathy Study Research Group 1991) (Arfken 1998; Klein 1984; Lee 2017; Lloyd 1995; Porta 2001; Roy 2006; WESDR). The Nielsen 1984 study devised a scheme with four grades of background retinopathy, based on presence and number of red dots, haemorrhages and/or hard exudates, cotton wool spots, maculopathy and/or vitreous haemorrhage, and four grades of PDR, classified based on location of new vessels and presence of fibrovascular membranes, recurrent vitreous haemorrhage, neovascular glaucoma and/or retinal detachment. The Lee 2021 study categorised the level of DR according to the International Clinical Diabetic Retinopathy Disease Severity Scale (Wong 2018).

226Six studies undertook multivariable regression (Arfken 1998; Lee 2017; Lee 2021; Lloyd 1995; Porta 2001; WESDR: Klein 1998, 2008) (Table 25). All found DR severity at baseline to be an independent predictor of PDR, with higher DR severity increasing the risk of PDR. We assessed the certainty of the evidence as moderate, suggesting DR severity at baseline is likely to be associated with risk of progression to PDR (Table 24).

227Some studies undertook univariable analyses. In the WESDR study on T1D, in participants with moderate NPDR or worse in at least one eye at baseline, the risk of developing PDR was 42.4%, 63.6%, and 67.7% at four, 10 and 14 years, respectively (WESDR: Klein 1989b, 1994b, 1998). In the Roy 2006 study, in participants with moderate NPDR, 54.2% progressed to PDR during the six‐year follow‐up period. Porta 2001 reported that 17%, 40%, and 79% of participants with minimal, moderate, and severe NPDR, respectively, developed PDR during the seven‐year study duration. However, except for the Lloyd 1995 study, which accounted for HbA1c level, no other risk factors were included in the models (Porta 2001; Roy 2006; WESDR: Klein 1989b, 1994b, 1998)).

228In T2D, the WESDR study found that, after four years of follow‐up, no participants in the older‐onset group taking insulin and with DR severity of less than level 21/21 (microaneurysms only or retinal haemorrhages or soft exudates in absence of microaneurysms) developed PDR. With increasing DR severity, progression to PDR generally increased. In the older‐onset group not taking insulin, progression to PDR increased significantly in participants with advancing levels of severity from 31/31 (microaneurysms and one or more of the following: venous loops 31 µm or greater; questionable soft exudate, intraretinal microvascular abnormalities (IRMA) or venous beading; and retinal haemorrhages) onwards (WESDR: Klein 1989a). At ten years' follow‐up, the trend in progression to PDR with increasing DR severity at baseline was "statistically significant" in both groups (using and not using insulin) (P < 0.001), according to the publication. In those with moderate NPDR or worse in at least one eye at baseline, 61.8% in the group diagnosed with DM at 30 years or older taking insulin and 50% in the group not taking insulin developed PDR (WESDR: Klein 1994b). The primary prognostic factor of interest in the Silva 2015 study, including 100 participants with T1D or T2D followed for four years, was the effect of predominantly peripheral lesions on the development of PDR, but an increase in the proportion of eyes developing PDR was observed with advancing DR severity at baseline. The percentages of participants developing PDR with mild, moderate, severe, and very severe NPDR at baseline were 6.1%, 13.3%, 36.4%, and 100%, respectively (Silva 2015).

229Four studies assessed the influence of the presence of individual features of DR on the subsequent progression to PDR: two studies in people with T1D (WESDR; Verdaguer 2009), and two studies with a mixed population of people with T1D and T2D (Lee 2017; WESDR: Klein 1995c). Meta‐analysis was not possible due to their heterogeneity.

230Only two studies undertook multivariable regression analyses (Lee 2017; WESDR: Klein 1995c) (Table 26). To determine the effect of DR features on the incidence of PDR, Lee 2017 conducted a sub‐analysis of eyes with severe NPDR (n = 2823). The sub‐analyses included a total of 715 eyes, 240 eyes, and 169 eyes with IRMA, venous beading, and dot/blot haemorrhages in four quadrants, respectively. Lee 2017 established that the percentages of progressors to PDR by one, three, and five years were elevated in participants with IRMA (10.5%, 31.7%, 49.0%, respectively), followed by dot/blot haemorrhages (5.9%, 34.7%, 40.8%) in four quadrants and venous beading (5.0%, 17.2%, 39.9%). In multivariable Cox regression analysis, IRMA, but not dot/blot haemorrhages in four quadrants, was "statistically significantly" associated with increased risk of developing PDR compared to those with venous beading in two quadrants, according to the publication (Lee 2017) (Table 24).

231A report by the WESDR study in a mixed population of people with T1D or T2D found that the difference and ratio in the number of microaneurysms in the worst affected eye between baseline and the four‐year follow‐up were "statistically significantly" associated with incidence of PDR, as quoted in the report (WESDR: Klein 1995c). For an increase of one retinal microaneurysm at the four‐year follow‐up, there was a 4% increased risk of developing PDR (WESDR: Klein 1995).

232We rated the certainty of the evidence on the effect of DR features on development of PDR based on multivariable regression analyses as very low due to risk of bias, inconsistency, and imprecision (Table 24).

233The remaining two studies undertook univariable analyses only (Klein 1984; Verdaguer 2009). The Klein 1984 study (in people with T1D, follow‐up of six years), including an undetermined number of participants with moderate NPDR at baseline, evaluated the effect of hard exudates, cotton wool spots, IRMA, venous beading, and haemorrhages/microaneurysms (equalling or exceeding those in Standard Photo #3 as depicted by the DR Study Research Group; Diabetic Retinopathy Study Research Group 1981b) on the subsequent risk of PDR. The proportion of participants who progressed was identical in those with or without hard exudates (present 8/18; absent 11/25, 44%), and marginally increased in those with cotton wool spots (present 17/34, 50%; absent 2/9, 22%) and IRMA (present 11/18, 61%; absent 8/25, 32%) when compared with those without these features. There were too few with venous beading (n = 3) and haemorrhages/microaneurysms (n = 1) for an evaluation of risk in relation to these features (Klein 1984). The Verdaguer 2009 study (in 39 participants with T1D, follow‐up of 18 years) found in univariable analysis that hard exudates conferred a protective effect on the risk of PDR (OR 0.13, 95% CI 0.02 to 0.71; P = 0.011) (Verdaguer 2009).

234The WESDR study examined the correlation between microaneurysms and progression to PDR, in 236 participants who had only microaneurysms at baseline, over four (WESDR: Klein 1989d) and 10 (WESDR: Klein 1995c) years. At four years, the ratio of total number of retinal microaneurysms at follow‐up divided by the number present at baseline was associated with the development of PDR (P < 0.001). All eyes that developed PDR had a ratio of three or more (WESDR: Klein 1989d). Similarly, at 10 years, progression to PDR was more common in eyes with three or more microaneurysms at baseline (P < 0.05) (WESDR: Klein 1995c).

235The Valone 1981 study included 136 participants with T2D (136 eyes) with NPDR in one eye and PDR in the fellow eye, and followed them up for at least three months (average follow‐up 34.5 months, 95% CI 29.4 to 39.6 months). They found that PDR developed in 58% (n = 73) of participants after an average follow‐up of 23.9 months (95% CI 18.8 to 29.0 months). A "statistically significantly" higher percentage of participants younger than 40 years of age (67%) compared to those older than 60 years (36%) developed PDR, according to the publication, but the analysis did not correct for other risk factors (Valone 1981).

236In the Vesteinsdottir 2010 study, in a mixed population of people with T1D and T2D, 28 out of 48 (58%) developed PDR in the second eye within five years of its diagnosis in the first eye. Kaplan‐Meier analysis revealed that participants with T1D were at higher risk of developing PDR simultaneously in both eyes and within a short period when compared to those with T2D, but univariable analysis only was undertaken (Vesteinsdottir 2010).

237Only two studies evaluated the effect of retinal vessel caliber on progression to PDR in participants with T1D (Roy 2006; WESDR: Klein 2004, 2007) (Table 27).

238Roy 2006 found that increased central retinal vein equivalent (CRVE), defined as the average diameter of retinal venules measured at close proximity to the optic nerve head (but not central retinal artery equivalent, CRAE), was an independent predictor of the six‐year progression to PDR (P = 0.03) in univariable and multivariable models adjusted for HbA1c, age, sex, socioeconomic status, BMI, proteinuria, CRAE, ocular perfusion pressure, and refractive error (P = 0.03), with wider vein caliber found to increase the risk of PDR development (Roy 2006). DR severity at baseline was not included in these models.

239In the WESDR study, larger CRVE and smaller arteriolar‐venular ratio (AVR), but not CRAE, were statistically significantly correlated with greater four‐, 10‐, and 14‐year incidence of PDR (Klein 2004). In multivariable analysis (n = 871), larger venular diameters, but not arteriolar diameters, were associated with increased four‐year incidence of PDR (RR 4.28, 95% CI 1.50 to 12.19; P = 0.006), when controlling for sex, duration of diabetes, HbA1c, mean arterial blood pressure, antihypertension medication use, and DR severity at baseline. PDR was four times more likely to develop at four years in participants in whom the CRVE was in the fourth quartile range at baseline compared with participants in the first quartile range (WESDR: Klein 2004).

240In the T2D population, the WESDR study found larger CRVE ‐ defined as the CRVE in eyes in the fourth quartile compared with those in all other quartiles ‐ was associated with an increased 10‐year cumulative incidence of progression to PDR (P < 0.004). However, this relationship did not remain statistically significant in multivariate analyses (n = 889) controlling for age, HbA1c, and DR severity at baseline (P = 0.57). CRAE had no influence on the 10‐year development of PDR in univariate (P = 0.11) or multivariate (P = 0.78) analyses (WESDR: Klein 2007).

241We rated the overall certainty of the evidence for the effect of retinal vessel caliber as low, but the evidence suggests larger central retinal venular diameter may be associated with increased risk of progression to PDR in T1D (Table 24).

242Only two studies evaluated the potential effect of IOP on incidence of PDR (Valone 1981; WESDR: Moss 1994), and both undertook multivariable regression analyses (Table 28). IOP was not found to be an independent predictor of incidence of PDR in either study. However, we rated the evidence as very uncertain due to risk of bias, inconsistency, and imprecision (Table 24).

243Valone 1981 also compared the IOP in 44 participants with T2D evaluated over three years, and found no significant difference in the percentage of participants with higher IOP who remained with NPDR (53.8%) compared to those who progressed to PDR (61.3%) during the study (P > 0.05).

244Fourteen studies evaluated the effect of BMI on progression to PDR: four in people with T1D (Grauslund 2009; Porta 2001; Styles 2000; WESDR: Klein 2008); nine studies in people with T2D (Gui 2013; Kim 1998; Kim 2014; Lee 1992; Lee 2021; Nelson 1989; Okudaira 2000; WESDR: Klein 1997; Zavrelova 2011); and two in a mixed population of participants with T1D or T2D (Janghorbani 2000; Keen 2001).

245Six studies undertook multivariable regression (Grauslund 2009; Keen 2001; Kim 1998; Lee 2021; Nelson 1989; WESDR: Klein 1997, 2008) (Table 30). Only the WESDR study, reporting outcomes at 25 years in participants with T1D, found BMI (per increase of 4 kg/m2) to be an independent predictor for the development of PDR, but DR severity at baseline was not accounted for in the model (WESDR: Klein 2008). BMI was not found to be an independent predictor of PDR in the other studies. However, we rated the evidence based on multivariable regression analyses to be very uncertain due to risk of bias, inconsistency, and imprecision (Table 29).

246It was only possible to undertake meta‐analysis of unadjusted effect estimates combining three studies (Kim 2014; Lee 2021; Okudaira 2000). The pooled HR was 0.99 (95% CI 0.96 to 1.03), which was consistent with the findings from multivariable regression analyses which also did not find BMI to be predictive of PDR. The prediction interval ranged from 0.78 to 1.26 (Figure 4).

247All other studies, which undertook univariable analyses only, did not find BMI to be statistically significantly associated with progression to PDR (Gui 2013; Janghorbani 2000; Kim 2014; Lee 1992; Okudaira 2000; Porta 2001; Styles 2000; Zavrelova 2011).

248Thirteen studies evaluated the relationship between smoking and development of PDR: four studies in people with T1D (Grauslund 2009; Porta 2001; Styles 2000; WESDR: Moss 1991, 1996, Reports XVII, XXII); nine studies in people with T2D (Gange 2021; Gui 2013; Kim 2014; Lee 1992; Nelson 1989; Okudaira 2000; WESDR: Moss 1991, 1996; Zavrelova 2011); and two studies in a mixed population of people with T1D or T2D (Janghorbani 2000; Keen 2001). Meta‐analysis was not possible due to study heterogeneity.

249Six studies undertook multivariable regression (Gange 2021; Grauslund 2009; Gui 2013; Keen 2001; Nelson 1989; WESDR: Moss 1991, 1996, Report XVII) (Table 31). In all but one (Gui 2013), smoking was not found to be an independent predictor of development of PDR. Two studies including higher numbers of participants found smoking to have a protective effect in preventing PDR but neither included DR severity at baseline in their models (Gange 2021; Keen 2001). However, we rated the evidence based on multivariable regression analyses to be very uncertain due to risk of bias, inconsistency, and imprecision (Table 29).

250The remaining studies undertook only univariable analyses. In people with T1D, some studies compared the smoking status in “progressors” and “non progressors” to PDR but did not find a significant difference (Porta 2001; Styles 2000). In a mixed population (n = 3468) of people with T1D and T2D with follow‐up of 5 years, Janghorbani 2000 found a negative association between current smoking and incidence of PDR (RR 0.68, 95%CI 0.55 to 0.84; P < 0.001), but a positive one between ex‐smokers and PDR (RR 1.36, 95% CI 1.14 to 1.62; P < 0.001) in univariable analysis. However, according to the publication, there was no "statistically significant" mean difference in the percentage of participants progressing to PDR in the non‐, ex‐, and current smoker groups (Janghorbani 2000).

251Only two studies considered the influence of alcohol consumption on the development of PDR (Styles 2000; WESDR: Moss 1994a). Of these, only the WESDR study undertook multivariable regression analysis. The study found that alcohol consumption was not an independent predictor for progression to PDR in the total population, or in male and female subgroups in either the younger‐onset or older‐onset groups (WESDR: Moss 1994a) (Table 32) (Table 29).

252In the small Styles 2000 study (n = 52) in people with T1D followed for 45 years, there was no "statistically significant" difference between units of alcohol consumed per week by progressors to PDR compared to non‐progressors, as quoted in the publication (P = 0.31).

253Three studies explored the influence of a medical history of cardiovascular disease on progression to PDR (Jeng 2016; Lee 2021; Porta 2001). None found it to be a prognostic factor for PDR. Two of these studies undertook multivariable regression analyses (Jeng 2016; Lee 2021), but meta‐analysis was not possible due to study heterogeneity (Jeng 2016; Lee 2021; Porta 2001).

254A selection of other, less commonly evaluated, prognostic factors were investigated in individual studies, including: physical activity (WESDR: Cruickshanks 1995); myopia and ocular perfusion pressure (WESDR: Moss 1994a); alanine aminotransferase, haemoglobin, white blood cells, and platelets (Lee 2021); peripheral circulatory disorders (Gange 2021); peripheral lesions (Silva 2015); absence of Achilles tendon reflexes (Nelson 1989); cataracts (Verdaguer 2009); prothrombin time (Gui 2013); and oscillatory potential (Simonsen 1980).

255The WESDR study found no association between physical activity and development of PDR, even for those with more severe DR at baseline (WESDR:Cruickshanks 1995).

256In the report on ocular factors and progression of PDR in the WESDR study, in multiple logistic regression analyses controlling for HbA1c, DR severity, and age at baseline in the younger‐onset group, myopia (≤ 2.00 dioptres) was found to be protective (OR 0.40, 95% CI 0.18 to 0.86), but not in the older‐onset group when controlling for HbA1c, duration of DM, and DR severity at baseline (OR 0.40, 95% CI 0.04 to 4.16) (WESDR: Moss 1994b). The study also investigated ocular perfusion pressure (mmHg), which was calculated from IOP, and blood pressure; these were not found to be significantly associated with progression to PDR in the younger‐onset (OR 1.21, 95% CI 0.76 to 1.94) or older‐onset, insulin‐taking (OR 1.04, 95% CI 0.64 to 1.71) groups (WESDR: Moss 1994b).

257In a Cox proportional hazard model, adjusting for HbA1c, DR severity at baseline, gender, age, and BMI, Lee 2021 assessed the effect of alanine aminotransferase (U/L), haemoglobin (g/dL), white blood cells (103/µL), and platelets (103/µL) in 2623 participants with T2D at six years of follow‐up. Only haemoglobin was found to be an independent predictor of PDR (HR 0.84, 95% CI 0.74 to 0.96; P = 0.008) (Lee 2021).

258In multivariable regression analysis, including HbA1c, age, sex, ethnicity, education level, income, smoking, hypertension, dyslipidaemia, renal and neurological disease, morbid obesity, and insulin use as covariates, Gange 2021 found that peripheral circulatory disorders were associated with development of PDR (OR 1.88, 95% CI 1.25 to 2.83; P = 0.003) whereas diabetic ketoacidosis was not (OR not reported, P > 0.05) (Gange 2021).

259Silva 2015 found that peripheral lesions (haemorrhages, venous beading, IRMA, and NVE) identified on ultrawide field imaging were not "statistically significantly" associated with development of PDR, after adjusting for the previous two years’ HbA1c levels, DR severity at baseline, diabetes duration, and diabetes type in 109 participants with T1D or T2D followed for approximately four years (Silva 2015).

260In the Nelson 1989 study on Pima Indians with T2D aged 35 years or older, the absence of Achilles tendon reflexes was associated with PDR (incidence‐rate ratio 4.4, 95% CI 1.3 to 14.9) in a Cox’s proportional hazard model controlling for age, sex, and diabetes duration (Nelson 1989).

261In a small but long‐term study of 39 participants followed for 18 years, Verdaguer 2009 compared the presence of cataracts in groups with NPDR and PDR but did not find a significant difference (P = 0.117).

262Gui 2013 found that the mean difference in prothrombin time between non‐progressors (10.6, SE 2.11) and progressors (18.4, SE 3.05) to PDR was not significant (P = 0.21) in 190 participants with T2D followed for approximately two years.

263The Simonsen 1980 study explored the value of oscillatory potentials in detecting participants with T1D at risk of developing PDR. A total of 137 participants were followed at six to eight and 13 to 15 years. In all participants with PDR, oscillations were significantly reduced or extinguished (despite normal latencies of a‐ and b‐waves).

264Only three studies ‐ Roy 2006; Varma 2010; WESDR ‐ evaluated prognostic factors associated with progression specifically to HRC‐PDR, as defined by the DR Study Research (DRS) Group (and in our protocol) by the presence of NVD more than one‐fourth to one‐third disc area in size, or NVD/NVE of any size associated with vitreous or pre‐retinal haemorrhages (Diabetic Retinopathy Study Research Group 1991). None of these studies, however, looked at the specific risk of progression from PDR to HRC‐PDR.

265The WESDR study reported the effect of DR severity in T1D on development of HRC‐PDR at four, 10 and 14 years in participants with NPDR at baseline (WESDR: Klein 1989b, 1994b, 1998). Varma 2010 investigated the influence of age at baseline and duration of PDR on progression to HRC‐PDR, in participants with T1D and T2D (mixed cohort) over a four‐year period, but all had NPDR at baseline. The Roy 2006 study assessed the influence of gender and retinal vessel caliber on progression to HRC‐PDR, but some participants had no DR or NPDR at baseline.

266There were insufficient data to explore the impact of studies at high risk of bias or retrospective studies on the effect sizes observed, as we had originally planned. For the same reason, we were unable to explore sources of heterogeneity between studies.

267This systematic review found HbA1c and DR severity at baseline to be independent predictors for the development of PDR in people with T1D or T2D, with higher levels of HbA1c and retinopathy increasing the risk of PDR. Included studies used different biomarkers indicative of renal disease (nephropathy, proteinuria, albumin excretion rate, ACR, eGFR, creatinine), with most pointing towards a possible increased risk of progression to PDR in people with impaired kidney function. Age at diagnosis of diabetes, elevated triglyceride levels, and larger retinal venular diameters may also possibly be associated with progression to PDR in people with T1D. There was no clear evidence that duration of diabetes had an influence on the development of PDR when HbA1c and DR severity were included in the models. Neither do SBP, DBP, total cholesterol, LDL or HDL levels, nor gender, ethnicity, BMI, socioeconomic status, smoking, and alcohol consumption appear to be associated with progression to PDR.

268Despite 59 studies (87 reports) being included in the review, the great heterogeneity in study design, prognostic factors evaluated, how they were measured, the lack of adjustment for potentially important risk factors in some studies and their consideration in statistical models, as well as weaknesses in the quality of reporting, significantly limited our ability to undertake meta‐analysis and even interpret with certainty their results (Figure 2).

269HbA1c was identified as an independent predictor of PDR in almost all studies in which multivariable regression analyses were undertaken, with the exception of Cho 2019 and Lee 2021, both retrospective cohort studies including Asian populations with T2D (Cho 2019: n = 1527; Lee 2021: n = 2623). Besides differences in ethnicity when compared with the other studies, HbA1c levels at baseline in these studies appeared to be lower, which may at least partly explain these discrepant results (Cho 2019; Lee 2021).

270The finding that elevated HbA1c contributes to the development of PDR, even when taking into consideration DR severity, challenges the hypothesis of 'retinopathic momentum', which suggests that once DR progresses far enough, no intervention will halt its relentless progression (Diabetes Control and Complications Group 1993). A review by Liu and colleagues similarly proposed that intensive glycaemic control may not be beneficial if DR severity is worse than moderate NPDR (Liu 2020). The fact that HbA1c levels are a risk factor for PDR is also supported by interventional RCTs, which found that a more rigorous glycaemic control has beneficial effects in reducing risk of progression of DR, albeit this being more pronounced in people with T1D when compared to those with T2D (ACCORD 2011; Diabetes Control and Complications Group 1993; Turner 1998). Furthermore, results of a meta‐analysis of 16 interventional RCTs concluded that intensive glycaemic control significantly delayed progression to PDR or requirement of laser treatment for PDR in people with T1D (OR 0.44, 95% CI 0.22 to 0.87; P = 0.018) (Wang 1993). It is important to note that rapid changes in glycaemia should be avoided, as these may lead to acceleration of the progression of DR, including development of PDR (Diabetes Control and Complications Group 1998). This aligns with the findings of a recent meta‐analysis of RCTs that the use of newer glucagon‐like peptide‐1 receptor agonists (GLP 1 RA) (e.g. liraglutide, semaglutide and dulaglutide) was associated with an increased risk of rapidly worsening of DR (Yoshida 2022).

271DR severity at baseline was also found to be a consistent and significant independent predictor of PDR development. In this regard, the ETDRS found that severity of DR features at baseline was the most important factor predicting progression of DR (ETDRS 1991b). Interestingly, when DR severity was included in multivariable regression models in many studies included in our review, the effect of other risk factors, initially potentially linked to PDR, was no longer observed. This suggests that the retina portrays the effects of the systemic environment and, as a result, changes in the retina have more prognostic value than that of the risk factors themselves.

272Renal impairment, determined by different means in the various studies included, was generally found to be an independent predictor of PDR. Thus, three out of four studies evaluating nephropathy using multivariable regression found this diabetic complication increased the risk of progression to PDR. Similarly, five of the eight studies evaluating proteinuria, two of the three studies using “albumin excretion rate” or “albumin creatinine ratio” or “creatinine”, and the two studies that used “eGFR” found, in multivariable regression models, that deranged kidney function was associated with increased risk of development of PDR. Heterogeneity in the populations studied and in the various definitions used for renal impairment may at least partly explain discrepancies in the results observed amongst different studies. Given that similar histopathological alterations are present in retina and kidney in people with diabetes ‐ including basement membrane thickening, endothelial cell dysfunction/loss, and loss of pericytes in the retina and their kidney counterpart (podocytes), amongst others ‐ it is not surprising that disease in both organs may be interrelated (Wong 2014). Indeed, the overall prevalence of PDR in a cohort of 15,409 participants over 19 years of age with T1D or T2D was found to be 5.5% if chronic kidney disease was present compared to 1.8% if it was not (Park 2015). A study of 1214 participants with T2D found a profound difference in the prevalence of PDR in those on dialysis (31.7%) compared to those not requiring it (1.9%) (P < 0.001) (Boelter 2016).

273Evidence on the relationship between triglycerides and PDR was relatively limited, but suggested that increased triglyceride levels in T1D, but not in T2D, may be associated with development of PDR. Other studies not included in this review have suggested a possible association between triglyceride levels and PDR. Thus, in a mixed population of 2651 participants with T1D or T2D, the ETDRS found that increased triglyceride levels were associated with progression to HRC‐PDR in multivariable models controlling for HbA1c, type and duration of DM, DR severity at baseline, age, gender, ethnicity, weight, visual acuity and presence of diabetic macular oedema (DMO) (Davis 1998). Moreover, in an Irish cohort study including 2770 participants with T2D, triglycerides were positively correlated with increased risk of referable DR (HR 1.10, 95% CI 1.03 to 1.18; P = 0.004) (Smith 2020). It should be noted, though, that 'referable DR' includes PDR but also higher stages of NPDR and DMO. However, the Diabetes Control and Complications Trial (DCCT), which included 1441 participants with T1D, did not find triglyceride levels to be an independent predictor of progression to PDR in multivariable regression controlling for HbA1c, duration of DM, DR severity at baseline, randomised treatment, age, sex, and smoking (Miljanovic 2004). Unlike cholesterol, levels of triglycerides are greatly influenced by whether or not blood samples obtained for their analysis are gathered following fasting. Thus, a fasting blood sample is required to ensure accuracy of results; these were obtained in some of the studies included in this review in which the potential relationship between triglyceride levels and PDR was tested, but not reported in others.

274It is possible that age at diagnosis of DM may also determine risk of development of PDR. Although this conclusion is drawn from results of only two studies in which multivariable regression models were used, in one of these by Porta and colleagues, the effect of younger age at diagnosis was observed even when DR severity at baseline was included in the model, suggesting it is indeed an important risk factor, at least in people with T1D (Porta 2001).

275Similarly, in people with T1D, larger retinal venular diameter was also found to be associated with an increased risk of PDR in two studies (Roy 2006; WESDR: Klein 2004). In one of these, the effect remained after controlling in a multivariable regression model for DR severity at baseline (WESDR: Klein 2004). This finding is in accordance with studies evaluating pathogenic mechanisms of disease in DR. These showed that the arteriolar vasoconstriction observed in early DR stages with subsequent blood flow reduction is followed by vasodilation and enhanced blood flow which could then hasten the development of PDR (Stitt 2016).

276Duration of diabetes in people with T1D was only found to be an independent predictor of the development of PDR when DR severity at baseline was not included as a covariate in multivariable regression models (Grauslund 2009; Janghorbani 2000; Kalter‐Leibovici 1991; Lloyd 1995; Porta 2001). In T2D, results were inconsistent and none of the studies using multivariable regression models included DR severity at baseline. It may be challenging to determine the exact date of diagnosis of DM in people with T2D, and, thus, its relationship to PDR. In this regard, studies have suggested that glucose dysregulation can precede the diagnosis of T2D by up to 20 years (Sagesaka 2018). Similarly, in nearly all studies evaluating DBP, SBP, and total cholesterol, these were only independently associated with progression to PDR when DR severity at baseline was not included in multivariate regression models. This suggests that the influence of these factors may not be evident beyond their effect already imprinted in the retina.

277Our findings with regard to the effect of blood pressure (BP) supported those of a Cochrane Review (Do 2015). It included 15 interventional RCTs involving participants with T1D and T2D, and did not find a beneficial effect in reducing blood pressure to prevent progression to PDR (estimated RRs 0.95, 95% CIs 0.83 to 1.09), although there was a benefit in delaying the incidence of DR (Do 2015). In the UK Prospective Diabetes Study (UKPDS), the beneficial effects of BP control were only seen when baseline BP was very high (160/94 mmHg) (UK Prospective Diabetes Study Group 1998). However, it is possible that in a real‐world clinical setting, individuals’ control of BP may be as poor as it was in the UKPDS. Under these circumstances, there would be an overall benefit of reducing levels of BP in order to reduce the risk of development of other diabetic complications.

278Aside from the possible effect of triglycerides in people with T1D, as reviewed above, dyslipidaemia ‐ although evaluated extensively in relation to incidence and progression of DR ‐ was only rarely studied with regard to its relationship with the incidence of PDR, limiting available data.

279In studies undertaking multivariable regression, BMI was not associated with progression to PDR, with the exception of the 25‐year prospective cohort study by Klein and colleagues (WESDR: Klein 2008), which found BMI (per 4 kg/m2) to be an independent predictor of the incidence of PDR in people with T1D. However, DR severity at baseline was not included in the model. Discordance in the relationship between obesity and PDR may be due to BMI being an inaccurate indicator of central fat distribution, which has been found to be associated with more severe stages of DR in those with T2D (Man 2016; Raman 2010). Indeed, in Parente and colleagues' 15‐year observational cohort study, waist‐to‐height ratio and waist circumference were more significant indicators of progression to "serious diabetic eye disease" than BMI (Parente 2022).

280Our review did not find evidence for an effect of smoking on the risk of development of PDR. Although nicotine can cause vasoconstriction in normal circumstances, which would be expected to reduce blood flow through the retina, it is possible this effect may not alter the blood flow in the diabetic retina given that vasodilation of arteriolas is known to be part of the dysregulation present in the diabetic retina in more advanced stages of disease (Mills 2021; Schmetterer 1999).

281Strengths of this review include our comprehensive and systematic search of the literature for relevant studies, with no language or date restrictions, and our detailed scrutiny of all references listed in all included studies. We believe this rigorous search process means it is unlikely we missed relevant literature. We conducted the review according to Cochrane’s standards.

282The review's limitations relate to features of the included studies, such as the heterogeneity in study design, populations, follow‐up period, type of analyses undertaken, and adjustment (or lack thereof) for other potentially important risk factors. Thus, study heterogeneity greatly restricted our ability to undertake meta‐analysis. Furthermore, anomalies in the reporting of important baseline characteristics of study participants, such as the level of DR severity at baseline, ethnicity, and HbA1c and blood pressure levels, resulted in some uncertainty when interpreting outcomes of various studies in the context of the populations investigated. Moreover, some studies did not specify the direction of the effect of particular prognostic factors studied, making it impossible to interpret findings. Many studies undertook only univariable regression analyses when evaluating specific risk factors for progression to PDR, thus disregarding the effect of other potentially important factor. Based on our review, HbA1c and DR severity should always be corrected for in future PDR prognosis studies. Other potential limitations of the data presented relate to limitations inherent to the techniques used to diagnose PDR: most studies used standard fundus photography. Thus, it is very possible that cases of early PDR missed by standard fundus photography could have been identified by the use of fundus fluorescein angiography.

283Substantial resources are required to conduct prognosis studies, not only with regard to costs, but also the time and effort invested by participants and researchers. Thus, it is crucial that future studies are rigorously planned, and appropriately statistically analysed and reported, to ensure maximal benefit to people with diabetic retinopathy is realised. Guidelines provided by the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) initiative (Patzer 2021; www.tripod‐statement.org/scope), as well as those provided by the Quality In Prognosis Studies (QUIPS) tool (developed to assess risk of bias in prognosis studies), could help researchers to design more robust prognosis studies.

284Increased HbA1c is likely to be associated with progression to PDR, and therefore, maintaining adequate glucose control throughout life, irrespective of stage of DR severity, may help to prevent progression to PDR and risk of its sight‐threatening complications. Renal impairment in people with T1D or T2D, as well as younger age at diagnosis of DM, increased triglyceride levels, and retinal venular diameters in people with T1D may also be associated with increased risk of progression to PDR. Given that more advanced DR severity is associated with higher risk of progression to PDR, the earlier the disease is identified, and the above systemic risk factors are controlled, the greater the chance of reducing the risk of PDR and saving sight.

285There is evidence to support that elevated glycated haemoglobin (HbA1c) and more advanced diabetic retinopathy (DR) severity at baseline are independent risk factors for the development of proliferative diabetic retinopathy (PDR) in people with type 1 (T1D) or type 2 (T2D) diabetes. Evidence for other risk factors is less compelling, although it suggests that renal disease in people with T1D and T2D, and younger age at diagnosis of diabetes, higher triglyceride levels, and larger retinal venular diameters in people with T1D may also increase the risk of PDR.

286Despite the large number of cohort studies that have been undertaken over the years, the great heterogeneity in study design, prognostic factors evaluated and how they were measured, the lack of adjustment for potentially important risk factors in some studies and their consideration in statistical models, as well as weaknesses in the quality of reporting, significantly limited our ability to interpret their results.

287Evidence from this review suggests it is likely that maintaining adequate glucose control throughout life reduces the risk of developing PDR. People with T1D or T2D and renal disease may be at increased risk of developing PDR. Maintaining triglyceride levels within the normal range may also reduce the risk of progression to PDR in people with T1D. Research has shown that fenofibrate, which lowers triglycerides, has other additional beneficial effects in the retina independent of the lipid reduction (Chew 2003; Keech 2007; Stewart 2018). However, at present, this drug is not licenced in the UK for the treatment of DR.

288More robust research is necessary to adequately determine prognostic factors associated with PDR and, specifically, progression from PDR to high‐risk characteristics DPD (HRC‐PDR) in order to identify individuals at higher risk of developing sight‐threatening complications who may benefit from even earlier interventions.

289We identified HbA1c and DR severity at baseline as being the most significant risk factors for development of PDR. Thus, it is essential that future studies adjust for them in their prognostic models.

290The heterogeneity in study design, characteristics of populations included, follow‐up period, types of analyses undertaken, and adjustment (or lack thereof) for other potentially important prognostic factors greatly limited our ability to undertake meta‐analysis. Researchers should be mindful of this. Establishing a core outcome set for prognosis studies in the field of DR, as advocated by the Core Outcome Measures in Effectiveness Trials (COMET) initiative (Mokkink 2016), and homogenising the instruments to measure these outcomes, as proposed by the COnsensus‐based Standards for the selection of health Measurement Instruments (COSMIN) initiative (www.cosmin.nl), would greatly facilitate synthesis of data to derive more meaningful conclusions. We encourage these developments.

291Frequently in this review, prognosis studies were not eligible for inclusion because study authors reported only a combined outcome of development of PDR or diabetic macular oedema (DMO; what is called 'sight‐threatening' DR), or generalised progression of two to three steps on the Early Treatment Diabetic Retinopathy Study (ETDRS) scale. Given that the risk of blindness conferred by each of these conditions is very different and that indeed total blindness (central and peripheral) still occurs in present times only as a result of complications of PDR, it is essential clinicians and researchers look into each of these outcomes separately in future studies.

292Lastly, establishing a database of prognosis studies in diabetes and DR would facilitate the identification of such studies in medical database searches and would thus aid and advance research efforts in this field.

293Protocol first published: Issue 11, 2020