Diabetic retinopathy, a progressive eye complication of diabetes mellitus, which affects 9.3% of the people globally, is considered the fifth leading cause of global blindness [1,2]. Results from a recent systematic review estimated that globally, the prevalence of retinopathy among patients with diabetes is 35% [3]. Nearly all patients with type 1 diabetes and more than 60% of the patients with type 2 diabetes develop retinopathy within 20 years of diabetes onset [4]. However, like many other diabetic complications, retinopathy remains asymptomatic until significant damage has occurred [4,5]. Periodic eye screening is essential for diagnosing the disease in early stages and to enable timely initiation of treatment [6]. Current recommendations are that retinal screening occur once a year for all patients with diabetes and that more frequent examinations take place if abnormal findings are identified [7]. However, periodic retinal screening is not available in many low- and middle-income countries although they account for 75% of the global burden of diabetic retinopathy [3,8]. The reported barriers are multifactorial. A recent systematic review reported that poor knowledge and attitudes to asymptomatic eye screening are prevalent both among health care providers and patients [9]. At the health care system level, the lack of equipment, insufficiently skilled professionals, nonfunctioning referral mechanisms, and inadequate data within national management information systems are the main barriers [10]. Moreover, routine diabetic retinopathy screening is not always feasible for eye consultants, given their availability versus the load of patients with diabetes.

As an alternative to eye consultants, different cadres of non-ophthalmologists such as general practitioners, opticians, and diabetologists have been successfully involved in diabetic retinopathy screening in many high-income countries [11]. Particular attention has been focused on developing simple algorithms and technologies suitable for non-ophthalmologists. Among these screening tools, digital fundus photography has been identified as one of the best and lowest cost options [12]. A digital camera allows repeated images to be taken until the best one is captured. Final retinal images can be stored and sent for expert opinion by using a web-based interface [13]. With this technology, high-income countries have shown increased diagnostic accuracy even without pupil dilatation (mydriasis) [12,14]. However, the nonmydriatic approach has shown low accuracy (12%-25%) in the Southeast Asian context [6]. Patients having a dark iris and reporting to a hospital at an advanced age with comorbid eye diseases such as cataracts are the commonly reported explanations for poor vision in the nonmydriatic approach [6].

Like many low- and middle-income countries, Bangladesh demonstrates a substantial diabetic retinopathy disease burden and an array of health system challenges that complicate the routine implementation of diabetic retinopathy screening services. According to the International Diabetic Federation statistics, around 8.4% of the total population in Bangladesh had diabetes in 2017 [15], which puts the country among the top 10 high diabetes burden countries in the world [15]. At the same time, Bangladesh has a critical shortage and “geographic maldistribution” of health workforce with an increased concentration in urban areas even though 70% of the population resides in the rural region [16]. Eye care services are provided predominantly by ophthalmologists, and diabetic retinopathy screening programs are not readily available under the current health system [17]. Given the growing burden of diabetes in Bangladesh, blindness prevention has become a strategic priority [17]. To support this effort, the “Integrated Model of Care for Diabetic Retinopathy within the Health System of Bangladesh” was initiated as a collaborative program between the Fred Hollows Foundation (FHF), a nongovernmental organization and the Government of Bangladesh [18]. This program seeks to establish a care pathway for diabetes and diabetic retinopathy by training non-ophthalmologist health cadres to conduct diabetic retinopathy screening by using digital fundus photography [18].

The aim of this study was to test the diagnostic accuracy of detecting any grade of diabetic retinopathy by non-ophthalmologists using digital fundus cameras against the gold standard diagnosis of ophthalmologists. We also explored the risk factors of diabetic retinopathy in the rural sites under investigation.

Conducted in 6 peripheral health facilities in Bangladesh, this validation study showed that non-ophthalmologists can detect the presence of diabetic retinopathy with reasonable diagnostic accuracy. The sensitivity and specificity of using a single-field macular photograph after pupillary dilatation against the referral gold standard diagnosis by an eye consultant were 86.6% (483/558) and 78.6% (705/897), respectively. The degree of diagnostic agreement between a non-ophthalmologist and an ophthalmologist was excellent with a desk-based camera and satisfactory with a hand-held camera.

To our knowledge, this is the first ever study in Bangladesh to engage health care personnel other than eye consultants in routine retinal examination. The reported true positive diabetic retinopathy case detection rate was within the international standard of more than 80% [21,22]. However, our overall specificity (79%) was lower than the international recommendation of 90%-95% [21,22]. This implies that 19.9% (173/869) of the total diabetic retinopathy negative cases were referred to an eye consultant when they did not have diabetic retinopathy [26]. In generic terms, lower specificity implies greater health system burden with a greater number of false positive cases referred to the next level [26]. Nevertheless, in a country like Bangladesh where no formal diabetic retinopathy screening services are available, a large proportion of people with diabetes remain undiagnosed until opportunistic diagnosis occurs at a very advanced stage [28]. Considering the increasing burden of diabetic retinopathy, this will result in increasing eye care–related costs and a greater risk of blindness [29]. By training and engaging non-ophthalmologist health staff in diabetic retinopathy screening, limited resources can be maximized and coverage increased [25]. However, evidence to support the incorporation of non-ophthalmologists into the diabetic retinopathy screening pathway remains scarce in low- and middle-income settings [30]. Some studies in India, Sri Lanka, and Pakistan report satisfactory test accuracy for non-ophthalmologist screening [6,9,31,32]. Using digital fundus photographs and a nonmydriatic approach, “physician graders” in Sri Lanka showed a sensitivity of 88.7% and specificity of 94.9% [10]. In Pakistan, optometrists showed 72% sensitivity and 86.3% specificity [32]. A decision must also be made about which instrument and imaging technique to use and whether to dilate or not dilate [33]. We validated the diagnostic agreement of hand-held versus desk-based cameras by calculating the AUROC and kappa values. The AUROC was “close to 1” for both instruments [26], suggesting a fair-to-excellent amount of agreement. However, in low- and middle-income country settings, instrument costs should also be considered. Here, the hand-held camera performs better—being comparatively less expensive than a desk-based camera and easier to carry and employ in the context of community-level screening [34].

The variation in test accuracy by instrument type further emphasizes the need to explore other factors influencing diagnostic test accuracy. Here, the gradeability of images plays an important part in successful diabetic retinopathy screening. In this study, a lower rate of ungradable images was reported (56/1511, 4.1%) than that reported in similar health facility–based studies in the South Asian context. A technical failure rate of 7.5% was noted in another study in Bangladesh [35], which rose to 12% in Sri Lanka [10]. The higher level of gradeability observed in our study is probably a function of the exclusion criteria or the image-taking technique that was employed. Considering the higher prevalence of cataracts in low-income country settings [29,36], we made the presence of cataract an exclusion criterion in our sample selection and chose to deploy a mydriatic approach for imaging. Although the gold standard for diabetic retinopathy screening is a 7-field photograph using a nonmydriatic approach [37], the applicability of this technique in low-income country settings is highly controversial [38]. When using a nonmydriatic approach, the body’s autonomic nervous system becomes hyperactive with reflex pupillary constriction in the second eye after taking an image of the first eye [39,40]. Further, a Brazilian study showed that longer screening time and more referrals to ophthalmologists occurred in the absence of pupil dilation due to a larger number of ungradable images [41]. Further supporting our approach is the evidence of increased provider compliance when using a single-field photograph [42].

To identify other factors affecting diabetic retinopathy diagnosis, we compared test accuracy across different cadres of non-ophthalmologist personnel. We observed that nurses and paramedics showed higher accuracy in detecting any form of diabetic retinopathy than the nonclinical trained staff [43]. The relatively poorer performance of the nonclinical trained staff may be due to their lack of comfort with the hand-held camera. We expect that with more onsite supportive training, performance can be increased. Supporting the call for more hands-on training, a study suggested that the diagnostic accuracy of the fundus camera is highly dependent on the user’s technique in taking a correct image and their ability to do proper grading—both of which can be improved with more hands-on support [44]. Variations in the test accuracy by type of provider were similarly observed in a systematic review of 22 observational studies [45]. Noting that the sensitivity of detecting any form of retinopathy using a mydriatic approach ranged between 87% and 100% for general practitioners, >91% for optometrists, and 89% and 93% for ophthalmologists or their assistants, the authors of a study concluded that using appropriate technology and ensuring quality are more important than the type of provider in diabetic retinopathy screening programs [45].

We also explored factors predicting retinopathy occurrence among patients with diabetes. Increased duration of diabetes and obesity were identified as significant predictors of diabetic retinopathy. The odds of having diabetic retinopathy was found to increase 2 times among patients with diabetes for 5-10 years and 4 times among those with diabetes for more than 10 years compared to the odds of having diabetic retinopathy in those with diabetes for less than 5 years. While the increased duration of diabetes for the development of diabetic retinopathy is well-established [3,8,46,47], the observed negative association between obesity and diabetic retinopathy in this study was inconsistent with that reported in the literature. Only 1 study conducted in an urban slum in India reported a similar result; those in the overweight category had lower odds of diabetic retinopathy (OR 0.6, 95% CI 0.4-0.9) than people with normal BMI [31]. In general, however, obesity in the Asian context has been identified as a risk factor for developing diabetic retinopathy [1].

Although results from this pilot study are supportive of mainstreaming non-ophthalmologist health staff in diabetic retinopathy screening services, some limitations in the diagnostic processes need to be considered before generalizing this recommendation to other contexts. One important limitation was our decision to exclude patients with cataract and to employ a mydriatic and single-field macular photographic approach due to the pilot nature of this research. In this regard, further clinical trials may be useful for determining the test accuracy by type of health professional, type of instrument used, use of pupil dilation, and number of fields chosen for screening. Finally, recommendations emerging from a study conducted in hospitals with specialized diabetes or eye departments may be less pertinent in nonspecialist hospitals where the prevalence of diabetic retinopathy is likely lower.

Considering the rising burden of diabetes in Bangladesh, routine retinal screening of patients with diabetes is not feasible by eye consultants alone. Our study results suggest that non-ophthalmologist staff such as nurses and paramedics could function as frontline health workers in diabetic retinopathy screening programs. As the first step, the engagement of non-ophthalmologist cadres in diabetic retinopathy screening should be limited to categorization based on the presence or absence of diabetic retinopathy only. Their involvement in the more technical area of diabetic retinopathy grading requires further specific training and health system level support. Regarding instrument choice, although the desk-based camera shows better accuracy in detecting diabetic retinopathy, the choice of instrument type should be a function of the capacity of the health facility and the health care provider performing the diagnosis.