Using artificial intelligence-assisted retinal imaging devices operated by non-specialist workers in detecting diabetic retinopathy in Sri Lanka Authors: Nilmini Wijemunige, Moditha Madhuwanthi Dissanayake, Tanya Vimarshani Gamage, Hasitha Rajakaruna, Pavithra Maniyangamage, Niroshani Wisidagama, Malinda De Silva, Manjula D Nugawela, Mapa Prabhath Piyasena, Sobha Sivaprasad, Sarath Samarage, Shanti Dalpatadu, Ravindra Prasan Rannan-Eliya Published: 2026 ABSTRACT AbstractBackgroundDiabetic retinopathy (DR), a microvascular complication of diabetes, is an important cause of preventable blindness and can cause a significant reduction in the quality of life of working-age adults. Sri Lanka has one of the highest prevalences of diabetes globally, and like many low- and middle-income countries, faces significant barriers to DR screening. Easy-to-use, relatively cheap handheld retinal fundus imaging devices with automatic interpretation using artificial intelligence (AI) presents a new avenue for DR screening by non-specialist healthcare workers. This study piloted the use of one such device by non-specialist healthcare workers in a low-resource setting and aimed to evaluate image gradability and the diagnostic test accuracy (DTA) of the AI interpretation in this context.MethodsAdults with and without diabetes were recruited from a research clinic. Non-specialist healthcare workers received brief training to capture macular- and optic-disc-centred retinal images using an AI-assisted handheld retinal imaging camera (Remidio FOP NM-10), before and after mydriasis. The device generated reports identifying referrable DR. Participants also had visual acuity assessed and underwent slit lamp biomicroscopy by an ophthalmologist. We evaluated image gradability of the captured images pre- and post-dilation, as well as the DTA of the AI reports using the ophthalmologist’s findings as the gold-standard.ResultsRetinal images were captured on both eyes of 119 participants, including 49 with diabetes. The proportion of participants that had ungradable images of at least one eye prior to mydriasis was 46.2%, and this fell to 15.1% after mydriasis. A higher percentage of older individuals (59% of those aged 55–64 years and 70% of those aged 65–74 years), 85% of those with cataracts and 58% of those with poor visual acuity had at least one eye with ungradable images pre-mydriasis, with these percentages reducing by 40–70% with mydriasis. The sensitivity and specificity for referral of patients compared to the gold-standard was 73.3% and 79.8% respectively, with image gradability a major driver of misclassification.ConclusionsThe AI-assisted retinal imaging camera used by non-specialist healthcare workers after minimal training, in a low-resource setting shows promise as a feasible tool to identify DR in dilated eyes.Supplementary InformationThe online version contains supplementary material available at 10.1186/s44263-026-00279-6.Keywords: Diabetic retinopathy, Diabetes, Screening, Sri Lanka, AI, Fundoscopy Background Globally, the prevalence of type 2 diabetes is high and rising across all regions. An estimated 589 million adults aged 20–79 years worldwide (11.1% of all adults in this age group) have diabetes [1]. Over the past decade, the burden of diabetes, measured by disability-adjusted life years, has risen faster in low- and middle-income countries (LMICs) than in high-income countries [2]. With the increasing prevalence of diabetes, there is a concurrent increase in complications of diabetes [3]. One such complication is diabetic retinopathy (DR), an important cause of preventable blindness, which, when treated early, can greatly improve quality of life [4]. Globally, 1.9% of moderate or severe visual impairment and 2.6% of blindness in 2010 were due to DR [5] with 22% of people with diabetes mellitus estimated to have DR [6]. Furthermore, DR is the largest cause of preventable blindness in working-age adults [7] leading to visual impairment, poor quality of life, and low levels of psychosocial well-being. Studies such as the Early Treatment Diabetic Retinopathy Study (ETDRS) [8] have shown promising results that early detection and treatment of DR, and effective control of diabetes can make blindness from DR almost entirely preventable. However, systematic screening for DR proves to be a significant challenge, particularly in LMICs, due to limited resources and infrastructure [9]. The Sri Lanka Health and Ageing Study (SLHAS) found that Sri Lanka has one of the highest prevalences of diabetes in the world, at 23% of adults aged 18 years and above [10]. A small number of studies, none nationally representative [11–13], suggest that the prevalence of DR in Sri Lankan people with diabetes is about 27%–53%, and the prevalence of sight-threatening diabetic retinopathy is about 4%. The local reference standard for DR screening is retinal screening with dilated fundoscopy [14]. In practice, people with diabetes receiving treatment in larger, tertiary-level public health facilities are often referred to eye clinics for screening at the time of diagnosis of diabetes. At the primary care level, people with diabetes are referred for DR screening to consultant eye surgeons or other primary screeners (physicians trained in ophthalmology). However, Sri Lanka has few ophthalmologists–only 5 per million population compared to and average of 11 in the eight South Asian Association for Regional Cooperation (SAARC) countries [15]. A survey conducted in the Western Province reported that, although there was adequate capacity to deliver eye care services and conduct a DR screening program, human resources, and infrastructure distribution were not properly aligned to the population’s need, resulting in the lack of a systematic DR screening program [16]. Another study conducted in the Western Province estimated that annual diabetic eye screening visits cover only 17% of people with diabetes and that only 11% of people with diabetes who are likely to need photocoagulation therapy receive it [17]. Barriers to accessing the existing DR screening services were multifactorial and included institutional factors, with overcrowded clinics, long waiting times, and poor organisation of care, resulting in poor patient experience [18]. The ability to screen for DR in a way convenient to patients can greatly improve the impact of screening programs [7]. Diabetic retinopathy screening could potentially be increased using retinal imaging devices with in-built automatic interpretation to detect diabetic retinopathy, particularly if they can be used effectively by non-specialist healthcare workers, the typical cadre at primary and secondary healthcare facilities. Early detection would facilitate healthcare workers to triage people for further examination and treatment by ophthalmologists. One such device is the Remidio Fundus on Phone Non Mydriatic-10 (Remidio FOP NM-10), which is a low-cost smartphone-enabled retinal imaging system that utilizes precision optics to illuminate and capture retinal images [19]. The integrated artificial intelligence (AI) system included in the smartphone allows the capturing of high-quality retinal images with a 45º Field of View, and automatic detection of diabetic retinopathy and glaucoma. This study was a pilot to inform the design of clinical protocols for a subsequent multi-centre study, DIAGNOSE-Diabetic Retinopathy (DIAGNOSE-DR). DIAGNOSE-DR aims to test the use of a digital retinal camera with AI interpretation in low-resource clinic settings, operated by non-specialist healthcare staff, in Sri Lanka. DIAGNOSE-DR also aims to assess the prevalence of DR in Sri Lanka. This pilot study was registered on Open Science Framework (10.17605/OSF.IO/EN5RU) [20]. The aim of the pilot study was to test one such digital retinal camera—the Remidio FOP NM-10 device—in a clinic setting, assessing its use by non-specialist workers with a healthcare background, and assessing the need for mydriasis in capturing images, thereby informing the design for the multi-centre study. The objectives of this paper are to (1) calculate the proportion of participants and eyes for which gradable images are captured pre- and post-dilation; (2) compare the gradability of retinal images captured pre- and post-mydriasis; (3) identify any associations between the gradability of images and characteristics of participants or image capturers; (4) assess the Diagnostic Test Accuracy (DTA) of the Remidio FOP NM-10 generated AI report for diabetic retinopathy after dilation compared to slit lamp biomicroscopy by an ophthalmologist; and (5) review discrepancies between the slit lamp biomicroscopy and AI report for all eyes where gradable images were captured. Methods The study was a non-randomized single-centre, cross-sectional feasibility study conducted at the Eye Care Centre, Faculty of Medicine, University of Colombo, in Sri Lanka. This clinic conducts community eye care services, education, and research projects under the management of its Director, a consultant ophthalmologist (MMD). Study procedure Participants Patients aged 18 years and above who are registered at the Eye Care Centre (ECC), Faculty of Medicine, University of Colombo, were recruited. Two groups of participants were recruited, with and without diabetes, with a target of at least 50 participants in each group. A participant was classified as a person with diabetes, if they had a previous diagnosis of type 1 or type 2 diabetes mellitus on medical records (typically a clinic book or medical notes brought by participants) and had been diagnosed with diabetes for at least three months. For the sample of people without diabetes, participants must report no known history of diabetes, nor have any diagnosis of diabetes on available medical records. The DIAGNOSE-DR project includes individuals without diabetes to enable estimation of baseline retinopathy prevalence, which is done in DR prevalence studies [21]. Including individuals without diabetes in this pilot study therefore allows feasibility to be assessed in this group, who will be included in the main DIAGNOSE-DR study alongside individuals with diabetes. All participants provided informed written consent, which was obtained by trained research staff under the supervision of investigators (MMD, SS, NW). People with mental or cognitive impairment who could not give informed consent, and pregnant women were excluded, the latter because of the need for mydriasis. People who could not safely have their eyes dilated (history of acute angle closure glaucoma, hypersensitivity to mydriatic eye drops, using eye drops for glaucoma, penetrating eye injury in past six months, self-driving or other activity after dilatation), and people with apparent eye infections were excluded for safety reasons. Image capturers and training Non-specialist healthcare workers were recruited to capture images. While junior or undergraduate doctors and nurses, health promotion graduates and biomedical science graduates were eligible, three main categories of image capturers were utilized; medical students, medical graduates and non-medical graduates. They underwent a structured, short-duration training program prior to data collection to capture monoscopic retinal fundus images (retinal images) using the Remidio FOP NM-10 and create AI reports for diabetic retinopathy. The training consisted of two sessions of 2 h per session. Image capturers were trained on the operation of the retinal imaging device, patient positioning and communication, image quality criteria (field definition, positioning the camera clarity, artefact recognition), and the process of generating the AI reports. Each trainee was required to capture ten satisfactory images with adequate focus, field, disc quality and image quality with the absence of artefacts) on volunteer participants. They also needed to demonstrate the ability to independently operate the device and generate the AI reports, before being cleared to be image capturers in the study. Initially, trainers were trained by MPP in December 2022 on techniques for capturing digital retinal images using a hand-held digital imaging device by adapting the protocol developed for Sri Lanka by Piyasena and Gudlavellati in the year 2018 [22]. The trainers (TVG, HR, NW) subsequently received further training on using the device, including on-site training from the manufacturer of the device (TVG, NW). Data collection Retinal images were captured for two fields (optic-disc-centred and macula-centred) using the Remidio FOP NM-10 device, both prior to and after mydriasis. Mydriasis was achieved using tropicamide 1% optical solution with participants waiting 20 min, or more if needed, to ensure that the pupils were adequately dilated. Automatic interpretation AI reports for diabetic retinopathy were generated on the device for each set (pre- and post-mydriasis) of retinal images. These reports state if a participant has signs of retinopathy if moderate non-proliferative DR or more severe DR is detected (“referrable DR”). The cut-point for detection of DR is not configurable [23]. Additionally, participants underwent slit lamp biomicroscopy examination by the ophthalmologist (MMD) who recorded whether the participant had diabetic retinopathy. The ophthalmologist was blinded to all retinal images taken by the device and AI reports. Data were collected on participant characteristics including age, gender, ethnicity, history of ophthalmic conditions such as cataracts. Visual acuity was recorded with and without glasses using a Snellen chart. We defined poor visual acuity as having unaided visual acuity of 6/18 or less. Data on the image capturers, including age, gender, and professional status were also collected. Data were collected using a computer-assisted personal interviewing platform, iFormBuilder (Zerion Software Inc., Herndon, VA, USA), with built-in skip logic and checks for unlikely values during data entry. The images taken were evaluated by the ophthalmologist at a later date for (a) quality of images where a level of gradability was assigned based on the proportion of macular-centred images that could be graded (less than 50%, 50%, 75%, 100%) using image quality grading guidelines validated in Sri Lanka by Piyasena and Gudlavalleti [22] and (b) presence of DR including non-proliferative and proliferative diabetic retinopathy using all available retinal images for each participant. Images were considered to be ungradable if the level of gradability was < 50%. Statistical analysis Descriptive statistics were used to summarize the study participants’ characteristics, including age, sex, cataracts and vision impairment. The significance of differences between people with and without diabetes was calculated using Student’s t-test for age as a continuous variable, and Fisher’s exact test for the binary variables of sex, cataract and poor visual acuity [24, 25]. The proportion of participants with at least one ungradable eye and with both eyes ungradable were calculated, along with the proportion of eyes with ungradable images. We used McNemar’s chi-square test to compare the pre- and post-dilation proportions, accounting for measurements within the same individual before and after mydriasis [26]. We used Fisher’s exact test to assess differences in the proportion of participants with ungradable images across levels of participant characteristics. To check for associations between gradability and participant and image capturer characteristics, we calculated the proportion of individuals with at least one ungradable eye and with both eyes ungradable for subgroups of participants characteristics (age group, gender, diabetes status, cataract status, poor unaided visual acuity), along with the proportion of eyes that were ungradable by characteristics of the image capturers (sex, and whether they were medical students, junior doctors or non-medical graduates). Finally, we ran a mixed effects logistic regression to identify if any of the participant-level or image-capturer characteristics were predictive of capturing ungradable retinal images [27]. To assess the DTA, we calculated the sensitivity, specificity and area under the (receiver operating characteristic) curve (AUC) of two endpoints compared to detecting DR on slit lamp biomicroscopy. The first endpoint was identifying individuals for referral, either due to a result of DR by the AI report or due to < 50% gradability of images captured for one or both eyes, to test the real-world use of the camera for DR screening. The second was identifying eyes with DR, analysing only gradable images, to assess the accuracy of the AI interpretation on adequately captured (gradable) images. We calculated 95% confidence intervals using exact binomial confidence intervals for sensitivity and specificity, and asymptotic normal approximation for the AUC [28]. All statistical analysis was carried out in Stata [29]. The data and analysis program are available in Supplementary Material 1 and Supplementary Material 2. When there was disagreement between the AI report and reference standard, further analysis of the images and data collected were undertaken to identify and classify the reasons for false negatives, to understand reasons for potential under-performance by the AI software, some of which may be population specific, such as increased pigmentation in the Sri Lankan population and cataracts. Results Of the 123 participants that enrolled, 119 had at least one image captured prior to and after mydriasis, including 51 people with diabetes and 68 people without diabetes (Fig. 1). One participant had pre-dilation images, however, due to a data collection error, post-dilation images for this person are not available. A total of 114 of the 119 people were interviewed. Participant flowchart of the study The characteristics of the participants are shown in Table 1, with 82.4% aged 45 years and above, and 51.3% male. There were 51 people with diabetes (42.9% of the sample), and 16.8% of people reported a history of cataract in at least one eye, or had cataracts mentioned in their medical records. Participants with diabetes were older than those without diabetes (p = 0.003). Table 1Characteristics of participants with both pre- and post-dilated imagingPeople with diabetesPeople without diabetes p All participantsN = 51%N = 68%N = 119% Age category 0.00320–3412.0811.897.635–4435.9913.21210.145–54917.61623.52521.055–642549.02942.65445.465–741325.545.91714.375 and above00.022.921.7 Sex 0.2Male3058.83145.66151.3Female2141.23754.45848.7 Cataract 0.3No4078.45986.89983.2Yes1121.6913.22016.8 Poor visual acuity (6/18 or less) 0.4No2447.13855.96252.1Yes2752.93044.15747.9 Total 51 100.0 68 100.0 119 100.0 Notes: p-values show the significance of differences between people with and without diabetes for each characteristic, using Student’s t-test for age as a continuous variable, and Fisher’s exact test for sex, cataract and poor visual acuity Characteristics of participants with both pre- and post-dilated imaging Notes: p-values show the significance of differences between people with and without diabetes for each characteristic, using Student’s t-test for age as a continuous variable, and Fisher’s exact test for sex, cataract and poor visual acuity Gradability of the images captured by dilation status are show in Table 2. A significant improvement is seen in gradability after mydriasis: the images from at least one eye were ungradable in 46.2% participants prior to mydriasis, dropping to 15.1% after mydriasis (difference 31.1%, p < 0.001). When analysing images taken on individual eyes, 35.7% of eyes were ungradable prior to mydriasis, dropping to 10.1% after mydriasis (difference 25.6%, p < 0.001). Table 2Comparison of proportions of ungradable images by dilation statusPre-dilation (%)Post-dilation (%)Difference (%)McNemar’s 𝛘2 p Analysis by individuals (n = 119) Individual with at least one eye that is ungradable46.215.131.155.9< 0.001 Analysis by eyes (n = 238) Number of eyes that are ungradable35.710.125.645.9< 0.001 Comparison of proportions of ungradable images by dilation status Table 3 shows the percentage of participants that had at least one eye that was ungradable, and both eyes ungradable prior to and after mydriasis by participant characteristics. A higher percentage of individuals had ungradable eyes pre- and post-dilation for one or both eyes in older age groups. The increasing percentage of ungradable eyes in older age groups is statistically significant for pre-dilation images (Fisher’s Exact test p < 0.001). A higher percentage of people with cataracts in one or both eyes had ungradable eyes, with 85.0% of people with cataracts having at least one ungradable eye compared to 38.4% of people without cataracts (Fisher’s Exact test p < 0.001). The difference in gradability by cataract status, though still present, was not statistically significant post-dilation. There was some variation seen by visual acuity, where a higher percentage of people with poor unaided visual acuity of 6/18 or less had at least one ungradable eye compared to people without poor visual acuity (57.9% vs. 35.5%, Fisher’s Exact test p = 0.02). No statistically significant variation was seen by sex or diabetes status of the participant. Table 3Percentage of participants within each subgroup who had at least one or both eyes ungradable, before and after dilationParticipant characteristic N At least one eye ungradable (%)Both eyes ungradable(%)Pre-dilationPost-dilationPre-dilationPost-dilation Age category 20–34911.1***0.0*0.0*0.035–44128.38.30.00.045–542528.08.016.04.055–645459.313.035.25.665–741770.641.235.311.875 and above2100.050.050.00.0 Sex Male6149.29.829.58.2Female5843.120.720.71.7 Diabetes No6852.914.726.54.4Yes5137.315.723.55.9 Cataract No9938.4***12.121.2*4.0Yes2085.030.045.010.0 Poor visual acuity (6/18 or less) No6235.5*11.321.03.2Yes5757.919.329.87.0 Total 119 46.2 15.1 25.2 5.0 Notes: Significance of Fisher’s exact test between participant category levels denoted by *** p < 0.001, ** 0.001 < p ≤ 0.01, * 0.01 < p ≤ 0.05 Percentage of participants within each subgroup who had at least one or both eyes ungradable, before and after dilation Notes: Significance of Fisher’s exact test between participant category levels denoted by *** p < 0.001, ** 0.001 < p ≤ 0.01, * 0.01 < p ≤ 0.05 The proportion of ungradable eyes by different categories of image capturer characteristics are shown in Supplementary Material 3, Table S1. There was little variation seen by sex or qualifications. The results of the mixed effects logistic regression of factors associated with ungradable eyes are shown in Supplementary Material 3, Table S2. In pre-dilated eyes, each additional year of participant age was associated with higher odds of ungradable eyes (odds ratio 1.1; 95% confidence interval (CI) 1.1–1.2, p = 0.004). There was no other patient-level or image-capturer-level characteristic that had a statistically significant association with ungradable eyes. Table 4 shows the DTA of the AI report generated by the camera compared to the gold-standard–slit lamp biomicroscopy. The sensitivity and specificity for identifying individuals after mydriasis who required referral, defined as those detected with DR by the AI software, or with at least one ungradable eye, were 73.3% (95% CI: 44.9–92.2) and 79.8% (95% CI: 70.8–87.0) respectively, with an AUC of 0.77 (95% CI: 0.64–0.89). Without mydriasis, the AUC fell to 0.60 (95% CI: 0.47–0.73) (Supplementary Material 3: Table S3). For identifying eyes with referrable DR, after analysing only gradable images after mydriasis, the sensitivity and specificity was 61.1% (95% CI: 35.7–82.7) and 96.9% (95% CI 93.5–98.9) respectively, with an AUC of 0.79 (95% CI: 0.67–0.91). Table 4Diagnostic test accuracy of camera in detecting referrable DRPerformance in:SensitivitySpecificityAUCIdentifying individuals who require referral for DR or due to ungradable images (“real-world”)73.3(44.9–92.2)79.8(70.8–87.0)0.77(0.64–0.89)Identifying eyes with DR, analysing only gradable images (AI-performance)61.1(35.7–82.7)96.9(93.5–98.9)0.79(0.67–0.91)Notes: 95% confidence intervals shown in parenthesis Diagnostic test accuracy of camera in detecting referrable DR 73.3 (44.9–92.2) 79.8 (70.8–87.0) 0.77 (0.64–0.89) 61.1 (35.7–82.7) 96.9 (93.5–98.9) 0.79 (0.67–0.91) Notes: 95% confidence intervals shown in parenthesis Of the 21 participants screened as false positives—that is, people who would be screened as needing a referral, but did not have DR on slit lamp examination—15 were due to at least one eye having ungradable images (Supplementary Material 3: Fig. S1). A breakdown of false positives and negatives considering only gradable eyes, is shown in Supplementary Material 3: Fig. S2. Each false positive and false negative was reviewed retrospectively by the ophthalmologist. Regarding false negatives, where DR was seen on slit lamp biomicroscopy but not visible on the retinal images, it is likely that DR signs seen on the slit lamp biomicroscopy were obscured by cataracts or poorly focused images in the retinal images captured. For the false positives, it appears that that the AI interpretation picked up features not related to DR in four eyes. However, there were two eyes where DR changes were clearly seen in the retinal images but not picked up on slit lamp biomicroscopy: in one, a haemorrhage was in a less common location, nasal to the disc. Discussion This study provides key information on the feasibility of using the Remidio FOP NM-10, operated by non-specialist healthcare workers with minimal training, to screen for DR in a low-resource setting. Firstly, it finds that mydriasis is recommended, in line with a study in a tertiary general medical clinic in Sri Lanka, using a different handheld retinal camera (Zeiss-Visuscout100) [30], which found that 24% and 30% images of captured images were ungradable for two different physician image capturers prior to mydriasis, and that this dropped to 5% and 7% respectively, following mydriasis. Another study in Iran on participants with dark irides also demonstrated that mydriasis similarly improved the gradability of images captured by a Canon CR2-45NM fundus camera, and also demonstrated that the quality of images after dilation with one drop of tropicamide 1% was comparable to the quality obtained after three drops of tropicamide 1% [31]. Furthermore, the experience from our investigators (SP, MPP) suggests that mydriasis is necessary to capture images of sufficient quality for South Asians using the Remidio FOP NM-10, however, published studies are limited. Given the small but present risk of angle closure glaucoma when dilating eyes, we wanted to confirm the need for mydriasis more formally for the Remidio device. This pilot study demonstrated that mydriasis significantly improved the gradability of retinal images captured using an AI-assisted hand-held retinal imaging camera operated by non-specialist healthcare workers. Prior to dilation, nearly half (46.2%) of the participants had at least one eye that was ungradable, which reduced to 15.1% with dilation. Post-dilation, both eyes were ungradable in only 5% of participants. Along with previous studies [30–33], our findings reinforce the need for mydriasis in Sri Lankans–as a population with dark irides–when images are captured by non-specialist healthcare providers with fairly minimal training, and suggests that mydriasis with one drop of tropicamide 1% should be sufficient. The gradability of images was poorer in older people, and to a lesser extent, in people with an unassisted visual acuity of 6/18 or worse, though small sample sizes limit precision and these findings should be interpreted with caution. The gradability was also poorer in people with cataracts The finding of poorer image gradability in older people, people with cataracts and those with poorer visual acuity is consistent with other studies [30, 33, 34]. A significant improvement in gradability with mydriasis was seen in these three groups of people, confirming that mydriasis is both helpful and important in increasing DR screening for these groups as well. Furthermore, the low sensitivity and specificity of a non-mydriatic screening workflow reinforces the importance of mydriasis in the study group. Our study, though not powered to adequately assess this, did not find any marked differences in gradability of images based on the characteristics of the individuals who captured the images, which may suggest that it is reasonable to train a variety of healthcare workers, including undergraduate medical trainees, junior doctors and non-physician graduates to capture retinal images for DR screening. The preliminary estimates of DTA of using the AI-assisted retinal camera in a screening setting were promising. In the “real-world” camera-based screening system, individuals would be referred if either the camera detected referrable DR, or if they had one or more ungradable eyes, which would require referral for further evaluation by a specialist. The system demonstrated good sensitivity (73.3%) and specificity (79.8%) for identifying individuals requiring referral when both DR and ungradable images were considered. The major contributor to the sensitivity and specificity of this camera-based screening system was the gradability of images, rather than the AI itself: the sensitivity and specificity of the AI analysis of individual eyes rose to 90.0% and 96.9% respectively when excluding eyes with ungradable images. This finding is consistent with previous studies using a Remidio FOP NM-10 device, which found the sensitivity for detecting referrable diabetic retinopathy in gradable images to be above 80% and specificity above 90% [19, 23]. Most “false-positives” in referrals for patients in the real-world camera-based screening system—where those with DR, or one or more ungradable eyes would be referred—were due to the presence of ungradable images (15/21), suggesting that any modifiable factors which could improve the gradability of images captured could reduce such false-positives. Such modifiable factors may include better training of image capturers to capture a higher percentage of gradable images, as even after dilation, 15.1% of participants still had one or more eyes with ungradable images. This may be possible, given that a previous study in Sri Lanka, two trained physicians had a lower percentage of ungradable images on dilated eyes (4.9–7.0%) [30, 33]. In our study, our image capturers received a short training for approximately 4 h, with 6 of the 14 capturers taking less than 40 images during the pilot study. There were some instances (7/214 gradable eyes) where the AI detected retinopathy when none was detected on slit lamp biomicroscopy or on review of the captured images. This was mostly associated with hazy images related to cataracts, and the specificity remained high. In four participants, the AI did not detect retinopathy though the participants had DR on slit lamp examination (“false negatives”). On further examination of the retinal images, though these images were assessed as having a gradability of 50% or more, the images were blurred, with three of those cases having a cataract. Separately, there were two eyes where signs of DR were missed on slit lamp biomicroscopy but captured and identified by the AI, and confirmed on ophthalmologist review of the retinal images as well. The sensitivity and specificity results of this pilot study are similar to findings from a larger study in Sri Lanka of 700 people with diabetes, where the sensitivity was 78%–83% and specificity was 70%–76% [30]. However, this study used nine trained general physicians to capture retinal images using a hand-held retinal imaging camera (Zeiss-Visuscout100) in a tertiary general medical clinic, and two trained general physicians to grade the images. These results suggest that the use of AI-assisted retinal imaging by non-specialist healthcare workers, particularly with dilation, would be feasible in a low-resource clinic setting to detect people with DR. The risk of lower specificity is unnecessary referrals to an ophthalmologist for further evaluation of potential diabetic retinopathy, and potential stress for the patient while awaiting further evaluation, while lower sensitivity may result in patients with DR being missed, untreated, and the patient being falsely reassured. The risk of completely missing treatable DR may be reduced somewhat if screening is repeated at a reasonable frequency. However, further research with larger sample sizes and in routine programmatic settings is warranted to confirm these findings and to optimize protocols balancing sensitivity, specificity, and feasibility in resource-constrained environments. Logistically, the Remidio FOP NM-10 is practical, with no technical problems encountered during the pilot study. The device can be used with or without an internet connection to capture images and generate the AI report in real time, which is useful in settings with poor internet connectivity. The battery life of the device allowed it to be used continuously in a clinic setting lasting approximately three hours, and charging the device is also convenient. Additionally, images and reports can be easily shared through several modalities, allowing for teleophthalmology. A recent literature review lists the Remidio FOP NM-10 as one of four portable retinal cameras with AI integration, and the only one validated in a South Asian population [35]. The primary aim of the pilot study was to assess the use of the hand-held retinal imaging device in a controlled, research-clinic setting and to inform the design of the clinical protocols of the main study. Given the study’s requirements, limited budget and timeframe, the target sample size was set at 50 people with diabetes and 50 people without diabetes. Ultimately, we obtained pre- and post-dilation images on 51 people with diabetes and 68 people without diabetes. However, the modest sample size limits the ability to detect anything beyond major issues in capturing gradable images and factors which have large impacts on diagnostic test accuracy. Additionally, only one ophthalmologist performed the slit lamp biomicroscopy and review of images: future studies will include several ophthalmologists. Conclusions The study supports the feasibility of using an AI-assisted retinal camera operated by non-specialist healthcare workers to detect referrable DR in a clinic setting. It suggests that a real-world screening program would have a reasonable sensitivity and specificity, with the gradability of images the major driver of misclassification, rather than the AI itself. The gradability of images dropped with the age of the participant, the presence of cataracts, and possibly for participants with poor visual acuity. Importantly, the study confirms that mydriasis substantially improves the gradability of retinal images—including for the previously mentioned categories of patients—and finds that that the AI-interpretation, when restricted to only gradable retinal images, has a high specificity for DR. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1: Data file Supplementary Material 2: Analysis code. Supplementary Material 3: Document with results from extended analyses on image gradability by image capturer characteristics, and flowcharts of DR classification. Acknowledgements The team would like to acknowledge the Eye Care Centre, Faculty of Medicine, University of Colombo, and its staff for collaborating in this study; Professor Vajira H W Dissanayake, Dean of the Faculty of Medicine, University of Colombo, for his support of the project; and Yasodhara Kapuge, Raaidah Wahab Daniel, Kethmi Navodya, Kavisha Silva and Janisha Ratnayake for their contributions to previous versions of the study protocol. Abbreviations Computer Assisted Personal Interviewing Confidence Interval Diagnostic Test Accuracy Diabetic Retinopathy Eye Care Centre Early Treatment Diabetic Retinopathy Study Fundus on Phone Non Mydriatic-10 Low- and Middle-Income Countries Open Science Framework People without Diabetes Mellitus Sri Lanka Health and Ageing Study Author contributions MDN, SSi, RPR-E and NW were involved in securing funding. MMD, MDN, SSi, RPR-E and NW were involved in the conception of the research questions, and MMD, NW, TVG, HR, NiW, MDN, MPP, SSi, SSa, SD, RPR-E and MDS contributed to the study design. MPP provided initial training to the trainers on capturing digital retinal images. NW wrote the first manuscript draft. All authors reviewed and approved the final manuscript. Funding This study is supported by the Velux Stiftung Foundation (Grant number 573265). The sponsor did not and will not have any role in the study design; collection, management, analysis, and interpretation of data; writing of the report; or the decision to submit the paper for publication. Data availability The data underlying this study and the analysis code used are available in Supplementary Material 1 and Supplementary Material 2. Declarations Ethics approval and consent to participate This study involved was approved by the Ethics Review Committee, Faculty of Medicine, University of Colombo (EC-23-116). The version of the protocol approved by the ERC is DDR-PROT-1.04. Participants gave written informed consent to participate in the study before taking part. The research adhered to the ethical principles of the Declaration of Helsinki. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1: Data file Supplementary Material 2: Analysis code. Supplementary Material 3: Document with results from extended analyses on image gradability by image capturer characteristics, and flowcharts of DR classification. Data Availability Statement The data underlying this study and the analysis code used are available in Supplementary Material 1 and Supplementary Material 2.