International Journal of Retina (IJRETINA) 2025.
Volume 8.
Number 1.
P-ISSN.
E-ISSN.
EVALUATING THE ACCURACY OF ARTIFICIAL INTELLIGENCE (AI)-INTEGRATED,
SMARTPHONE-BASED SCREENING FOR DIABETIC RETINOPATHY:
SYSTEMATIC REVIEW
Shofia Medina Samara.
Ariyoga Kun Laksono Resident of Ophthalmology Department.
General Hospital of Saiful Anwar Malang.
Indonesia Faculty of Medicine.
Universitas Brawijaya.
Malang.
Indonesia Abstract Background: Diabetic retinopathy (DR is the most common microvascular complication of diabetes that can cause vision problems and blindness that poses a significant health risk and financial burden, increasing the needs to effectively screen and manage diabetic eye disease.
The current method of screening for diabetic eye disease relies on human experts to analyze the results.
Alternatively, recent advancements in artificial intelligence (AI) especially deep learning (DL) and retinal imaging using smartphones offer a promising solution for both patients and ophthalmologists, potentially improving patient compliance and making telemedicine more efficient for DR screening.
Purpose : To represent on accuracy of AIAcintegrated process in smartphone-based DR screening and to compare the various study methods and settings used to achieve this accuracy.
Method: Literature search on current DR screening programs was conducted following the Preferred Reporting Items for Systematic Reviews and Meta-Analyzes (PRISMA) framework on Google Scholar.
Scopus.
Web of Science.
PubMed.
Medline, and Embase with most recent search was updated on June 1st, 2024.
Key information was extracted from the studies included author names, journal, year of publication, country, sensitivity, specificity, positive and negative predictive values .
f availabl.
, study methods, and settings.
Result: The study identification process resulting in 9 selected studies.
The performance metrics reported included intergrader/intramodality agreement, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).
The sensitivity of AI in detecting DR ranged from 77-100%, while specificity ranged from 61.
4 - 95.
PPV and NPV were reported less frequently, with ranges of 48.
1 - 92.
92% and 91.
3 - 99.
46%, respectively.
Intergrader agreement was within range = 0.
45 Ae 0.
Conclusion: The studies reviewed in this paper collectively represents the potential of smartphone based integrated with AI in revolutionizing DR screening.
The high sensitivity and specificity achieved by various AI algorithms, often exceeding the standards set by regulatory bodies like the FDA and ETDRS, highlight their accuracy in detecting DR and its severity levels.
The accessibility and userfriendliness of smartphone-based retinal imaging further enhance the coverage of DR screening, particularly in underserved areas with limited resources and internet connectivity.
Keywords: artificial intelligence, smartphone, diabetic retinopathy, screening Cite This Article: SAMARA.
Shofia Medina.
LAKSONO.
Ariyoga Kun.
Evaluating the accuracy of Artificial Intelligence (AI)-integrated.
Smartphone-based screening for Diabetic Retinopathy: Systematic Review.
International Journal of Retina, [S.
ISSN
Available <https://w.
com/index.
php/ijretina/article/view/294>.
Date accessed: 05 mar.
doi: https://doi.
org/10.
35479/ijretina.
Published by: INAVRS https://w.
org/ | International Journal of Retina https://ijretina.
Telemedicine offers a way to make fundus
INTRODUCTION
Diabetic eye disease, a Correspondence to:
Shofia Medina Samara.
General Hospital of Saiful Anwar Malang.
Indonesia, shofiamedina@gmail.
screening more accessible by allowing patients to be common issue for people screened at convenient with Type 1 or Type 2 eliminating the need to travel to far-off hospitals for diabetes, is a major reason an eye examination with an ophthalmologist.
While for vision loss in working-age telemedicine makes DR screening more accessible, it In 2017, an estimated 425 million adults still requires human experts to analyze the images.
worldwide had diabetes, double the number in 1980.
However, recent progress in deep learning (DL)- and this figure is expected to reach 629 million by based artificial intelligence (AI) presents a potential Diabetic retinopathy (DR) roughly affected 30% to 45% people with diabetes and is the most ophthalmologists, as a way of detecting retinal common microvascular complication where 10% of images which may be sightActhreatening,10 potentially these cases are vision-threatening, meaning they could significantly impair vision, even blindness.
and locations, telemedicine more efficient for DR screening.
The increasing prevalence of diabetic eye disease DL in AI software can automatically analyze the is a worldwide problem that poses a significant retinal images and provide recommendations for health risk and financial burden for both individuals follow-up care or referrals, reducing the workload of and societies, particularly in developing countries This increased convenience and where it is becoming more common.
This surge in participation in DR screenings and detections.
healthcare providers to effectively screen and While some research has shown that AI can manage diabetic eye disease.
A fully implemented screen for diabetic retinopathy in national diabetic eye screening program (DESP) by developed country like UK.
UK successfully screened a vast majority of diabetic expensive, desktop-based fundus cameras that may patients resulting in decrease of DR as the primary not be accessible in rural areas.
cause of blindness in working-age adults in the UK, highlighting the effectiveness of such programs in preventing and treating DR-related vision loss.
Despite 3,11 it often relies on Retinal imaging using smartphones has proven to be an effective and affordable method for diabetic retinopathy screening.
13,14
Similarly, using AI to compliance with recommended guidelines is low devices has also been shown to be valid for screening because of a lack of understanding about the disease in community settings.
15 This study aims to represent and its potential complications, difficulty accessing on accuracy of AIAcintegrated process in smartphone- based DR screening and to compare the various One study from study methods and settings used to achieve this coverage for these exams.
Indonesia reflects that despite being knowledgeable and having a positive attitude towards DR screening, general practitioners (GP)s did not consistently METHOD implement it in practice because of limited The author evaluated the current state of experience, lack of confidence in diagnosing fundus smartphone-based DR screening programs by abnormalities, and lack of equipment in primary searching Google scholar and PubMed via Medline for open-access studies published in English using Published by: INAVRS https://w.
org/ | International Journal of Retina https://ijretina.
these keywords: Audiabetic retinopathy, artificial between the review authors were resolved by intelligence.
DR screening, smartphoneAy.
In addition, discussion until a consensus was reached.
we sought reference lists and publicly accessible commercially available diabetic retinopathy (DR) Data Extraction and Analysis The systematically gathering and combining relevant screening algorithms.
A comprehensive study was done to compare the accuracy and methods on the pre-existing research on smartphone-based.
AI-integrated of diabetic retinopathy screening.
The PRISMA (Preferred Reporting Items for Systematic Reviews and MetaAnalyze.
framework guided the review process by establishing a set of predefined criteria and data from the selected studies.
Data extraction included meticulously reviewing each study to identify information such as the study's methods, setting, sample size, the AI software or tools used, intergrader/intermodality agreement .
appa value.
and the study's accuracy as indicated by sensitivity, specificity, and predictive values.
The PRISMA diagram (Figure .
outlines In our analysis, we divided the results into two the search strategy used, including the criteria for categories to assess the accuracy of AI-integrated, including and excluding studies.
The most recent smartphone-based devices in diagnosing any level search was conducted on June 1st, 2024.
Key of DR and the more severe, referable DR (RDR) information extracted from the studies included defined as moderate non-proliferative DR with author names, journal, year of publication, country, diabetic macular edema (DME), severe NPDR, or study type, sensitivity, specificity, positive and worse, regardless of DME presence.
This assessment negative predictive values .
f availabl.
, and study was based on clinical grading using retinal images methods and settings.
without any OCT examination.
16 Sensitivity and specificity data from studies reporting detailed test Eligibility criteria outcomes such as true positive (TP), false positive We sought to include studies that had outcome (FP), true negative (TN), and false negative (FN) were measures of accuracy for sensitivity and specificity visually summarized in a forest plot using RevMan .
nd predictive values, if availabl.
for smartphone- Studies without this detailed information were excluded from the forest plot.
This approach allowed Only studies in English were included.
for a clear and concise presentation of the accuracy Studies which not stated the accuracy on DR of smartphone-based devices in diagnosing any DR screening are excluded.
Researches on data set and RDR.
AI-integrated without the patient are excluded.
Conference abstracts, review articles, letters to the editor, editorials, and correspondence notes were excluded.
Risk of Bias The searching databases, resulting in 24 records.
These reports were screened for eligibility, leading to the We used the QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studie.
tool to assess the risk of bias and applicability of all included studies.
Two reviewers (SMS and AKL) performed assessment of selected studies independently.
Any disagreements RESULTS inclusion of 9 studies in the review [Figure .
Twelve reports were excluded for reasons such as focusing on data sets rather than patients, not using smartphone, not stating sensitivity and/or specificity, and studies using smartphone but not integrated with AI.
Ultimately, 9 studies were selected for Published by: INAVRS https://w.
org/ | International Journal of Retina https://ijretina.
Figure 1 Data selection steps using the Preferred Reporting Items for Systematic Review and MetaAcAnalyzes.
detailed analysis and synthesis in the review.
Forest The study by Malerbi .
, which categorized plot contains the summary of sensitivity and outcomes as "more than mild DR .
tmDR)," was specificity from five studies in detecting any DR incorporated into the RDR forest plot due to the [Figure .
and five studies in detecting RDR [Figure equal definition of RDR.
Published by: INAVRS https://w.
org/ | International Journal of Retina https://ijretina.
Any DR Figure 2 Forest plot showing sensitivity and specificity range of AI software in detecting any severity of diabetic retinopathy (DR) from smartphone-based images RDR Figure 3 Forest plot showing sensitivity and specificity range of AI software in detecting referable diabetic retinopathy (RDR) from smartphonebased images Table 1 illustrates the risk of bias assessment of Table 2 summarized studies evaluating the included studies using the QUADAS-2 tool.
Most performance of various smartphone-based.
AI studies employed low risk of bias on reference integrated screening process in detecting DR.
The studies were arranged chronologically, starting with because this study did not state exact reference used the research published in 2018 progressing towards to grade DR.
Some studies were deemed high risk of more recent publications.
The studies employed bias in index test because the images used were labelled by a single grader.
Risk of bias in patient International Clinical Diabetic Retinopathy (ICDR) selection from couple of studies were scored high scale being the most common.
Intergrader and because the study did not present demographic of the patient.
Overall, the applicability concern of the inconsistencies in image interpretation.
standard except from Wroblewski, et al.
studies is low.
Published by: INAVRS https://w.
org/ | International Journal of Retina https://ijretina.
Table 1 Risk of Bias Assessment Using Quadas-2 Tools However, the accuracy results were promising, with sensitivity ranging from 83.
3% to 98.
84% and specificity 4% to 95.
5 PPV and NPV were also generally high, indicating the reliability of smartphone-based screening in identifying individuals with and without DR.
Table 3 provides a summary of different study settings DR screening.
data from nine studies conducted in India, the USA.
Brazil, and Mexico, which explored the use of smartphone-based AI software for diabetic retinopathy screening.
The studies used various AI software (Medios AI.
EyeArt.
PhelcomNe.
, smartphones (HTC One, iPhone .
, and image capture methods .
ith and without pupil dilatio.
The number of fundus images taken and the fields of view varied across the studies.
Some studies specified the healthcare workers involved in the screening process .
rained technicians, medical students, health worker.
, while others did not.
Published by: INAVRS https://w.
org/ | International Journal of Retina https://ijretina.
Table 2 Summary of selected study showing study samples and accuracy on DR screening.
CI: Confidence Interval.
DR: Diabetic Retinopathy.
NPDR: Non-Proliferative Diabetic Retinopathy.
ICDR: International Clinical Diabetic Retinopathy.
STDR: Sight Threatening Diabetic Retinopathy.
DME: Diabetic Macular Edema.
RDR.
Referable Diabetic Retinopathy.
PDR: Proliferative Diabetic Retinopathy.
ETDRS: Early Treatment Diabetic Retinopathy Study.
DESP: Diabetic Eye Screening Programme.
vtDR: Vision-threatening Diabetic Retinopathy.
mtmDR: more than mild Diabetic Retinopathy Author Sample Size Sengupta, et Rajalakshmi, et al.
135 individuals 233 eyes 296 individuals 2408 images DR classification ICDR18.
DESP19,20 Intergrader Agreement any DR = 0.
VTDR = 0.
ICDR,18 Referable DR (RDR) defined as moderate NPDR and above Natarajan, et Sosale, et al.
231 individuals ICDR18 Any DR = 0.
297 individuals ICDR.
DME18 Any DR = 0.
DME = 0.
Tyson, et al.
69 individuals 119 eyes Sosale, et al.
922 patients modified Airlie House classification system used in ETDRS24,25 and ICDR18 ICDR.
DME18
Jain, et al.
1378 individuals ICDR18 = 0.
45 A 0.
= 0.
= 0.
79 - 0.
Any DR 95.
8% .
DME 97% .
PDR 78.
1% .
STDR 99.
1% .
RDR 99.
3% .
Accuracy results, 95% CI Specificity Any DR 89.
1% 89.
Any DR 80.
2% .
DME 75.
8% .
PDR 89.
8% .
STDR 80.
4% .
RDR 68.
8% .
Any DR 92% .
RDR 88.
4% .
Any DR 95.
45% .
Ae97.
RDR 86.
73% .
8Ae90.
Any DR 85.
2% .
RDR 100% .
Any DR 86.
78% .
9Ae90.
RDR 98.
84% .
6Ae.
Any DR 92.
RDR 75.
Any DR RDR 99.
Per patient/eye RDR 87.
0% .
8% .
Per patient/eye RDR 78.
6% .
5% .
Any DR 83.
3% .
RDR 93% .
STDR 95.
Per Patient/Eye Any DR 89.
1% .
/88.
6% .
5Ac92.
RDR 100% .
7Ac100.
/100% .
Any DR 95.
5% .
RDR 92.
5% .
Any DR 87.
RDR 78.
Any DR RDR 97.
Any DR Medios 95% EyeArt 89% Any DR Medios 93% EyeArt 93% Intramodality Agreement any DR =0.
STDR = 0.
RDR = 0.
= 0.
Malerbi, et Wroblewski, et al.
824 individuals 3255 images 248 patients 2130 images ICDR18 Any grade DR by a Sensitivity Any DR 93.
mtmDR 97.
8% .
Per Patient/Eye Any DR 94.
4% .
/94.
8% .
8Ac95.
RDR 89.
7Ac91.
/91.
7Ac92.
mtmDR 61.
4% .
Any DR Medios 94% .
EyeArt 94% .
Any DR Medios 94% .
EyeArt 86% .
PPV
NPV
Any DR 89.
DME 67.
PDR 48.
STDR 75.
RDR 74.
Any DR DME 98% PDR 97.
STDR 99.
RDR 99.
Published by: INAVRS https://w.
org/ | International Journal of Retina https://ijretina.
Table 3 Summary of selected studies showing study settings, methods, study samples, tools, and images used in DR screening.
AI: Artificial Intelligence.
Author Country Screener AI Software Used Smartphone Used Pupil DIlation Yes Sengupta, et al.
India Not specified Medios AI28 HTC One with Remidio FOP (Fundus on Phon.
29 Rajalakshmi, et India Not specified EyeArtTM .
Not specified Yes Natarajan, et al.
India Health worker Medios AI28 Not specified Yes Sosale, et al.
Tyson, et al.
Sosale, et al.
Jain, et al.
Malerbi, et al.
Wroblewski, et
India
Trained technician14
Medios AI28
USA
EyeArtA .
Medios AI28 Yes India Medical student and medical intern Trained technician14 Iphone 6 Remidio Non Mydriatic (NM) FOP29 Iphone with Retinascape31 Iphone 6 with Remidio NM FOP29 India Healthcare workers Medios AI28 Smartphone with Remidio NM FOP29 Brazil 9 examiners including med students 3 graduate students PhelcomNet.
Modified Xception32 Medios AI,28 EyeArtTM .
Smartphone with Eyer.
Phelcom Technologies33 Smartphone with Remidio FOP29 Yes Mexico Yes Fundus Images 45A Field of View (FOV) 3 fields .
osterior pole .
acula-centere.
, nasal field, and superotemporal fiel.
45A FOV 4 fields .
acula centred, disc centred, superior-temporal and inferior-tempora.
Anterior segment 3 fields .
osterior pole .
ncluding disc and macul.
, nasal and temporal fiel.
3 fields .
osterior pole .
acula centre.
, nasal and supero-temporal fiel.
5 sequential images .
entral, inferior, superior, nasal, and tempora.
2 images, disc and macula centred 3 fields .
osterior pole, nasal and temporal 45A FOV, 2 images of posterior pole .
acula and disc centre.
3 fundus fields .
osterior pole .
isc and macul.
, nasal, tempora.
Published by: INAVRS https://w.
org/ | International Journal of Retina https://ijretina.
DISCUSSION
In this review, we reported the sensitivity, examination of the false positives showed that the algorithm often mistook normal variations in fundus intergrader/intermodality agreement and compared pigmentation or image imperfections for signs of the different study settings from nine studies.
These studies showed from good sensitivity of 67.
3 - 100%
and specificity ranged from 61.
4 - 95.
There is a wide range of PPV with 48.
1 - 92.
92%, but display astounding NPV range of 91.
3 - 99.
This smartphone-based screening integrated with AI for DR might need some work to increase the detection differentiating those with no disease.
The U.
Food and Drug Administration (FDA) requires superiority cut-offs for AI algorithms used in DR screening to demonstrate a minimum sensitivity of 85% and specificity of 82.
34 Most studies reviewed shown that AI algorithms can meet or exceed these accuracy requirements.
In a study by Jain et al.
, the AI system demonstrated 100% sensitivity for detecting RDR, but only 89.
55% specificity.
This resulted from 136 false positives .
5%), including 55 misdiagnoses of mild NPDR and other non-DR conditions such as glaucoma, retinitis pigmentosa, ageAcrelated macular degeneration, gliosis, macular scars and asteroid This high false-positive rate .
nd the low PPV of 33.
33%) might be attributed to the combination of the AI's extreme sensitivity and potential image quality issues due to lessexperienced operators.
While a high rate of false positives may result in more patients being referred unnecessary anxiety, it can also be beneficial by Within the study of Rajalakshmi et al.
, the EyeArt software demonstrated a high sensitivity exceeding 95% for detecting DR.
STDR, and RDR in retinal images captured using the FOP smartphone The accuracy achieved in this study is comparable to that of the Google AI algorithm, which demonstrated high sensitivity and specificity for detecting RDR in both the EYEPACS-1 .
sensitivity, 93.
4% specificit.
and Messidor-2 .
sensitivity, 93.
9% specificit.
identifying other eye conditions other than DR that require medical attention.
Overall, the intergrader agreement .
uadratic weighted kapp.
reported in selected studies is moderate to high (= 0.
Sengupta, et al.
2018 compared photographic modality (Remidio FOP vs.
Topcon desktop fundus camer.
and found intergrader agreement using Remidio FOP is moderate in detecting any DR (= 0.
CI = 0.
, not too far off from using Topcon (= 0.
Lower specificity is found in study of Rajalakshmi = 0.
Both imaging methods showed et al.
2018 in detecting RDR .
8%) because AI substantial agreement in diagnosing VTDR with algorithm tends to overestimate the presence of Remidio FOP (= 0.
CI = 0.
and Topcon moderate NPDR in retinal images.
This misdiagnosis (= 0.
CI = 0.
However, the slightly frequently occurs because the AI system mistakenly higher agreement for Topcon may be partially identifies certain retinal features not associated with attributed to the misclassification of three cases of DR, such as drusen.
RPE atrophy.
RPE hypertrophy.
PDR as no DR (R.
by one grader using the Remidio telangiectatic vessels, and retinal vein occlusion, as FOP.
This error occurred despite adequate image indicators of DR.
Specificity of not mtmDR in Malerbi et al.
is also somewhat lower .
4%) than previous report, 34,36 and from further neovascularization of PDR was not captured within the limited field of view of the images.
Published by: INAVRS https://w.
org/ | International Journal of Retina https://ijretina.
In this review, the Medios AI software was This may not be feasible in countries with primarily used for retinal image analysis, working limited resources, particularly in outreach settings seamlessly with both the Remidio FOP and Remidio where access to trained personnel, reliable internet.
NM-FOP applications installed on the smartphone and expensive equipment may be limited.
12,21,26
used for capturing retinal images.
By utilizing the CoreML OpenGL capabilities, image processing occurred directly on the device's graphics processing unit (GPU), eliminating the need for an internet connection to a The AI algorithm was run offline by the technician on the smartphone itself after image The technician was trained to retake images if the AI indicated poor quality.
The AI initially assessed image quality, then provided a binary output such as DR present or No DR.
17 Furthermore.
AL software enhances the captured images with visual map highlighting potential lesions on the retinal images, assisting healthcare providers in their assessments and educating patients about potential issues of their eyes.
13 Even though Remidio FOP only utilized 4 FOV compared to previous study with 7 FOV using Digital Fundus Camera (Zeiss FF450 Plu.
37, it produced high sensitivity of 92.
7% .
8Ae .
in grading Any DR, 87.
9% .
2Ae92.
for stDR with high specificity of 98.
4% .
3Ae99.
for any DR 9% .
7Ae98.
for stDR.
Tyson, et al.
202023 used smartphone with Retinascape31 conjugated with EyeArtA .
system30 has relatively lower score of intergrader agreement in detecting any DR (= 0.
45 A 0.
compared to study by Rajalakshmi, et al.
201821 that used the same AI software (= 0.
7 - 0.
The lower specificity reported by Tyson, et al.
202023 for detecting RDR of 78.
6% .
8 - 94.
3%) per patient 5% .
7 - 86.
9%) per eye is because the gold-standard examination by a retina specialist.
Another reason is because EyeArt AI system was trained from a dataset of conventional retinal pictures, which may limit its ability to identify certain pathologies in smartphone Incorporating smartphone images into the training data could improve the algorithm's Tyson, et al.
202023 emphasizes that studies that validate new screening modalities by comparing clinician grading of mobile device images to traditional images may have erroneously high sensitivity due to the assumption that both methods are equivalent while recent research has EyeArt's high accuracy was also reflected in the shown that there is a notable difference between the UK's National DESP program, where it screened two methods.
38Ae40 Tyson also states that it is critical 30,000 patients across three regions.
EyeArt showed for researchers to rely dilated examination as the 7% .
8 - 96.
5%) sensitivity with 95% CI for gold standard to make clinical diagnosis when detecting sight-threatening retinopathy, though its validating the sensitivity and specificity of new specificity was lower of 68% .
-69%) for no diagnostic tools, particularly when using them in .
combined with non-referable retinopathy.
However, combined with its high sensitivity.
EyeArt still provided significant cost savings for the NHS.
3 The large number of people screened also marked the reproducibility of EyeArt in greater population.
While the results are promising, it is worth noting that these studies were carried out under ideal conditions where experienced professionals used desktop fundus cameras and had stable internet Rajalakshmi, et al.
stated that retinal photography, evaluated and interpreted by eye doctors specializing in the retina or other trained professionals, is a widely acknowledged and 37,41,42 However, there is shortage of trained professionals to evaluate retinal images in countries like India.
If even available, busy schedules Published by: INAVRS https://w.
org/ | International Journal of Retina https://ijretina.
have resulted in delays in providing DR grading and all 248 patients despite having poor quality images.
These delays can result in whereas EyeArt is lower with 63% .
of miscommunication, loss of follow-up, and ultimately One of the reasons for the high ungradable hinder the timely management of sight-threatening rate by EyeArt is that the camera operators were not DR.
That is along the current trend of tele- ophthalmology and telemedicine, use of AI software to evaluate retinal images for automated DR grading as it has potential to reduce the workload and costs for healthcare providers in screening the increasing number of individuals with diabetes,45 resulting only those who have stDR and RDR ophthalmologist or retina specialist.
trained and unfamiliar with the specific image and patient criteria required by EyeArt.
In 2022.
Malerbi et al.
made a study that used the Eyer camera from Phelcom Technologies to capture fundus images.
These images were then analyzed remotely using the EyerCloud platform with a deep learning-enhanced method called PhelcomNet.
This method assigned a prediction score .
between 0 and 1, indicating the likelihood of diabetic Along the trend of using artificial intelligence retinopathy (DR).
Notably, the device grouped "no DR" and "mild DR" together, while all other DR (AI) that accurately detect and grade DR in digital fundus images 13,21,34,46 combined with the feasibility severities were considered "more than mild DR" in of DR detection using smartphone-based fundus comparison to the reference standard.
27,33 Despite photography.
Wroblewski, et al.
202317 from Mexico having high sensitivity of 97.
8% and NPV 98.
7% for made a noninterventional, retrospective analysis to mtmDR, it produced lower specificity of 61.
4% and compare the diagnostic accuracy of the offline PPV of 48.
The team for obtaining images consist Medios AI software and online EyeArt AI software in of a mix of trained professionals and inexperienced detecting DR.
The analysis used a single set of patient images taken with the Remidio-FOP camera individuals in six hours.
Since image quality from in a field setting.
Medios AI have a sensitivity of 94% portable devices depends heavily on the operator and specificity of 94% for detecting any DR when skill,49 this combination of factors may have affected including poor quality images, and 99% and 88% the quality of the images obtained in the study.
when excluding poor quality images, respectively.
These results are comparable with previous study from India.
13,14 For EyeArt AI analysis, the sensitivities and specificities are 94% and 86% with all gradable images and 95% and 88% after excluding poor quality images.
These results are similar to.
from previous studies detecting for any DR and stDR conducted in India,21 English,3 and America.
Surprisingly, the two AIs achieved similar levels of sensitivity and specificity, despite being trained using different methods.
15,47 International DR data set where Medios were trained from patients from Indian lineage22,26,46 while American and Northern Mexican lineage were used for EyeArt.
15,48 It is The ease of use and portability of smartphonebased retinal cameras have been demonstrated by the successful acquisition of gradable images by a variety of operators, including healthcare workers 12,13 trained technicians,14,22,26 medical interns and students,17,23,27 with minimal to no experience in ophthalmic examination with reliable result of sensitivity and specificity.
This highlights the accessibility and user-friendliness of smartphonebased retinal imaging compared to traditional tabletop fundus cameras, which are expensive and require specialized training.
Dilated imperative to note that from the same image sets ophthalmologist, the main screening method for DR using Remidio FOP.
Medios AI was able to evaluate screening, timeAcconsuming as the patients need to Published by: INAVRS https://w.
org/ | International Journal of Retina https://ijretina.
wait till their pupil is dilated, and accessibility to In 2023.
Lestari and colleagues conducted a study pharmacologic agents and ophthalmologist are not to evaluate the knowledge, attitudes, and practices readily available.
The gold standard method defined of general practitioners (GP) in Jakarta regarding the by the ETDRS group to screen photographically for screening of diabetic retinopathy in primary care DR is by stereoscopic color fundus photographs in settings and revealed a discrepancy between GPs' seven standard fields, and this presents challenges theoretical understanding and positive attitudes in terms of time, accessibility, and cost.
It requires a skilled photographer and special costly equipment, implementation of the practice, which was found to and is also timeAcconsuming to the patient, limiting be poor.
Most GPs referred patients for ophthalmic expanded use, especially in resource-constrained without attempting it themselves, believing DR 51,52 screening was not their responsibility.
Limited While most study use pupil dilation for part of the screening process.
Sosale et al .
22,26 and Jain et al .
12 opted for non-dilated using Remidio NM FOP.
The sensitivity and specificity using Remidio NM FOP is considerably high, even exceeding mandated superiority cut-offs by FDA.
These results are comparable with a study of Alfejri .
52 that compare non mydriatic fundus camera with optical experience, lack of confidence in diagnosing fundus abnormalities, and lack of equipment in primary care also hinder GPs from conducting DR screening.
8 The author hoped with emergence of AI-integrated screening process combined with the convenience of smartphone to capture retinal images will expand healthcare coverage, particularly for preventive coherence tomography (OCT) in screening for DR.
The strength of this study is that we used This study shows 75.
2% .
3Ac80.
sensitivity, 96.
comprehensive search methodology that was .
8Ac96.
specificity, 75.
8% .
9Ac81.
of PPV and specified beforehand in our studyAos design and gave 8% .
7Ac96.
NPV from a total 2406 patients in comparative details about different settings and Riyadh.
Saudi Arabia.
This confirmed that using non methods used in selected studies that provide mydriatic DR screening, both from smartphone or options if it were to be applied to a health care fundus camera can produce high sensitivity and However, limitations are noted such as time-effective difference reference methods, difference definition screening process and eliminating pharmacologic of DR grading such as RDR, vtDR, and mtmDR.
availability barrier in performing examination.
Secondly, we did not specify whether the AI software Indonesia, the country where the author came from, is the world's largest archipelagic country with a population of over 280 million, making it the fourth most populous country in the world.
population-based study in Indonesia revealed that 1% of adults with type 2 diabetes in both urban and rural areas had DR, with 26.
3% having the more severe VTDR.
45 However.
Indonesian healthcare services still heavily focus on curative efforts rather than promotive and preventive services,54 resulting used in smartphone retinal imaging was compared to actual retinal examination by ophthalmologist or assessing the digital retinal image obtained from tabletop fundus camera.
However, in DR screening, it is important use a system that can be implemented existing workflow especially in primary population of underserved diabetic patients to get limited study to focus on preventive strategy, such as DR screening.
Published by: INAVRS https://w.
org/ | International Journal of Retina https://ijretina.
CONCLUSION
& Alhumidan A.
Barriers for adherence to The studies reviewed in this paper collectively diabetic retinopathy screening among Saudi represents the potential of smartphone-based Cureus Vol.
11:6454, .
integrated with AI in revolutionizing DR screening.
The high sensitivity and specificity achieved by Alwazae M.
Al Adel F.
Alhumud A.
Almutairi A Eppley SE.
Mansberger SL.
Ramanathan S & various AI algorithms, often exceeding the standards Lowry EA.
Characteristics associated with set by regulatory bodies like the FDA and ETDRS, highlight their accuracy in detecting DR and its severity levels.
The accessibility and user-friendliness diagnosed diabetes.
Ophthalmology.
Vols.
126, 1492Ae1499 .
smartphone-based enhance DR screening coverage, particularly in underserved areas with limited resources and Yeni Dwi Lestari.
Gitalisa Andayani Adriono.
Rizka Ratmilia.
Christy Magdalena & Ratna internet connectivity.
Sitompul.
Knowledge, attitude, and practice Funding and endorsement None.
screening among general practitioners in Conflicts of interest of Indonesia.
BMC Prim Care 24, 114 .
primary health centres in Jakarta, the capital This review had no conflicts of interest.
Sicong Li.
Ruiwei Zhao & Haidong Zou.
REFERENCES