Applications of Artificial Intelligence for the Detection, Management, and Treatment of Diabetic Retinopathy.

Abstract

Rates of diabetic retinopathy (DR) and diabetic macular edema (DME), a common ocular complication of diabetes mellitus, are increasing worldwide. There is a substantial burden concerning the detection and management of this condition, particularly in low-resource settings, due to limitations such as the time, cost, and labor associated with current screening and treatment methods. Artificial intelligence (AI) is a modality of pattern recognition that has the potential to combat these limitations in a reliable and cost-effective way. This review explores the various applications of AI on the screening, management, and treatment of DR and DME. AI applications for detecting referable DR and DME have been the most thoroughly researched applications for this condition. While some studies exist using AI to stratify DR patients based on the risk of progression, predict treatment outcomes to anti-VEGF therapy, and explore the utilization of AI for clinical trials to develop new treatments for DR, further validation studies on larger datasets are warranted.

Introduction

The worldwide prevalence of diabetes mellitus (DM) continues to increase rapidly and is expected to reach over 592 million people by 2035.^{1, 2} The most common microvascular complication of DM is diabetic retinopathy (DR), the leading cause of new cases of blindness in adults aged 17–34 in the world.³ According to a meta-analysis performed on 35 studies between 1998 and 2008, the prevalence of DR and vision-threatening DR was 34.6% and 10.2% respectively.⁴ In this same study, the prevalence of proliferative diabetic retinopathy (PDR) was 7.0% and the presence of diabetic macular edema (DME) was 6.81%.⁴ Interventions exist to effectively halt or reverse retinopathy in many cases and thus the key to preventing vision loss is expedient recognition and treatment of retinopathy. However, given the high prevalence of DM worldwide, ensuring that all patients are screened and have effective follow-up adds a significant burden on healthcare systems. This paper will explore how artificial intelligence (AI) can minimize the burden of DR by addressing issues with detection, monitoring progression, and predicting treatment outcomes for DR and DME.

In computer science, machine learning is the study of using machines to make intelligent decisions by extracting patterns and learning from data. A branch in machine learning called deep learning is largely inspired by our knowledge of the human brain. Specifically, deep learning consists of artificial neural networks consisting of a relatively large number of layers or “neurons,” hence the term “deep.”⁵ Deep learning has performed remarkably well in image processing tasks and has outperformed other computer vision methods including in medical imaging, with reported accuracy similar to that of expert humans in some cases.^{6, 7} Figure 1 describes the traditional method of staging diabetic retinopathy whereas Figure 2 describes a deep learning model.

Figure 1 –

Image Classification by Trained Personnel: This figure shows the process of having trained personnel use their experience and traditional algorithms to determine which stage of DR a fundus image is.

Figure 2 –

Deep Learning: In deep learning, photos of DR and healthy images are initially introduced to the machine. Through complex mathematical equations resembling neural networks, the machine identifies features from the input data. The model creates a map using these features so that it can determine the proper DR staging when presented with an unknown fundus image using what it learned during the training phase.

1. Diabetic Retinopathy Detection and Screening

Screening for DR is important for early detection and prevention of vision loss. Nationwide screening programs identify referable DR (characterized by moderate or severe diabetic retinopathy or the presence of diabetic macular edema) and refer them to ophthalmology clinics where patients can receive a full ophthalmic examination. Estimates predict that over $500 million could be saved in healthcare related costs if DR is detected earlier in the United States (U.S.) alone.⁸ Traditionally, the gold standard for DR screening is considered a funduscopic examination by a trained ophthalmologist.⁹ However, due to the rapidly increasing rates of diabetes and the costs associated with screening (monetary, time, labor), ensuring all diabetic patients receive the required screening and treatment proves to be a challenge.

Telemedicine has become a popular method for diagnosing DR during screening by taking color fundus photographs and sending them to ophthalmologists to grade the images remotely and has been shown to improve screening rates, treatment rates, and reduce vision loss due to diabetes.^{10, 11} Even with imaging using only telemedicine, it is estimated that 32 million retinal images would have to be evaluated to ensure that the need for DR screening is met.¹² Recent advances in AI provide promise that an automated diagnostic tool may be developed so that screening will be accessible to people in a wide range of settings.

Using AI to screen for “referable diabetic retinopathy” is one of the most advanced applications of AI within ophthalmology. Many researchers have developed algorithms that are able to “learn” to diagnose DR and tested the accuracy of their programs on repositories of retinal images with high sensitivity and specificity results.^{6, 13–21} (Table 1) The FDA approved a program in April 2018 known as IDx-DR which uses pictures from a nonmydriatic retinal camera (NW400, Topcon Medical Systems, Oakland, NJ) and then uses an AI program to diagnose DR in a clinical setting.²² The imaging system on the IDX-DR is costly, and thus there are numerous studies exploring the use of similar AI programs on smartphone based fundus cameras such as the Remidio Fundus on Phone (FOP).^23–27 The use of automated AI programs on smartphones provides a relatively low cost and effective method of screening DR. Several studies have validated screening programs for diabetic retinopathy in the clinical setting, both with the use of smartphone based fundus cameras and traditional fundus imaging, in international settings including Singapore,^{28, 29} United Kingdom,³⁰ India,²⁵ Australia,³¹ and Thailand.^32–34

Table 1:

Identification of Referable DR (Moderate NPDR or Worse)

Author	Title	Year	Sensitivity	Specificity	AUC	Accuracy	Validation Data Set	Type of Image
Abramoff et al13	Improved automated detection of diabetic retinopathy on a publicly available data set through integration of deep learning	2016	96.8	87	0.98	Not Reported	1748 (Messidor 2)	Fundus
Abramoff et al14	Pivotal trial of an automated AI-based diagnostic system for detection of diabetic retinopathy in primary care offices	2018	87.2	90.7	Not Reported	Not Reported	819 patients (Acquired Images)	OCT
Bellemo et al31	Artificial intelligence using deep learning to screen for referable and vision-threatening diabetic retinopathy in Africa: a clinical validation study	2019	92.3	89	0.973	Not Reported	1574 patients (Acquired Images)	Fundus
Bhaskaranad et al15	The Value of Automated Diabetic Retinopathy Screening with the EyeArt System: A Study of More Than 100,000 Consecutive Encounters from People with Diabetes.	2019	91.3	91.1	Not Reported	91.20%	850908 images (Acquired Images)	Fundus
Gargeya et al16	Automated identification of diabetic retinopathy using deep learning	2017	94	98	0.97	Not Reported	75137 (Kaggle Images)	Fundus
Gulshan et al11	Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs	2016	87–90.3	98.1–98.5	Not Reported	Not Reported	128175 (EYEPACS-1 and Messidor-2)	Fundus
He et al17	Artificial intelligence-based screening for diabetic retinopathy at community hospital.	2019	91.2	98.8	0.95	Not Reported	889 patients (Acquired Images)	Fundus
Heydon et al32	Prospective evaluation of an artificial intelligence enabled algorithm for automated diabetic retinopathy screening of 30 000 patients	2020	95.7	54	Not Reported	89.4	30405 (Acquired Images)	Fundus
Keel et al29	Visualizing deep learning models for the detection of referable diabetic retinopathy and glaucoma.	2018	92.3	93.7	Not Reported	Not Reported	96 Patients (Acquired Images)	Fundus
Kermany et al18	Identifying medical diagnoses and treatable diseases by image-based deep learning	2018	97.8	97.4	0.99	96.60%	4686 patients	OCT
Natarajan et al24	Diagnostic accuracy of community-based diabetic retinopathy screening with an offline artificial intelligence system on a smartphone	2019	100%	81.90%	Not Reported	Not Reported	231 patients (Acquired Images)	Smartphone fundus
Rajalakshmi et al23	Automated diabetic retinopathy detection in smartphone-based fundus photography using artificial intelligence.	2018	99.3	68.8	Not Reported	Not Reported	296 patients (Acquired Images)	Smartphone Fundus
Ramachandran et al19	Diabetic retinopathy screening using deep neural network	2018	84.6/96	79.7/90	0.90/0.98	Not Reported	485 Images (Otago)/1200 Images (Messidor)	Fundus
Ruamviboonsuk et al30	Deep learning versus human graders for classifying diabetic retinopathy severity in a nationwide screening program.	2019	96.8	95.6	0.987	Not Reported	7517 patients (Acquired Images)	Fundus
Sahlsten et al43	Deep Learning Fundus Image Analysis for Diabetic Retinopathy and Macular Edema Grading.	2019	89.6	97.4	0.987	Not Reported	7304 images (Messidor)	Fundus
Takahashi et al35	Applying artificial intelligence to disease staging: Deep learning for improved staging of diabetic retinopathy	2017	Not Reported	Not Reported	0.81	96%	496 (Acquired Images)	Fundus
Ting et al36	Development and validation of a deep learning system for diabetic retinopathy and related eye diseases using retinal images from multiethnic populations with diabetes.	2017	90.5	91.6	0.936	Not Reported	71896 images (Singapore National Diabetic Retinopathy Screening Program)	Fundus
Xu K et al85	Deep Convolutional Neural Network-Based Early Automated Detection of Diabetic Retinopathy Using Fundus Image.	2017	95.8	80.2	0.945	94.50%	400 (Kaggle Images)	Fundus

2. Assessing Progression of DR and DME Using AI

Once referable diabetic retinopathy is detected, the main goal is to prevent progression to vision threatening diabetic retinopathy, which is defined as severe non-proliferative diabetic retinopathy (NPDR), PDR or the presence of DME.⁴ In order to do this, it is essential to identify the stage of diabetic retinopathy and monitor for progression. As management regimens vary by the severity of diabetic retinopathy, it is crucial to 1) assess the severity of diabetic retinopathy by staging DR and determining the presence of DME, 2) identify which patients are at high-risk for advancing to vision threatening diabetic retinopathy, and 3) establish ideal screening intervals based on a patient’s risk stratification scores.

2a. Staging DR and Evaluating for Presence of DME

Vision threatening diabetic retinopathy (VTDR), particularly PDR or the presence of DME, is commonly associated with severe vision loss and thus identifying VTDR to initiate prompt treatment (i.e. anti-VEGF and/or laser treatment) is crucial while managing a diabetic patient.³⁵ However, studies have shown that reliability among graders staging DR is only 70%.³⁶ Deep learning algorithms developed by researchers have found reliability of grading up to 95% and sensitivities for staging ranging from 90.5% to 97%.^{6, 29, 32, 37, 38} (Table 2)

Table 2:

Staging Diabetic Retinopathy

Author	Title	Year	Sensitivity	Specificity	AUC	Accuracy	Validation Data Set	Type of Image	Type of Staging
Abbas et al35	Automatic recognition of severity level for diagnosis of diabetic retinopathy using deep visual features	2017	92.2	94.5	0.924	Not Reported	750 images (Combination of DIARETB1, FAZ, Messidor, Prv-DR datasets)	Fundus	Identifies 5 Severity Levels of DR
Gulshan et al11	Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs.	2016	84.0/87.8	98.8/98.2	Not Reported	Not Reported	8788 Images (EyePACS-1)/ 1745 Images (Messidor-2)	Fundus	Identification of severe or worse DR
He et al17	Artificial intelligence-based screening for diabetic retinopathy at community hospital.	2019	80.4	99.4	0.9	Not Reported	889 patients (Acquired Images)	Fundus	Identification of PDR
Rajalakshmi et al23	Automated diabetic retinopathy detection in smartphone-based fundus photography using artificial intelligence.	2018	99.1/78.1	80.4/89.8	Not Reported	Not Reported	296 patients (Acquired Images)	Fundus	Identification of severe or worse DR
Ruamviboonsuk et al30	Deep learning versus human graders for classifying diabetic retinopathy severity in a nationwide screening program.	2019	93.6	98.2	0.991	Not Reported	7517 patients (Acquired Images)	Fundus	Identification of severe or worse DR
Ting et al36	Development and validation of a deep learning system for diabetic retinopathy and related eye diseases using retinal images from multiethnic populations with diabetes	2017	100	91.1	0.958	Not Reported	71896 images (Singapore National Diabetic Retinopathy Screening Program)	Fundus	Identification of severe or worse DR

Diabetic macular edema (DME) is a complication of DR characterized by retinal thickening in the macula due to fluid accumulation. Center-involved (ci)- DME is the primary cause of moderate vision loss for diabetic patients and is an indication for referral and treatment. The standard for identifying ci-DME was established by several major clinical trials measuring macular thickening based on optical coherence tomography (OCT).³⁹ OCT machines are more costly than fundus cameras and thus the presence of hard exudates within one optic disc diameter on fundus photographs is commonly used as a proxy for photographic screeing.⁹ However, this alternative method of determining ci-DME has been shown to have variable results.⁴⁰ Several researchers have developed a deep learning algorithm using fundus photographs to determine the presence of ci-DME, validated by OCT or fundus images, with a similar sensitivity and higher specificity than human graders.^41–44 (Table 3) Sahlsten et al developed a program that accurately graded diabetic macular edema according to a commonly used diabetic macular edema severity scale (PIMEC).⁴⁵ Lastly, similar to fundus images, conclusions drawn from OCT images are limited by interobserver variability amongst readers and are prone to error given the large number of images.⁴⁶ Several studies have presented methods to utilize deep learning to detect, differentiate, and quantify macular fluid objectively for a range of exudative macular diseases, including DME.^47–49 (Table 3)

Table 3:

Detecting Diabetic Macular Edema

Author	Title	Year	Sensitivity	Specificity	AUC	Accuracy	Validation Data Set	Type of Image	Type of Staging	Comments
Alsaih et al42	Machine learning techniques for diabetic macular edema (DME) classification on SD-OCT images.	2017	87.5	87.5	Not Reported	Not Reported	32 volumes (Singapore Eye Research Institute Database)	OCT	Detects DME	*Each volume has 128 B-Scan slices
Chan et al41	Fusing Results of Several Deep Learning Architectures for Automatic Classification of Normal and Diabetic Macular Edema in Optical Coherence Tomography.	2018	93.75	93.75	Not Reported	93.75	115 volumes (from Singapore Eye Research Institute and Chinese University of Hong Kong dataset)	OCT	Detects DME	*Each volume has 128 B-Scan slices
Gulshan et al11	Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs.	2016	91/90.4	99/98.8	Not Reported	Not Reported	8788 Images (EyePACS-1)/ 1745 Images (Messidor-2)	Fundus	Detects DME
Sahlsten et al43	Deep Learning Fundus Image Analysis for Diabetic Retinopathy and Macular Edema Grading.	2019	Not Reported	Not Reported	0.981	93.4	7304 images (Messidor)	Fundus	Identifies DME Stage based on PIMEC
Schlegl et al45	Fully Automated Detection and Quantification of Macular Fluid in OCT Using Deep Learning.	2018	Not Reported	Not Reported	0.94/0.92	Not Reported	1200 scans (Acquired Scans using Cirrus and Spectralis)	OCT	Detects and quantifies intraretinal cystoid fluid /subretinal fluid
Sun et al46	Fully automated macular pathology detection in retina optical coherence tomography images using sparse coding and dictionary learning.	2017	100/99.7	Not Reported	Not Reported	Not Reported	45 patients (Duke Spectral Domain)/678 Images (Acquired)	OCT	Detects DME	Also detects AMD
Syed et al47	Automated diagnosis of macular edema and central serous retinopathy through robust reconstruction of 3D retinal surfaces.	2016	100	96.7	Not Reported	98.9	90 Volumes (Armed Forces Institute of Opthalmology)	OCT	Detects DME	Also detects central serous retinopathy
Varadarajan et al39	Predicting optical coherence tomography-derived diabetic macular edema grades from fundus photographs using deep learning.	2020	85	80	0.89	81	1033 images (Thailand Dataset)	Fundus	Detects ci-DME

2b. Risk Stratification for Progression to VTDR

Another useful modality of AI in DR is risk stratification. Since progression can be mitigated by intensive glycemic management and, when necessary, intervention with laser photocoagulation, intraocular corticosteroids, or anti-vascular endothelial growth factor (VEGF) agents, identifying which individuals are most likely to progress to VTDR is critical.^50,51 While studies have theorized several risk factors (i.e. age, gender, HbA1c levels, blood pressure, proteinuria) and predictors (i.e. microaneurysm turnover), there are still many unknown factors contributing to an individual’s risk profile, making it difficult for ophthalmologists to know which individuals are more likely to progress quickly.^52–58 Several studies have used AI to help in this task. (Table 4) Arcadu et al used deep learning to predict DR progression, defined as 2-step worsening according to the Early Treatment Diabetic Retinopathy Study severity scale, using a 7-field color fundus photograph taken during the initial visit. They compared their predictions to DR severity scores assessed by masked, well-trained human reading center graders at various time points. The deep learning model accurately predicted progression of DR at 6,12, and 24 months with an area under the curve of 0.68, 0.79, and 0.77 respectively.⁵⁹ Similarly, several studies have utilized retinal vasculature to predict progression from early NPDR to PDR with AUCs as high as 0.968 using deep learning.^60–62 Another study by Bora et al determined the risk of diabetic patients with no DR developing DR within two years using deep learning with an AUC up to 0.71.⁶³ Identifying high-risk patients can help determine which patients need more intensive treatment, support clinical decision-making, and potentially aid in recruitment for clinical trials. While further studies are necessary on larger data sets to further these results, these studies provide promise that an algorithm to identify higher risk patients to determine an ideal management schedule is possible.

Table 4:

Progression of Diabetic Retinopathy or Diabetic Macular Edema

Author	Title	Year	Sensitivity	Specificity	AUC	Accuracy	Validation Data Set	Type of Image	Type of Staging	Comments
Arcadu et al40	Deep Learning Predicts OCT Measures of Diabetic Macular Thickening From Color Fundus Photographs.	2019	87.50%	96.50%	0.97 (Central Subfield Thickness) and 0.91 (Central Foveal Thickness)	Not Reported	17997 Images (RIDE/RISE DME Database)	Fundus	Predicts OCT Measures of Diabetic Macular Thickening based on Fundus Photographs	Predicts central subfield thickness >250 um and Central foveal thickness >250 um
Arcadu et al57	Deep learning algorithm predicts diabetic retinopathy progression in individual patients.	2019	66%, 91%, 79%	77%, 65%, 72%	0.68,0.79,0.77		529, 528, 499 patients (Acquired Images)	Fundus	Predicts Progression of DR at 6, 12, 24 months	DR progression defined as 2-step worsening on the EDTRS scale
Bora et al61	Predicting Risk of Developing Diabetic Retinopathy using Deep Learning	2020	Not Reported	Not Reported	0.71	Not Reported	4762 images (EyePACS and Thailand Dataset)	Fundus	Predicts risk of Progression to DR
Leontidis et al58	A new unified framework for the early detection of the progression to diabetic retinopathy from fundus images.	2017	Not Reported	Not Reported	0.821–0.968	Not Reported	127 images	Fundus	Progression from no DR to DR	Identified retinal vascular geometry as a feature that predicts progression from no DR to DR
Orlando et al60	Proliferative diabetic retinopathy characterization based on fractal features: Evaluation on a publicly available dataset.	2017	Not Reported	Not Reported	0.093	Not Reported	1200 images (Messidor dataset)	Fundus	Progression to PDR on patients with no NPDR or mild NPDR	Identified higher fractal dimension has a predictor for PDR
Welikala et al59	Automated detection of proliferative diabetic retinopathy using a modified line operator and dual classification	2017	86.2	94.4	0.9632	Not Reported	60 images (Messidor and Acquired)	Fundus	Identifies progression to PDR	Detects PDR via detecting new blood vessels

2c. Establishing Optimal Screening Intervals Based on Risk Profile

Determination of high-risk patients can also help establish optimal screening intervals for these patients, which would improve cost-effectiveness of screening. Several studies have established the cost-effectiveness of population-based screening programs that have been optimized for individuals based on their risk assessment.^64–66 AI can be used to help determine the risk for progression of patients with DR as well as define the ideal screening frequency based on this risk profile. For example, in order to account for a variety of factors such as risk profile, resource availability, an individual’s lifestyle, and genetics, the current guidelines for DR screening intervals are reported in ranges (i.e. 3–6 months for moderate NPDR). By using AI to develop a more individualized screening interval, low-risk patients can alleviate the burden on the health system by spacing out screening intervals whereas high-risk patients can have more frequent follow-up to ensure adequate detection and treatment. Furthermore, just as recommendations for screening intervals vary based on the resource setting,⁶⁷ deep learning algorithms can be tested on a variety of different situations so that referral guidelines can be optimized based on a given set of parameters. Eleuteri et al used machine learning to develop a generalizable risk calculation engine to assign individualized screening intervals for patients with diabetic retinopathy based on risk.⁶⁸ Once more algorithms have been validated to accurately predict progression of DR, further research investigating optimal screening intervals using deep learning models integrating various risk factors and resource settings for diabetic retinopathy are warranted.

3. Predicting Treatment Outcomes with anti-VEGF for DR and DME with AI

After stratifying patients based on their risk of progression to VTDR, the next step will be to determine which patients will benefit the most from certain treatments. Management for patients vary greatly based on the stage of DR, the presence of DME, and their risk profile. Better understanding of which patients will benefit the most from different treatment options can help ophthalmologists develop targeted plans to improve treatment outcomes while minimizing complications.⁶⁹

For example, anti-VEGF agents such as aflibercept, bevacizumab, and ranibizumab are the main modalities of treatment for ci-DME. Although they have been shown to generally improve visual acuity, the therapeutic response to anti-VEGF treatment in these patients varies greatly due to the complicated pathogenesis of diabetic retinopathy.⁷⁰ Some patients, despite frequent anti-VEGF injections, continue to exhibit signs of disease activity (incomplete responders) and may benefit from earlier intravitreal corticosteroid treatment, however, reliable features for identifying these patients remain elusive. Identifying such factors that distinguish one type of patient from the other, including differences in phenotype and genotype, to predict response would be extremely beneficial considering the high-cost burden associated with anti-VEGF injections.⁷¹ AI has the potential to help determine these factors, providing the potential to select an individualized, more effective first-line therapy.

Several studies have attempted to develop deep learning algorithms to predict early anatomical and functional factors associated with favorable outcomes with anti-VEGF injections.^72–78 (Table 5) Takahashi et al used neural networks to predict which treatments would be most beneficial for patients with 96% accuracy but had a false negative rate of 12%.⁷³ Two studies used pre-treatment OCT scans to develop a deep learning algorithm that was able to predict responsiveness to anti-VEGF therapy for up to 95% of patients.^{74, 76} Other studies have identified specific anatomical features, such as presence of baseline subretinal fluid or intraretinal cystic fluid in the outer nuclear layer, that are associated with better outcomes after anti-VEGF treatment for patients with diabetic macular edema using deep learning techniques.^{77, 78} These studies highlight the potential for artificial intelligence to aid the clinical decision-making process in selecting treatment for patients with VTDR.

Table 5:

Determining Treatment Outcomes based on anti-VEGF Injections

Author	Title	Year	Sensitivity	Specificity	AUC	Accuracy	Validation Data Set	Type of Image	Type of Staging	Tx of DME/DR/Nonspecific?
Arcadu et al70	Deep learning algorithm for patient-level prediction of diabetic retinopathy (DR) response to vascular endothelial growth factor (VEGF) inhibition.	2019	87/81/84	72/82/82	.82/,85/.87	Not Reported	759 patients (RISE/RIDE data)	Fundus	Predicts >2 step improvement on Diabetic Retinopathy Severity Scale in response to ranibizumab treatment for 6 mo/12 mo/24 mo	DR
Chen et al73	A Novel Machine Learning Algorithm to Automatically Predict Visual Outcomes in Intravitreal Ranibizumab-Treated Patients with Diabetic Macular Edema.	2018	Not Reported	Not Reported	Not Reported	Not Reported	464 Eyes (DRCR.net database)	Patient data (sex, age, diabetes type or condition, systemic diseases, eye status, treatment time tables)	Reported correlation coefficients (0.7 at 52 weeks, 0.55 at 78 weeks, and 0.81 at 104 weeks)	DME
Gerendas et al75	Computational image analysis for prognosis determination in DME.	2017	Not Reported	Not Reported	Not Reported	R^2=0.21/0.23	629 patients (Acquired Images)	OCT	Predicts prognosis of DME patients undergoing anti-VEGF therapy	DME
Prahs et al72	OCT-based deep learning algorithm for the evaluation of treatment indication with anti-vascular endothelial growth factor medications	2017	90.1	96.2	0.968	95.5	5358 images (Acquired Images)	OCT	Classifies images that have no injection vs injection	Nonspecific
Rasti et al74	Deep learning-based single-shot prediction of differential effects of anti-VEGF treatment in patients with diabetic macular edema.	2020	80.1	85	0.866	Not Reported	127 patients (Acquired Images)	OCT	Response to VEGF was defined as at least 10% reduction in retinal thickness following treatment	DME
Roberts et al76	Quantification of Fluid Resolution and Visual Acuity Gain in Patients With Diabetic Macular Edema Using Deep Learning: A Post Hoc Analysis of a Randomized Clinical Trial.	2020	Not Reported	Not Reported	Not Reported	Not Reported	570 patients (Acquired Images)	OCT	Used an automated algorithm to quantify intraretinal and subretinal fluid to assess treatment outcomes with anti-VEGF treatment	DME
Takahashi et al71	Applying artificial intelligence to disease staging: Deep learning for improved staging of diabetic retinopathy.	2017	Not Reported	Not Reported	Not Reported	96	4,709 images (Acquired Images)	Fundus	Suggests treatments for DR based on staging	DR

4. Potential Impact on Patients and Clinical Workflow

With effective screening methods, artificial intelligence will decrease the potential need for less urgent ophthalmology office visits by identifying patients with referable DR. In addition to decreasing the number of appointments, this benefits the patient economically with reduced transportation and co-pay costs. Diabetic patients are particularly susceptible to loss to follow-up; the poorer a patient’s glycemic control, the more likely he or she is to miss appointments.⁷⁹ Minimizing the number of appointments will help increase rates of adherence as part of the multi-dimensional management of diabetes.⁸⁰ Tufail et al conducted a detailed analysis on the cost-effectiveness of automated DR screening and concluded that automated screening of DR was significantly cost-effective, particularly in rural settings.³⁰ Another factor to consider is the time it takes to schedule a follow-up appointment with a specialty service. One study found that the average wait time to schedule a follow-up appointment for patients in six high-volume clinics referred to an ophthalmologist from a primary care physician for DR was seven months.⁸¹ This delay in evaluation and treatment may lead to worse outcomes and higher costs due to managing DR at a later stage.

In addition to helping manage screening, diagnosis and treatment of DR clinically, AI can also help the clinical workflow. Through more effective screening, AI can also reduce the load on ophthalmologists’ offices, allowing for better utilization of resources for higher acuity patients. With the current projection of 592 million people having diabetes by 2035, around 3 million eyes will have to be evaluated every year, which would overwhelm the capacity of most healthcare institutions.⁸² This burden would be significantly alleviated if only moderate/severe DR patients had to be referred to an ophthalmologist.

Furthermore, the amount of time it takes to identify diabetic retinopathy is significantly reduced with AI; the median time required to manually stage DR in retinal images with seven fields was 12.8 min per patient whereas staging level for DR in Abbas et al’s study only took 2.12 seconds. ³⁷ AI/machine learning can also help clinicians quantify retinal vasculature using fluorescein angiography (FA), a task which is very time consuming and prone to errors due to contrast.⁸³ A study using this algorithm found it only took a mean of 22.1 seconds to calculate macular retinal vessel density from FA images and found it correlated with baseline visual acuity in patients with PDR.⁸⁴

Potential Impacts on Clinical Trials

Lastly, AI can help randomized control trials (RCTs) study new, more personalized therapies for diabetic retinopathy. Considering the complex pathogenesis of diabetic retinopathy and the varied clinical presentations that complicate the diagnosis, targeting specific therapies for subsets of patients may result in more individualized, effective therapies. For example, Roberts et al used deep learning to determine aflibercept was superior in specifically reducing baseline subretinal fluid, which serves as a potential explanation as to why aflibercept has been shown to be particularly beneficial in DME patients.⁷⁸ However, performing randomized controlled trials as therapies become more individualized becomes increasingly difficult as there is a smaller subset of individuals that are eligible for the study. AI can help identify certain subsets of individuals that may be eligible for the therapy, reducing the burden of having to screen a large number of individuals and making it easier to perform more focused trials and develop novel therapies.

Another benefit of utilizing AI for clinical trials is the standardization of outcome. When performing RCTs on novel drugs for diabetic retinopathy, the endpoint (i.e. severity of diabetic retinopathy or resolution of hemorrhage) is often determined by interpreting either a fundus image (i.e. decrease in exudates or microaneurysms) or an OCT image (i.e. decrease in intraretinal fluid). However, as mentioned earlier in this review, these endpoints are often subject to interobserver variability. AI has been shown to be able to determine these endpoints with higher sensitivities and specificities than human graders with minimal variability between readings. Lastly, given AI’s ability to accurately predict the outcome of a disease with and without established interventions,⁷⁴ future developments in AI may be able to predict natural progression of diabetic retinopathy to the extent that no control arm is necessary. This would further reduce the number of subjects required to perform RCTs.⁸⁵ Additionally, one of the biggest deterrents for patients joining an RCT is the prospect of randomization to a placebo,⁸⁶ and thus having artificial intelligence model the placebo arm may increase participation in these trials. For example, if a randomized control trial was being conducted on anti-VEGF therapy for patients at high risk for progressing to PDR, deep learning could be used to model a control arm to model the natural progression of diabetic retinopathy in this patient population without anti-VEGF treatment. This may also address an ethical issue encountered in RCTs where an intervention is clearly shown to be better for a certain patient subgroup. Though further research needs to be performed to validate that the modeled control arm acts similarly to the patient population subset, there is promise that AI can serve as an adjunct to RCTs to more effectively manage DR.

Conclusion

Using AI, there is potential for exponential gain using automated programs to screen, monitor and manage DR leading to reliable and cost-effective screening protocols. This will lead to an effective utilization of resources, help alleviate the growing burden of managing these patients, and improve treatment outcomes especially in light of the rising incidence of diabetes. While there are many technologies that hold promise to fundamentally change the landscape of DR management, further validation studies on larger datasets need to be performed, particularly to help with risk stratification and treatment prediction outcomes for DR. Similarly, more studies are necessary to evaluate how AI can help augment randomized clinical trials to further develop novel therapeutics for this condition. Another issue to be addressed in the clinical use of AI is interpretability since the basis for recommendations, in particular when using deep learning methods, may not be apparent the clinician using an AI system. Developments in explainable AI will lead to greater trust and utilization of AI and allow for clinicians to better judge when a recommendation by an AI system may not be appropriate for a particular patient. Ultimately, the hope is to integrate AI into DR screening programs resulting in an optimization of patients referred to specialists, timely diagnoses, and early treatment to preserve vision.