Key points are not available for this paper at this time.
With aging populations, health systems worldwide are struggling to provide adequate eye care at the population level, giving rise to projections of increasing levels of visual impairment (VI) and blindness from major eye diseases in the near future.1 The substantial burden of VI has overwhelmed global efforts to expand the physical capacity (eye clinics and hospitals and related facilities) and expert availability of eye care providers (eg, ophthalmologists, optometrists).2 This trend has been observed even in developed nations. Long waiting times to see an ophthalmologist in public hospitals are common in the United States, UK, Australia, Singapore, and Hong Kong. A study in England reported that a delay in care of 22 weeks resulted in harms such as permanent visual acuity deterioration and visual field loss in some patients that could have been avoided with earlier intervention.3 This reflects the urgent public health need for completely novel solutions to improve the accessibility and availability of eye care services at primary, secondary, and tertiary level for the larger population.4 Of the major eye diseases, diabetic retinopathy (DR), age-related macular degeneration (AMD), glaucoma, and cataract are major global causes of blindness. DR has alarming projections for prevalence despite established cost-effectiveness and guidelines for population screening.5–7 This is partly due to staggering growth in the burden of diabetes mellitus (DM),8 whereby DR is a common end organ manifestation of DM.9 Similarly, AMD and glaucoma are other leading causes of VI, driven by aging populations.10 Gaps in population-level screening and management of DR, AMD, and glaucoma are well known and have not been adequately tackled using traditional models of care, which have largely focused on provision of tertiary level eye care.11–13 DAWN OF DIGITAL HEALTH AND ARTIFICIAL INTELLIGENCE Technology may provide innovative solutions to address such gaps.14 The successful development of digital health and telemedicine solutions have provided new tools that improve the efficiency and accessibility of existing eye care services.4 Recent reports describe new models of care delivery enabled by technology, such as asynchronous teleophthalmology in England and Singapore.15,16 These enabled more targeted referrals from community-based primary eye care services to specialist ophthalmologists, enhancing right siting of patients.17 Successful use of these technologies in Scotland has been described in the Scottish Eyecare Integration project through electronic image exchange between community providers and clinicians of a centralized referral unit.18 Reported benefits include streamlined referral services, improved patient satisfaction, and generation of real-world data for provider training.19 Similarly, encouraging results have been described of digital ophthalmology services for patients with diabetes attending a specialist endocrinology outpatient clinic in Italy and rural remote eye screening services in suburban and rural Australia, whereby other common causes of blindness such as AMD and glaucoma were also incidentally detected early.20,21 Over in India, teleophthalmology has also been used to enhance access to eye screening and remote specialist review in rural settings using synchronous22 or mixed synchronous-asynchronous solutions,23 whereby the latter uses a mix of store-and-forward and/or video consultation.4 However, although these digital ophthalmology solutions have demonstrated improved clinical outcomes, their scalability is constrained by the need for additional infrastructure and manpower to operationalize them.4 The advent of artificial intelligence (AI) and its applications in ophthalmology promise to extend and expand the use of digital ophthalmology, providing the potential to improve the accessibility, availability, and productivity of existing resources and overall efficiency of eye care services.24 Telemedicine and AI are not constrained by the “where” (need for specialized eye hospitals) or who (availability of expert eye care manpower). Therefore, these solutions can be paired with other technology (eg, imaging devices) to quickly decentralize and scale-up eye screening and primary care services.24 Where relevant resources are available, this presents a timely solution as procurement cycles and the time required to train operators of these solutions are significantly less than that required to train ophthalmologists. Eye screening is well-suited for the application of AI due to well-understood relationships between clinical features and disease severity in the major eye diseases such as glaucoma25 and DR,26 and ready availability of imaging to effectively capture these features.27 IS AI IN OPHTHALMOLOGY RESEARCH READY FOR CLINICAL TRANSLATION AND IMPLEMENTATION? AI is a technology that was first described >50 years ago. In the past 5 to 10 years, with improving computing power and large datasets, deep learning (DL), a new branch of AI, has been developed. Research using newer AI techniques, including DL, has shown robust results that has surpassed human performance in many areas of medicine and healthcare. AI has possibly reached a “tipping point” in technology readiness level for translation to real-world application.24 Applications of AI in ophthalmology began with traditional feature-based machine learning (ML) techniques as described by Ruamviboonsuk et al45 in their article about the evolution of AI solutions for eye screening in this issue. Those ML techniques required expert ophthalmologists to label individual clinical features and severity in images to develop the AI solutions, sometimes referred to as "supervised learning."28 These labeled images were used to train AI algorithms to classify the images based on identification of those feature labels. DL is a new AI technique that is trained in a manner that bypasses the need for feature labeling by experts, sometimes referred to as “unsupervised learning.” DL instead involves using entire images labeled with clinical diagnosis (or severity) by experts, so the AI “self-learns” the predictive features for classification of diagnosis or severity, with better than traditionally accepted error rates. Recent reports demonstrated improved classification of ophthalmic imaging using DL compared with ML techniques,29 with the same in head-to-head comparisons.30 Recent research has demonstrated clinically acceptable performance of these DL algorithms in classifying ophthalmic imaging data such as color fundus photography (CFP) for various eye diseases such as DR.29 Other successful applications include that for classifying optical coherence tomography (OCT) scans.31 By providing granular 3-dimensional information as opposed to 2-dimensional information from CFPs, the incorporation of OCT imaging may improve the performance of existing CFP-based screening tools. For the major eye diseases of DR, AMD, and glaucoma, fortunately, retinal features that correspond to clinical outcomes and severity have been described,25,32 enabling the training of DL to classify ophthalmic imaging.33,34 DL systems have been developed for classifying DR in CFPs35–37 and OCT images,38 which have been summarized along with other applications in a recent review.29 Furthermore, DL systems that classify AMD in CFPs33,39,40 predict AMD progression using OCTs,41 and classify glaucoma in ophthalmic imaging have also been described,42,43 with some demonstrated in real-world datasets from screening settings such as teleophthalmology and glaucoma clinics.44 Other new DL studies and initiatives in Asia have also been described in the article on the evolution of AI-based solutions for eye screening contributed by Ruamviboonsuk et al45 in this issue. The authors suggest methods to reduce the cost of digital ophthalmology and/or AI-based ophthalmology services, such as the training of human graders to augment the digital services and reduce the need for specialist-trained ophthalmologists to support the services, freeing up ophthalmologists for more follow-up of patients requiring further assessment or interventional services. Independent researchers in Singapore and Korea have even applied a single deep learning system to simultaneously detect multiple referable eye diseases such as glaucoma, AMD and DR with high accuracy.36,46 In another completely different area of cataract detection and management, Goh et al47 discussed techniques for emerging applications of AI related to cataract, a relatively underexplored area in ophthalmology. These range from automating the selection of intraocular lenses during preoperative planning for cataract surgery, to detecting cataract in ophthalmic imaging such as CFP or slit lamp photography. HOW CAN AI MAKE AN IMPACT IN OPHTHALMOLOGY DIAGNOSIS AND SCREENING? Despite these research reports of excellent DL performance in clinical validation studies, few studies have evaluated suitable forms for its real-world implementation. This can either take on a “fully-automated” or “semi-automated” model (Fig. 1). The “fully-automated model” would function with no human provider involvement, whereby the AI system initiates referrals to ophthalmologists where necessary or flags patients as suitable for continued community-based monitoring. In contrast, a “semi-automated model” can take on various forms of human grader or ophthalmologist involvement to augment classification by the DL as a tool for triage of patients.FIGURE 1: Potential applications of a DL solution for DR screening. Illustration of the application of a DL solution for DR screening using imaging, comparing existing clinical practice with a fully automated AI model (replacement) and a semiautomated AI model (triage). AI indicates artificial intelligence; DL, deep learning; DR, diabetic retinopathy.A clear example of where an AI-based DL technology can potentially “fill” a growing gap is in DR screening. The increasing burden of DR has outpaced the long process of training eye care professionals, causing an expanding rift between health system capacity in screening and clinical demand.48 The traditional model for DR screening, and 2 possible models of AI-based tools for automated classification of DR are shown in Figure 1. Ideally, in some situations, AI based-DL algorithms can be implemented using a “fully-automated model” (as a replacement test), whereby classification is independently conducted by the AI with triage and referrals to ophthalmologists where required (Fig. 1). Alternatively, AI based-DL algorithms can be implemented in a “semi-automated model,” whereby a human assessor (doctors/optometrists/readers) mediates classification of imaging data that the AI labels as abnormal. Xie et al have described the cost-effectiveness of one such semiautomated model developed in Singapore whereby the AI is used to triage CFPs and those flagged as abnormal are reviewed by human graders using teleophthalmology.49 Apart from determining the ideal form of implementation, there are several other practical, technical, and sociocultural challenges that have yet to be addressed. These will be explored in detail in the following section. WHAT ARE CURRENT CHALLENGES IN IMPLEMENTATION OF AI IN CLINICAL PRACTICE? Practical Challenges The first major challenge in the implementation of validated AI solutions is the need for a whole solution suitable for practical application. This may require a combination of DL systems with clinically acceptable performance and interoperability of the solution to receive images of varying quality from commonly used devices. Supporting clinical guidelines for the clinical communication of DL systems classification to patients and for patient selection also need to be developed. Grassmann et al have previously illustrated the importance of appropriate patient selection in their report of misclassifications by a DL system that arose due to macular reflex in younger patients in their validation dataset, when the young were not represented in the DL systems training dataset.33 Furthermore, most existing DL systems have only been validated for independent classification of a single eye disease at a time. A need to maintain and switch between multiple DL systems for each possible eye disease is not practically feasible for providers without technical AI or software expertise. Fortunately, researchers are developing solutions for these challenges, such as Ting et al and Son et al who have demonstrated accuracy of a single DL systems to classify multiple referable eye diseases such as DR, glaucoma, and AMD.36,46 Other challenges common to any new technology also need to be addressed. For instance, as with any screening test, supporting systems and processes are needed to address misclassified patients such as false-positives or negatives. This relates to the need for standardized practices in the development, validation, reporting, and implementation of AI to avoid misclassification when applied to fundamentally different target populations.33 Finally, the ethical and legal challenges in applying DL systems for classification of clinical data given that it is not explainable also need to be addressed. In the event of a misclassification, does the liability lie with the clinician or the AI technology provider? Many questions like these have yet to be addressed and will contribute to resistance to adoption. Technical Challenges First, the need for suitable training data and external validation is a major technical challenge that needs to be addressed to facilitate generalizability and translation of these solutions. This has been further elaborated on by Goh et al47 in their article contributed to this issue. This also begets the next challenge—the laborious labeling of input data for the training process, requiring the involvement of expert practitioners that make human errors. Coupled with the need for careful patient selection, the repeated need for labeling imaging datasets for calibrating a DL system for each new population may delay adoption and contribute to set-up costs.50 However, emerging research has demonstrated that DL algorithms could be trained to automate the labeling of training data using data from a detailed 3-dimensional scan such as the OCT, to train a new DL algorithm to classify a different type of scan such as 2-dimensional CFPs.51 This could address the human limitations and help scale up the calibration process for improved performance when a DL system is applied to new target populations. Furthermore, the “black box” nature of AI and lack of explainability of DL techniques is another major technical challenge that needs to be addressed. Progress in the development of explainable AI for image classification includes emerging techniques such as “soft attention,”52 occlusion testing, and saliency maps to identify the pixels in images that drive the DL systems’ classification.53 New methods such as “weighted error scoring” are also being developed to help evaluate the impact of wrong automated classification decisions against grading by humans.54 However, these methods still require the development of expert consensus, standards, and guidelines on evaluating the performance of any DL systems for implementation in eye care services. Sociocultural Challenges There are sociocultural challenges globally in the adoption of AI technology in clinical practice. Asia is a good example. Asia is a vast continent with heterogeneous population biology and culture, along with unique differences in health spending and consumption patterns.55 Although national Asian health systems are rapidly developing with pockets of communities that have comparable resources and health systems to the west, resource-limited settings encompass a significant majority. These communities remain disadvantaged by considerable geographical variation in infrastructure availability, health care accessibility and population income inequality.56 These major social determinants of health would impact the effectiveness and adoption of any new health technology intervention.14 These present unique challenges in the implementation of disruptive solutions for AI-based eye screening programs, given that the needs of both high and low demand users need to be addressed for successful deployment of a disruptive population screening tool. First, a lack of infrastructure is a major challenge for implementing digital solutions in some regions due to large areas without reliable access to electricity and/or internet. Therefore, portable solutions with rechargeable power supply may be needed to facilitate mobile screening programs. Open-source geolocation tagging would help in the co-ordination of resources for follow-up initiatives by government or philanthropic organizations, such as cataract surgery outreach camps. Gaps in internet coverage necessitate the development of AI-solutions localized within screening hardware to provide real-time results effectively offline that be conveyed to patients at the point-of-care, enabling de-centralized care.57 The alternative would be “store-and-forward” device configurations that collect geotagged data during mobile screening and subsequently upload data to a cloud-based server for analysis once docked in a centralized terminal or connected to the internet. However, the latter would require new systems to centralize care co-ordination and communicate results to patients, unlike local solutions that deliver insights at the point of care—which can be more easily overlaid to clinical microsystems. Furthermore, specialists generally practice in tertiary hospitals with massive geographical catchment areas of referrals from primary eye care providers. Therefore, asynchronous teleophthalmology platforms may be an essential feature to ensure appropriate referrals of patients triaged to tertiary care.4 This will reduce false-positives and unnecessary travel that can amount to days between rural settlements and the nearest tertiary care provider. The major considerations for AI-based screening programs for resource-limited settings are summarized diagrammatically in Figure 2.FIGURE 2: Developing a model for implementation of AI-based eye screening in resource-limited settings. Illustrates considerations for a whole solution designed for AI-based population eye screening in resource-limited settings including AI-algorithms that are either (A) Localized within screening tools (top left) or (B) cloud-based with “store-and-forward” configuration (bottom left); (C) asynchronous teleconsultation systems for low cost review of referrals by tertiary providers (bottom right); (D) Cybersecurity to protect the various forms of patient data in storage or transit (top right); all of which are built around (E) mobile solutions with geolocation tagging to facilitate co-ordination over vast geographical catchment areas (center). AI indicates artificial intelligence.CONCLUSIONS Given all these significant challenges, it may be tempting to dismiss AI as a “passing fad” and difficult to see how AI is a “near-enough reality” with whole solutions for clinical implementation. However, AI and DL capabilities are rapidly improving with new potential solutions for technical challenges emerging all the time, some of which have been outlined above. Furthermore, research of AI in ophthalmology has transitioned from the development and validation of these tools towards the implementation. This is a key step required to scope out the potential practical and sociocultural challenges to develop targeted solutions based on the needs of users: patients and providers. Although it may not be “months away,” AI solutions in ophthalmology is likely a reality in this new decade. In conclusion, in narrow, well defined areas, AI has shown tremendous promise in expanding health systems’ capacity of eye screening. For one, this can be achieved by automating the classification of DR and other eye diseases in multiple clinical applications, highlighted by Ruamviboonsuk et al45, and in management of certain diseases like cataract as suggested by Goh et al47. Naturally, as the technology progresses from clinical validation to translation with increasing technology readiness level, new challenges in integration of AI solutions in clinical care pathways and health care systems will emerge. Potential targeted solutions are already being investigated and reported for many of the challenges outlined. Low-cost, simpler digital solutions, such as smart phone-based systems in resource-limited settings,58 may add to the ease of AI implementation. All factors considered, AI seems to be primed and ready for translation and implementation to revolutionize eye care beyond 2020.
Gunasekeran et al. (Sun,) studied this question.