A Review of Recent Developments in Artificial Intelligence and Big Data Technologies for Ophthalmology Referrals and Clinical Practice

Introduction

Smart technology refers to the use of devices or systems that integrate the computing, connectivity, or automation of tasks. Within ophthalmology, smart technology most commonly refers to use of artificial intelligence (AI)-powered tools, clinical decision support systems (CDSS), and big data simulations [1]. AI specifically refers to the imitation of human problem-solving by a machine, and encompasses several subtypes, each contributing to improved clinical efficiency [1]. For example, machine learning is a subset of AI capable of learning from experience without direct programming, allowing it to identify patterns and make predictions [1]. CDSS are guidance tools capable of assisting physicians during complex clinical decisions that are also presently being augmented with AI to enhance how users access and leverage this tool [2]. Chatbots are informational resource tools with user interface modules that enable text- or speech-based question and answer conversations to facilitate assistance with patient queries [3].

Beyond AI, big data analytics enhance ophthalmic care by extracting insights from large datasets that are too complex for manual analysis [4]. These analyses can predict future outcomes, optimizing administrative efficiency, and guide data-driven decision-making in healthcare [4]. Together, AI and big data have already demonstrated their ability to streamline high-volume, repetitive tasks including predictive inventory management [5], optimized appointment scheduling [5], and electronic data management [5], and are increasingly being applied in disease diagnosis, such as diabetic retinopathy [6]. The aim of this article is to review the clinical utility of emerging smart technology inclusive of AI, big data analytics, and CDSS relevant to ophthalmic referrals and ophthalmology practice.

Smart Technologies for Ophthalmic Referrals

CLINICAL DECISION SUPPORT SYSTEMS:

AI-driven CDSS have shown significant promise in managing ophthalmology referrals, often matching or surpassing the accuracy of physicians and optometrists. Tanya et al (2023) describe a CDSS for triage in primary care that analyzes ophthalmic history alongside a basic physical exam, including visual acuity and extraocular movements [7]. This particular CDSS was evaluated in comparison with primary care providers for diagnosing age-related macular degeneration, cataracts, and glaucoma and in its capacity to determine referral urgency [7]. The CDSS achieved a 53.1% diagnostic agreement with ophthalmologists, outperforming referring primary care providers, who achieved a 43.8% diagnostic agreement with ophthalmologists [7]. Furthermore, the same CDSS also correctly determined the indication for urgent referral in 68.8% of cases, compared with 46.9% for referring providers [7]. These results suggest that AI-driven CDSS have the potential to enhance ophthalmology referral efficiency by accurately detecting ophthalmic disease and minimizing false alarms for unnecessary ophthalmic findings, and therefore may be helpful in future healthcare systems.

FUNDUS ANALYSIS TECHNOLOGY:

Numerous efficient referral systems have been developed which analyze fundus images or EHR data to facilitate patient transfer from a primary care provider to the ophthalmologist. Specifically, autonomous AI systems have increasingly been used for screening asymptomatic diabetic patients in primary care settings. For example, Channa et al (2023) describe an autonomous AI screening tool that analyzes fundus images of diabetic patients taken during routine visits under care by a primary care physician [8]. Under current clinical guidelines, diabetic eye screening is typically conducted through physician-initiated referrals from primary care to ophthalmologists for in-person, dilated retinal exams [8]. However, this approach is not without limitations; patients with early or asymptomatic disease may go undetected due to missed referrals or lack of timely screening, while others may be referred despite having no clinically significant findings, leading to inefficient use of specialist resources. The authors’ model uses retinal fundus photographs to autonomously screen patients with diabetic retinopathy or macular edema in the primary care clinic, only sending high-risk patients to the ophthalmologist for additional care [8]. The authors found that replacing the traditional primary care referral sequence with an immediate point-of-care AI screening followed by selective referrals could prevent vision loss in over 27 000 diabetic patients over 5 years, underscoring the importance of streamlined ophthalmology referrals [8]. A variety of models also capable of analyzing fundus images have been developed to screen for diabetic eye disease [9] and support referral triage, such as RetCAD and IDx-DR, both of which have been evaluated in real-world screening environments for their detection of referrable diabetic retinopathy [10]. IDx-DR and RetCAD are 2 automated diabetic retinopathy image assessment systems (ARIAs), which are software packages designed to grade fundus photographs and improve the efficiency of diabetic retinopathy screening [10]. While IDx-DR is an FDA-approved ARIA, used in the United States, RetCAD is a contemporary software model more commonly used in the European Union [10]. In a recent study comparing these systems, RetCAD achieved a sensitivity of 89.4% and specificity of 94.8% for detecting referable diabetic retinopathy, while IDx-DR achieved a sensitivity of 99.3% and specificity of 68.9% [10]. Both models are intended for use in asymptomatic diabetic patients undergoing routine screening, and their high diagnostic accuracy enables immediate referral recommendations without requiring manual interpretation [10].

Other fundus analysis models are being developed which can detect diabetic retinopathy along with various other retinal diseases simultaneously, allowing for more efficient referrals to ophthalmologists. The Retinal Artificial Intelligence Diagnosis System (RAIDS), a deep-learning model trained to screen for 10 retinal conditions, achieved an accuracy of 98.1% for detecting referable diabetic retinopathy. RAIDS also accurately identified referable age-related macular degeneration (97.3%), possible glaucoma (94.3%), pathological myopia (98.4%), and retinal vein occlusion (97.4%), with accuracy exceeding 97% for most pathologies [11]. RAIDS also achieved a higher sensitivity for detecting any retinal abnormality (91.7%), compared with certified ophthalmologists (83.7%), junior retinal specialists (86.4%), and senior retinal specialists (88.5%) [11]. These multi-disease screening systems enhance the diagnostic yield of routine diabetic eye evaluations and improve referral accuracy by simultaneously evaluating for coexisting retinal pathology [11]. Additionally, because models like IDx-DR and RAIDS specifically assess for referrable diabetic retinopathy, these models specifically advise primary care physicians when to refer, or not refer, for ocular pathology [10,11].

EHR DATA ANALYSIS:

EHRs contain vast and valuable clinical data that are presently being leveraged against predictive machine learning models to screen and assess for ophthalmic diseases that may require ophthalmology referral in the general medicine setting. Young et al (2024) developed and compared various machine learning models that analyze patient EHR data to identify patients requiring ophthalmology referral [12]. Overall, these machine learning EHR models predicted ophthalmic disease based on history alone with an accuracy of 80% [12]. Additionally, the authors calculated odds ratios to quantify the likelihood of a patient having a referrable pathology when flagged by the models, which ranged from 2.4 for glaucoma to 5.7 for non-proliferative diabetic retinopathy, with an average of 3.9 across all diseases studied [12]. Thus, by implementing these models, primary care providers can rapidly identify patients under their care who have a nearly 4 times greater likelihood of having an undiagnosed ocular pathology, indicating a potential missed referral [12]. Use of machine learning EHR risk stratification tools at the start of each general medicine clinic day would further enable primary care providers to proactively identify high-risk or medically complex patients that may require more attention should referral for ophthalmology evaluation prove indicated.

Machine learning EHR models have also more recently been designed to facilitate referrals for complex conditions such as lens dislocation and low vision that may require a subspecialist within ophthalmology. Stein et al (2024) developed a natural language processing model that identified lens pathology by analyzing structured and free-text entries from the lens examination section of the EHR [13]. This model is valuable because ICD-10 codes for secondary lens pathologies are often omitted from the billing submission, despite being documented within the lens examination or patient encounter free-text fields of the EHR [13]. The authors found that their natural language processing models can correctly identify lens pathology, such as lens dislocation, pupillary membranes, and aphakia in 95.1% of tested entries [13]. Compared with clinician-submitted ICD-10 billing codes, the natural language processing algorithms demonstrated a substantially higher rate of detection for lens pathology (98% vs 49%) as well as a substantially lower false negative rate (2% vs 52%) [13]. These findings suggest that natural language processing EHR review can significantly improve detection of undocumented lens pathologies in the primary care clinic [13]. By scanning clinical notes and structured fields for missed diagnoses, this technology may better inform referral decisions and allow primary care providers to accurately manage patients at risk for lens-related complications [13].

Smart Technologies Within the Ophthalmology Clinic

BIG DATA ANALYSIS:

Kern et al (2021) used big data simulations to assess the impact of various clinical workflow changes on total waiting times (TWT) and overall clinic efficiency [14]. By modeling different appointment scheduling strategies, the authors identified an optimal approach that combined block scheduling with extended appointment windows [14]. When the model was implemented in real clinical practice, the mean TWT was reduced per patient from 229 ± 100 minutes to 183 ± 89 minutes, a 21% decrease achieved without additional staff or equipment [14]. The study highlights the potential for big data-driven scheduling systems to optimize patient appointments cost-effectively, allowing clinics to test workflow changes before real-world implementation.

EHR-BASED TOOLS:

Other smart technology tools, such as EHR alerts, have been explored to further improve patient attendance and reduce missed appointments. Missed appointments in ophthalmology clinics increase costs, delay diagnoses, and limit treatment of preventable disease [15]. Atta et al (2024) evaluated EHR alerts for reducing no-show rates in ophthalmology clinics [15]. The authors implemented an automated EHR messaging system capable of notifying patients 1 day after a missed appointment with an opportunity to reschedule [15]. In the study, the control group received only a standard mailed letter recommending rescheduling, while the intervention group received a mailed letter along with the additional EHR notification [15]. Within 30 days, 22.2% of patients who received the EHR message attended a follow-up appointment, compared with 11.6% in the control group [15]. Therefore, even minor optimizations within EHR systems can significantly enhance workflow effectiveness, reduce operational disruptions, and ultimately improve patient access to timely ophthalmologic care [15].

Much like in referrals from the primary care provider to the ophthalmologist, CDSS tools capable of analyzing EHRs can also be leveraged to enhance the process of referrals from an ophthalmologist to a subspecialist. For example, EHR-analyzing AI models have been designed to identify patients who may have low vision, defined as measurable impairment in visual acuity or a clinically significant loss in visual acuity causing reduced ability to perform daily activities [16]. Goldstein et al (2022) developed an AI-driven CDSS that analyzes EHR data to detect low-vision patients and generate electronic alerts for potential rehabilitation referrals within the institutional EHR system [16]. In a set of 40 000 EHR data encounters, the CDSS identified 8.9% of patients in the ophthalmology clinic as in potential need for low-vision rehabilitation and after ophthalmologist review, nearly 15% of patients with low-vision alerts prompted an actual referral order for additional low-vision services [16]. These findings demonstrate that, whether implemented in the primary care or ophthalmology setting, EHR-based AI can be integrated to more efficiently identify and manage patients at increased risk for ophthalmic disease.

CHATBOTS:

Chatbots are emerging as a valuable tool for the healthcare system capable of collecting patient data, providing personalized, evidence-based answers to patient questions, improving health care decision-making, and ultimately relieving time for physicians to focus on clinical tasks [23].

Large language models (LLMs) such as ChatGPT represent one form of AI-driven chatbot and are being explored in ophthalmology clinics to improve productivity by responding to patient questions. Bernstein et al (2023) compared answers to patient questions between LLMs, specifically ChatGPT-3.5, and ophthalmologists [17]. The authors gathered questions posted by patients on Eye Care Forum and the corresponding answers from American Academy of Ophthalmology (AAO) physicians [17]. The authors then used ChatGPT-3.5 to generate an answer to the same questions posed by patients on Eye Care Forum [17]. A panel of blinded expert ophthalmologists analyzed responses from AAO physicians and ChatGPT-3.5, finding no statistical differences between the groups in terms of incorrect information, alignment with medical consensus, or likelihood to cause harm [17]. The authors suggest that LLMs could be used to update patients on common ocular health concerns before meeting with an ophthalmologist, freeing time for increased focus on complex cases [17]. Thus, as AI-driven chatbots evolve, their role in ophthalmology may expand beyond administrative support to automating intake, triaging symptoms, and managing follow-up care, ultimately improving clinic efficiency.

“Dora” is an ophthalmology-specific AI-driven chatbot that has been implemented in the UK to conduct follow-up calls after cataract surgery, freeing time for in-person visits [18,19]. Preliminary studies of Dora have indicated high rates of patient satisfaction when interacting with Dora, including subjective feelings of time saved compared with in-person visits [18]. More recently, Meinert et al, in 2024, evaluated Dora’s cost-saving potential and safety profile in post-cataract surgery follow-ups. In this study, Dora autonomously conducted post-cataract surgery follow-up calls and was tasked with detecting patient symptomatology and recommending either immediate discharge or further review by an ophthalmologist [20]. Dora’s ability to detect 5 key potential patient symptoms, including redness, pain, vision changes, floaters, and flashing lights described by patients were compared with supervising ophthalmologists’ interpretation [20]. Dora achieved a high accuracy in determining if a patient had a symptom, ranging from over 97% accuracy for vision changes to over 99% accuracy for redness and flashing lights [20]. Additionally, only 1.3% of Dora’s symptom assessments were errors and only 0.7% were false negatives [20]. Dora recommended discharge for 60% of patients, with the remaining 40% being designated as requiring further review [20]. Among those discharged by Dora, roughly 3% were instead recommended for further review by a supervising ophthalmologist; however, none of these patients ultimately required clinical review after clinician call-back [20]. Additionally, 9% of patients recommended for discharge by Dora experienced unexpected management changes, but these patients had also been approved for discharge by supervising ophthalmologists, demonstrating no increased risk compared with standard care [20]. By successfully recommending patients for discharge without a face-to-face meeting, the authors found that Dora provided a cost savings of £35.18 per patient compared with typical care, mainly via a reduction of staff costs associated with the full clinical pathway [20]. As of 2022–2023, up to 600 000 cataract surgeries are performed in England each year [24], indicating a potential cost savings of over £21 million each year by implementing Dora for all cataract surgery follow-up calls. These findings indicate that chatbot-driven automation can safely enhance post-operative care while improving efficiency and cost-effectiveness in ophthalmology clinics.

DIAGNOSTIC TECHNOLOGY:

Within the ophthalmology clinic, smart technologies are being integrated into ophthalmic imaging workflows to bolster image quality assessment and accelerate disease diagnosis. Abramovich et al (2023) describe the development of FundusQ-Net, which is a deep learning-based tool designed to assess fundus image quality on a continuous 1–10 scale [21]. After training the model, FundusQ-Net was tested on an external data set and demonstrated a 99% accuracy in distinguishing between good and poor-quality images [21]. By automating the quality assessment of fundus images, this model reduces the burden on human graders, and prevents poor-quality images from entering diagnostic workflows, benefiting both ophthalmologists and further AI-assisted ophthalmic analysis [21].

Additionally, various AI-based diagnostic technologies are being developed to help ophthalmologists more rapidly distinguish between healthy patients and those with pathology. For example, Lin et al (2019) developed an AI model for rapid cataract diagnosis [22]. While the AI model was less accurate than expert ophthalmologists (87.4% vs 99.1% accuracy), it provided a significantly faster diagnosis (2.97 vs 8.53 minutes) [22]. Similarly, Abramoff et al (2023) studied the impact of implementing AI technology, specifically LumineticsCore, to hasten diabetic eye exams already being conducted in the ophthalmology clinic [25]. In the control group, all patients underwent a diabetic eye exam conducted by AI, followed by a mandatory confirmatory exam performed by an ophthalmologist [25]. In the intervention group, only patients flagged by AI as needing further evaluation received an additional ophthalmologist exam [25]. Retina specialists who were part of the control group achieved 1.14 completed care encounters per hour, while specialists in the AI intervention group achieved 1.59 completed care encounters per hour, indicating that AI can increase productivity within the ophthalmology clinic by 40% for diabetic eye exams [25]. These findings provide early evidence that AI models can be designed to increase diagnostic speed, and as accuracy continues to advance, their full potential in ophthalmology clinics may be realized.

Future Directions

The integration of smart technologies into ophthalmology has already demonstrated significant potential to enhance efficiency, but as smart technology continues to be developed, several key areas remain primed for further innovation. Within the referral process, a variety of triaging systems exist capable of analyzing clinical data, EHR data, and fundus images [7,8,12]. However, a model capable of analyzing all patient data within the triage process simultaneously is still lacking. Further research should focus on the development of multimodal AI models capable of integrating all available patient data, from the EHR system to the physical exam and fundus images, to allow for a more comprehensive patient assessment and to facilitate the continual improvement of referral accuracy.

While refining AI-driven triage models is essential for improving referral effectiveness, further efforts must validate smart technology’s scalability across diverse practice settings and global healthcare systems to ensure widespread ophthalmologic efficiency. For example, while the use of big data to reduce waiting times proved efficacious at the clinic level [14], further research is needed to explore the impact at the hospital-wide level and, ultimately, across entire healthcare systems. Similarly, other technologies, such as chatbots, must continue being evaluated for their adaptability beyond their countries of origin. Ongoing studies are investigating the implementation of Dora in Canada to assess its potential for improving productivity in different healthcare systems, highlighting the need for further research into the cross-border applicability of AI-driven tools [19].

Finally, future studies must focus on the cost-effectiveness of smart technology along with the potential for economic reimbursement for ophthalmologists who implement these models. Existing data indicate that smart technology can aid ophthalmologists by reducing visual complications experienced by their patients and increasing the number of patients that can be seen per day [8,25]. However, few studies have quantified these benefits in economic terms, leaving the financial impact of these technologies ambiguous. Additionally, in the United States, only a few AI models have received FDA approval and eligibility for insurance reimbursement, such as LumineticsCore and AEYE-DS, both of which are primarily used for diabetic retinopathy [25,26]. To facilitate broader adoption, additional AI models must be developed and validated for a wider range of ocular diseases, ensuring more practical and widespread integration of AI into ophthalmic practice with financial incentives for increased productivity.

Conclusions

Smart technologies, including AI, big data analytics, and CDSS, are revolutionizing ophthalmology by enhancing patient referral accuracy, optimizing clinic workflow, and increasing diagnostic efficiency. The widespread implementation of AI-based triage systems, big data analytics, chatbots, and diagnostic models may reduce administrative burden and patient wait times, optimize resource allocation, and enhance diagnostic efficiency in ophthalmic practice.

Despite these advancements, further research is needed to refine multimodal AI models capable of integrating diverse clinical data sources, validate the scalability of smart technologies across different healthcare settings, and quantify their long-term economic benefits. Expanding regulatory and reimbursement frameworks to encompass a broad range of ocular diseases beyond diabetic retinopathy will be critical for widespread adoption. The future of ophthalmic care will be shaped by the responsible integration of these technologies, with a focus on enhancing clinical efficiency while preserving patient safety and equity.