The Growing Challenge of Diabetic Eye Disease

Diabetic retinopathy (DR) remains one of the most consequential complications of diabetes mellitus, affecting an estimated 103 million people worldwide as of 2020. The condition progresses from mild non‑proliferative changes to proliferative diabetic retinopathy (PDR) with neovascularization and diabetic macular edema (DME), ultimately threatening vision. Despite advances in glycemic control and systemic management, nearly one in three people with diabetes develops some form of DR. The global diabetes epidemic, projected to affect 700 million individuals by 2045, will only intensify this burden. Healthcare systems face growing pressure to deliver timely, effective treatments that prevent vision loss while managing limited resources.

Traditional treatment paradigms have relied on monotherapy—anti‑vascular endothelial growth factor (anti‑VEGF) injections, corticosteroids, or laser photocoagulation—each targeting a single pathogenic pathway. However, the multifactorial nature of DR, which includes VEGF upregulation, inflammation, neurovascular impairment, and capillary leakage, often requires a combinatorial approach. This is where dual therapy strategies have gained traction, combining two agents or modalities to achieve synergy, reduce injection frequency, and improve anatomical and functional outcomes. Yet selecting the right combination for the right patient at the right time remains a clinical challenge—one that artificial intelligence (AI) is uniquely positioned to address.

Understanding Dual Therapy in Diabetic Eye Care

Dual therapy in diabetic eye care refers to the co‑administration or sequential use of two distinct therapeutic interventions to address the complex pathophysiology of DR and DME. Common combinations include:

  • Anti‑VEGF plus corticosteroid implant (e.g., intravitreal ranibizumab or aflibercept combined with dexamethasone or fluocinolone acetonide)
  • Anti‑VEGF plus focal/grid laser (e.g., ranibizumab plus laser photocoagulation)
  • Combination of two anti‑VEGF agents with different binding profiles (less common, but explored in refractory cases)

The rationale behind dual therapy lies in complementary mechanisms. Anti‑VEGF drugs neutralize VEGF‑A, reducing vascular permeability and neovascularization, but they do not address the inflammatory and neurodegenerative components. Corticosteroids suppress inflammatory cytokines and stabilize the blood‑retinal barrier, yet carry risks of cataract and intraocular pressure elevation. Laser photocoagulation reduces oxygen demand and seals leaking microaneurysms but can cause scotomata. By combining these approaches, clinicians can maximize efficacy while mitigating the adverse effects of higher‑dose monotherapy.

Clinical Evidence Supporting Dual Therapy

The Diabetic Retinopathy Clinical Research Network (DRCR.net) Protocol T compared ranibizumab, aflibercept, and bevacizumab monotherapies, establishing anti‑VEGF as first‑line for center‑involved DME. Subsequent studies demonstrated that adding laser or corticosteroid to anti‑VEGF can reduce injection burden while maintaining visual gains. For example, the VIVID‑EAST trials showed that aflibercept combined with laser achieved similar visual acuity improvements with fewer injections than aflibercept monotherapy at one year. Similarly, the FLUID trial found that adding dexamethasone implant to anti‑VEGF in eyes with persistent DME led to greater anatomical improvement, though intraocular pressure monitoring was required. However, not all patients benefit equally: some respond poorly to anti‑VEGF monotherapy and require early combination, while others may do well with anti‑VEGF alone. The ability to predict which patients need dual therapy upfront is where AI can transform clinical decision‑making.

The Role of Artificial Intelligence in Optimizing Dual Therapy

Artificial intelligence, particularly deep learning models, excels at identifying patterns in high‑dimensional medical data that elude conventional statistical analysis. In diabetic eye care, AI can optimize dual therapy strategies across three critical domains: diagnosis and disease phenotyping, predicting treatment response, and dynamic treatment planning.

AI‑Powered Diagnosis and Disease Phenotyping

Accurate staging of DR and DME is prerequisite to selecting the right therapy. AI algorithms trained on millions of retinal fundus photographs and optical coherence tomography (OCT) scans can grade DR severity with accuracy comparable to or exceeding retinal specialists. For instance, the FDA‑cleared IDx‑DR system detects more‑than‑mild DR with sensitivity above 87% and specificity above 90%. More advanced models now classify DME by morphological subtypes—diffuse, cystoid, serous—each of which may respond differently to anti‑VEGF versus steroids. A deep learning model developed at Singapore’s SERI can differentiate between predominantly inflammatory versus vasogenic edema on OCT, with an AUC of 0.91, guiding clinicians to choose corticosteroids as part of dual therapy.

By segmenting fluid volumes in OCT B‑scans, AI can quantify central subfield thickness and detect sub‑clinical changes weeks before visual acuity declines. A 2024 study in Ophthalmology Retina demonstrated that AI‑based volumetric analysis of intraretinal fluid predicted which eyes would develop chronic DME with 84% accuracy. This allows clinicians to identify patients with a predominantly inflammatory phenotype (e.g., large cystoid spaces) who are likely to benefit from upfront addition of a corticosteroid implant, rather than trying multiple anti‑VEGF injections first. The algorithm effectively phenotypes the disease, enabling precision dual therapy.

Predicting Treatment Response to Dual Therapy

One of the most promising applications of AI is predicting an individual patient’s response to specific dual therapy regimens. Researchers have developed machine learning models that integrate OCT imaging features, clinical variables (HbA1c, duration of diabetes, renal function), and genetic markers to forecast visual and anatomical outcomes. For example, a 2023 study published in JAMA Ophthalmology showed that a deep learning model using baseline OCT features had an AUC of 0.83 for predicting which DME patients would achieve a ≥10‑letter gain at 12 months on ranibizumab plus laser versus monotherapy. The model incorporated features like ellipsoid zone integrity and cyst location to stratify patients.

Another study from Moorfields Eye Hospital used a gradient‑boosting model combining age, OCT fluid location, and prior anti‑VEGF exposure to predict non‑response to monotherapy, with a sensitivity of 78%. These predictive models enable clinicians to avoid trial‑and‑error treatment, a common problem where patients undergo months of ineffective therapy before switching. Instead, AI can recommend starting dual therapy immediately for those predicted to have a suboptimal response to monotherapy, saving time, reducing injection burden, and preserving vision. Such approaches align with the goals of value‑based healthcare.

Dynamic Treatment Planning with AI

AI’s real‑time analytical capacity allows for adaptive treatment algorithms that adjust dual therapy combinations and dosing intervals as the disease evolves. Using reinforcement learning—a type of AI that learns optimal actions through feedback—models can propose a treatment schedule that minimizes the cumulative number of injections while maximizing visual outcomes. This is particularly valuable for long‑term management of chronic DME, where treatment fatigue leads to drop‑out and recurrence.

In practice, an AI system could analyze each follow‑up OCT scan and clinical visit to recommend: “Maintain current anti‑VEGF monotherapy,” “Add corticosteroid implant now given the increase in inflammatory biomarkers,” or “Consider switching to a combination of aflibercept plus laser due to persistent exudation.” Such dynamic decision support is already being tested in prototype Electronic Health Record (EHR) plugins at academic institutions like the University of Pittsburgh and Stanford. A 2024 proof‑of‑concept study used reinforcement learning to simulate treatment for 500 virtual DME patients and found that AI‑guided dosing reduced total injections by 30% while maintaining comparable visual gains to standard fixed‑interval regimens.

Technological Enablers: AI Models and Data Sources

The AI tools under development rely on diverse data inputs:

  • Imaging data: High‑resolution OCT, OCT angiography (OCTA), fundus autofluorescence, and ultrawide‑field imaging provide rich biomarker sets including capillary density, vessel tortuosity, and fluid volume.
  • Clinical data: Systemic factors (blood pressure, glycemic control, lipid levels), treatment history, and patient‑reported outcomes.
  • Genomic data: Single‑nucleotide polymorphisms (SNPs) linked to diabetic retinopathy risk and anti‑VEGF response (e.g., VEGFA, HTRA1).
  • Proteomic data: Levels of inflammatory cytokines in aqueous humor (e.g., IL‑6, MCP‑1) as potential biomarkers for steroid responsiveness.

Convolutional neural networks (CNNs) dominate image analysis, while gradient‑boosting machines and random forests are common for tabular clinical data. Multimodal AI that fuses imaging and clinical data is an active research frontier, with early models showing superior predictive performance over single‑modality approaches. For instance, a 2024 model from Google Health combined fundus photos, OCT, and EHR data to predict DME progression with an AUC of 0.92, outperforming any individual input. The trend toward federated learning allows institutions to collaboratively train robust models without sharing sensitive patient data, addressing privacy concerns while improving generalizability.

Challenges and Barriers to Clinical Adoption

Despite the promise, integrating AI‑guided dual therapy into routine practice faces several hurdles that must be systematically addressed.

Data Quality and Generalizability

AI models trained on high‑quality datasets from tertiary referral centers may not perform well in community clinics with different populations, camera models, or imaging protocols. The domain shift problem can lead to reduced accuracy and potentially harmful recommendations. For example, an algorithm trained on a predominantly White, affluent population may misinterpret retinal features in patients with Hispanic or African ancestry, where DR progression patterns differ. Rigorous external validation across diverse ethnicities, disease severities, and healthcare settings is essential. Several notable AI tools have failed replication in independent cohorts, highlighting the need for prospective, multicenter trials that include real‑world heterogeneity.

Regulatory and Reimbursement Issues

Only a handful of AI algorithms for diabetic eye care have received FDA clearance (e.g., IDx‑DR, EyeArt, and the recent LumiThera system for OCT analysis), and none are specifically approved for guiding dual therapy decisions. The regulatory pathway for an AI‑based treatment recommender is more complex than for a diagnostic tool, requiring evidence that the AI’s recommendation leads to better outcomes than standard care—a level of clinical validation that demands large, randomized controlled trials. Reimbursement models also lag: few payers in the US or Europe cover an AI consult for treatment selection, limiting real‑world deployment. The Centers for Medicare and Medicaid Services (CMS) have not yet established a specific code for AI‑assisted clinical decision support in ophthalmology.

Clinician Acceptance and Workflow Integration

Retinal specialists are trained to weigh multiple factors in treatment decisions; they may be skeptical of a black box AI recommendation, especially if it conflicts with clinical intuition. For AI to be adopted, it must provide explainable outputs—e.g., highlighting the specific OCT features driving the suggestion—and integrate seamlessly into existing EHR systems without adding extra clicks or delays. The human‑in‑the‑loop approach, where AI generates recommendations but the final decision rests with the clinician, has been shown to improve trust and adoption rates in early studies at the Doheny Eye Institute. Workflow integration also requires that AI outputs appear directly in the EHR alongside imaging data, rather than requiring a separate login or portal.

Privacy and Ethical Considerations

Training AI on large datasets raises concerns about patient privacy and data security. De‑identification, federated learning (where models are trained across institutions without sharing raw data), and adherence to HIPAA/GDPR are critical. Additionally, algorithmic bias—where AI underperforms in minority populations—must be actively monitored and mitigated. A 2023 systematic review in The Lancet Digital Health found that several DR screening AI models showed lower sensitivity in darker fundus pigmentation, underscoring the need for diverse training datasets. Ethical frameworks for AI in ophthalmology, such as those proposed by the American Academy of Ophthalmology, emphasize transparency, fairness, and accountability.

Future Directions: AI‑Driven Personalized Dual Therapy

The next decade will likely see AI evolve from a diagnostic aid to a true therapeutic partner. Promising avenues include:

  • Closed‑loop systems: AI that integrates continuous glucose monitoring data with retinal imaging to predict impending DME and automatically adjust dual therapy schedules via implantable drug‑delivery platforms. Early prototypes use reinforcement learning to optimize the timing of anti‑VEGF and steroid release based on glucose fluctuations and OCT fluid changes.
  • Virtual clinical trials: AI simulations that test dual therapy combinations in silico, accelerating drug development and identifying optimal regimens before costly phase III trials. Digital twins of patients—virtual replicas built from OCT and clinical data—can simulate thousands of treatment scenarios to find the best dual therapy for a given phenotype.
  • Predictive biomarkers: Deep learning on OCTA to detect early capillary dropout and predict which eyes will convert from non‑proliferative to proliferative DR, prompting pre‑emptive dual therapy with anti‑VEGF and laser to prevent high‑risk PDR.
  • Federated learning networks: Global collaborations that train robust AI models without moving sensitive data, ensuring broad applicability and fairness. Initiatives like the American Academy of Optometry’s AI consortium are exploring federated learning across multiple eye care sites.
  • Explainable AI dashboards: Tools that show clinicians exactly why a dual therapy recommendation is made, using heatmaps of OCT scans and presentation of similar patient cases from the training database. This transparency is key to building trust and facilitating regulatory approval.

A recent meta‑analysis in Ophthalmology (2024) examined AI in prediction of DME treatment outcomes and reported that models integrating OCT and clinical data achieved a pooled sensitivity of 82% for predicting worsening, with a specificity of 79%. While not yet standard of care, the trajectory suggests that within five years, AI‑guided dual therapy could become the norm for complex DME cases, especially in high‑volume centers where personalized decision‑making is both time‑consuming and critical.

Conclusion

Artificial intelligence offers a powerful lens to optimize dual therapy strategies in diabetic eye care, moving beyond one‑size‑fits‑all protocols to truly personalized medicine. By enhancing diagnostic accuracy, predicting treatment response, and enabling dynamic adjustments, AI can reduce the burden of frequent injections, improve visual outcomes, and preserve quality of life for millions of patients. The challenges of data quality, regulatory clearance, and clinician adoption are real but surmountable through continued interdisciplinary collaboration, investment in diverse data collection, and development of explainable AI systems. As research progresses and AI tools become more transparent and validated, their integration into clinical workflows will unlock the full potential of dual therapy—bringing us closer to a future where diabetic retinopathy no longer leads to preventable blindness. The next step is for clinicians, researchers, and regulatory bodies to work together to build the evidence base and infrastructure needed to make AI‑guided dual therapy a reality.