Proof of Good | SDG 3: Good Health and Well-Being

AI-Driven Diabetic Retinopathy Screening (Case Study: India)

In the pursuit of the United Nations Sustainable Development Goals, SDG 3: Good Health and Well-Being mandates universal access to quality healthcare and the reduction of preventable disease and disability. A leading cause of avoidable vision loss worldwide is diabetic retinopathy (DR): a microvascular complication of diabetes that progressively damages the retina and can lead to blindness if not detected early. Traditional screening programs require trained specialists to manually interpret retinal images, a process that is labor-intensive and often impractical in resource-constrained settings.

Innovative applications of data science, specifically artificial intelligence (AI) and machine learning (ML), are now transforming how DR is detected at scale. By automatically analyzing retinal fundus images and identifying disease severity with high accuracy, AI-based screening systems are improving early diagnosis, optimizing referral pathways, and directly contributing to better health outcomes. This case demonstrates how data, algorithms, and computational modeling can be leveraged for public health, supporting early intervention and reducing preventable morbidity in line with SDG 3.

Context and Challenge

According to epidemiological evidence, over 537 million adults worldwide are living with diabetes, and among those, around 145 million individuals are currently affected by diabetic retinopathy, representing approximately 27% of diabetic patients. A percentage that is prone to increase rapidly if not detected early. In India, a country with a rapidly rising diabetic population, the burden is particularly acute as population-level studies estimate that DR affects 13–18% of adults with diabetes, rising to 35–40% among those with poor glycemic control.

Early detection and timely treatment (such as laser photocoagulation or intravitreal therapy) can reduce the risk of severe vision loss by up to 90%, yet traditional DR screening programs depend on retina specialists, whose numbers are limited and unevenly distributed, especially in rural and underserved areas.

These constraints compromise eye care delivery, leading to delayed diagnosis, irreversible vision loss, and preventable disability which are a major barrier to achieving SDG 3’s goals of universal health coverage and disease prevention.

Data-Driven Solution

To overcome these challenges, technology developers ARTELUS™ designed an AI-Driven Diabetic Retinopathy Screening System (AIDRSS): a machine learning-powered diagnostic tool based on deep neural networks optimized for analyzing retinal fundus images.

Methodology and Algorithm:

The system was validated on a large multicenter dataset of 10,058 retinal images from 5,029 adult participants with diabetes across diverse healthcare settings in Kolkata, India. A deep learning model with 50 million trainable parameters was trained to classify retinal images according to the International Clinical Diabetic Retinopathy (ICDR) Scale into five categories (DR0–DR4). Techniques such as Contrast Limited Adaptive Histogram Equalization (CLAHE) were applied to enhance image quality, aiding feature recognition by the model. To ensure generalizability, the dataset was validated using five-fold cross-validation with an independent test set, ensuring no overlap between training and validation images. The system outputs both binary (DR present/absent) and multi-class severity predictions, enabling automated detection and triage.

Implementation and Results

The multicenter validation study tested AIDRSS performance against expert retina specialist assessments, revealing robust diagnostic capabilities in real-world health settings:

  • Overall Sensitivity: 92% (the percentage of true positive DR cases correctly identified)
  • Overall Specificity: 88% (the percentage of true negatives correctly excluded)
  • Sensitivity for Referable DR (DR3–DR4): 100% (all advanced, sight-threatening cases were correctly flagged)

These results indicate that AIDRSS can detect nearly all cases requiring urgent ophthalmic intervention while minimizing false positives, which is critical in resource-limited healthcare systems with constrained specialist capacity.

In the screened population, 13.7% had some form of diabetic retinopathy, and the prevalence rose to 38.2% among individuals with elevated blood glucose levels, highlighting significant unmet needs for early screening and DR management.

Outcomes and Impact

  1. Early detection at scale is achieved by automating DR screening with AI. AIDRSS enables early identification of disease often before symptoms appear, transforming population-level detection. All referable cases (DR3 and DR4: the most severe stages requiring prompt medical attention) were identified with 100% sensitivity. This early detection capability is vital because delayed diagnosis is strongly linked to permanent vision loss.
  2. Strengthened healthcare delivery with AI screening reducing reliance on scarce retina specialists as non-specialist providers can conduct initial screenings, and high specificity (88%) reduces unnecessary referrals, optimizing specialist time. AI thus improves resource allocation, enabling health systems to focus on high-risk patients and reduce clinical bottlenecks.
  3. The automated approach supports large-scale deployment across rural and urban settings where specialist resources are limited, accelerating progress toward universal access to eye care and inclusive health coverage.

The success of AIDRSS exemplifies how computational tools can fill gaps in public health infrastructure, particularly in low- and middle-income countries with growing chronic disease burdens.

Conclusion

The AI-Driven Diabetic Retinopathy Screening System case illustrates how data, AI, and machine learning can serve SDG 3 by delivering early, accurate, and scalable health interventions, especially in settings with limited clinical expertise. By processing more than 10,000 retinal images with advanced deep learning architectures and achieving 92% sensitivity and 88% specificity in disease detection, AIDRSS demonstrates the potential of AI to improve preventive healthcare and reduce preventable morbidity at population scale.

Importantly, the system’s 100% detection rate for referable diabetic retinopathy shows that AI can help health systems identify severe cases before irreversible harm occurs, effectively advancing SDG 3’s mission of ensuring healthy lives for all.

This case underscores that data-driven technologies are not just analytical tools, but also are impactful instruments for transforming healthcare delivery, expanding screening coverage, optimizing clinical workflows, and enabling earlier, more equitable health outcomes worldwide.