AI-Based devices in national screening programmes: barriers and challenges
Reviewed by Sofia Rokerya

Diabetic retinopathy (DR) may have long-term complications and is recorded as a leading cause of blindness. National DR screening programmes have effectively reduced severe visual loss by timely detection and subsequent treatment of sight-threatening proliferative DR and diabetic macular oedema. Diabetic retinopathy screening is a challenging and time-consuming task employing huge manpower and resources. The use of automated AI has been proposed to reduce the need for human input, while still maintaining a high diagnostic performance in grading DR. This article highlights the barriers and challenges of AI use in screening. Deep learning by convolutional neural networks is a method suitable for automated image analysis as it enables training of large data sets without hand-crafted human inputs besides ground-truth classification of DR. Results of sensitivities and specificities for DR detection have consistently been demonstrated and this has led to regulatory approvals of different algorithms. While the algorithms that have obtained regulatory approval so far have all presented excellent performance in binary classification of DR, they have not been sufficiently tested for the prediction of upcoming or progressive DR. Additionally, as the AI algorithms are trained on specific datasets, their performance can vary significantly when applied to a different target population or when used with dissimilar imaging techniques. Algorithms trained on one dataset may not perform well on another, limiting their applicability across different local and global healthcare settings. Effective integration of AI into screening programmes requires access to diverse, representative devices and datasets plus their valid comparison in variable clinical settings. The grading and referral standards among different countries may also require alignment before attempting to integrate AI into the existing clinical workflow. Ensuring seamless data exchange with electronic health records can be a challenge and presents an area of potential concern. Patient privacy and data protection require robust security measures to be in place. Other issues that may need to be streamlined are the regulatory and approval processes, as well as the financial implications for initiation and maintenance. Currently as it stands, one in four images are classified as ungradable by algorithms. This error is more than that of human graders. Moreover, the existing AI systems that have developed a deep learning platform capable of detecting common referable fundus diseases and conditions, raises ethical concerns as regards the margin and responsibility of error in case of misdiagnosis. Further ongoing work is needed to find ways to develop best practices to overcome these barriers. This process is deemed critical to improving access to care and promoting health equity for this vision-threatening disease.

The seven sins of automated diabetic retinopathy screening: barriers for clinical implementation of artificial intelligence-based devices in national screening programs.
Grauslund, J, Grzybowski A.
OPHTHALMOLOGICA
2025;248(3):137–40.