Malaysia’s AI-powered eye screening tool shows strong results in diabetes blindness prevention

Malaysia’s Ministry of Health has developed an artificial intelligence-based diabetic retinopathy screening system called DR.MATA, with early validation results suggesting it could significantly reduce the burden of diabetes-related blindness across the country’s public health network.

Presenting at the ADB-WHO Forum on Harnessing AI for Health Equity in Manila, Datuk Dr. Nor Fariza Ngah, Senior Consultant Ophthalmologist and Deputy Director-General of Health (Research & Technical Support) at the Ministry of Health, outlined the clinical rationale, technical development, and economic case for the system.

Diabetic retinopathy — damage to the retina caused by diabetes — is a leading cause of blindness globally, and one that is largely preventable through early detection and timely treatment. In Malaysia, more than four million people are living with diabetes, and Dr. Nor Fariza noted a growing trend of younger patients presenting with advanced complications. The condition disproportionately contributes to years lived with disability rather than years of life lost, meaning its impact on quality of life and workforce productivity is substantial.

Conventional screening in Malaysia relies on non-mydriatic fundus cameras operated at primary care level, but access to the equipment is uneven across facilities, staff turnover creates persistent training challenges, and results require specialist review before referral decisions can be made — introducing delays that can affect outcomes.

DR.MATA was developed to address these systemic gaps. Unlike many commercially available AI screening tools, which return only a binary present/absent result, the system grades diabetic retinopathy across three levels of severity — allowing clinicians to distinguish cases requiring urgent tertiary referral from those that can be safely monitored at primary care level, reducing unnecessary hospital referrals.

The algorithm was trained on approximately 15,000 annotated retinal images. In development, it achieved sensitivity of 88 to 92 percent, specificity of 93 percent, and overall accuracy of 84 percent. Validation conducted across five health centres and two hospitals returned sensitivity and specificity of around 92 percent, with multi-class accuracy of approximately 83 percent. Dr. Nor Fariza noted that image quality remains a key variable influencing both human and AI grading performance, and is an area of ongoing improvement.

The Ministry used Markov transition modelling to quantify the economic value of the intervention. Preventing 67 cases of blindness through early AI-assisted detection was estimated to generate savings of between 10 and 20 million Malaysian ringgit — a figure that comfortably exceeds the combined costs of cameras, AI software, and staff time required to run the screening programme. Dr. Nor Fariza characterised the system not merely as a clinical tool but as a cost-containment strategy for chronic disease complications, one that reduces downstream tertiary expenditure and preserves workforce productivity within a constrained public health budget.

The Ministry is now exploring whether the system can be adapted for use with smartphone-connected cameras, with the aim of extending screening reach to remote and underserved communities. Dr. Nor Fariza framed this as part of a broader commitment to ensuring that advances in health AI do not widen existing inequities in access to care.