Postdeployment data from India show strong diabetic retinopathy screening performance
A deep learning diabetic retinopathy algorithm delivered high diagnostic performance in routine use across 45 sites in Southern India, with results that matter most where screening volume is high and access to specialist ophthalmology is uneven. In this cross-sectional analysis from Aravind Eye Hospital, the automated retinal disease assessment system, or ARDA, was tested after deployment rather than in a tightly controlled development setting. That distinction matters at the bedside and in screening programs alike, because algorithms can behave differently once they meet real-world image quality, disease mix, and referral pathways.
Among 4,537 patients with adjudicated grades, the mean age was 55.2 years and just over half were men. Of the 3,941 patients with gradable photographs, 683 had any diabetic retinopathy, 146 had severe nonproliferative diabetic retinopathy or proliferative diabetic retinopathy, 109 had proliferative disease, and 398 had sight-threatening diabetic retinopathy. These are the lesions that determine whether a patient needs prompt ophthalmology evaluation, not simply a flag in a screening database.
Severe disease detection stayed above 96 percent on both sensitivity and specificity
For severe nonproliferative diabetic retinopathy or proliferative diabetic retinopathy, ARDA reached sensitivity of 97.0% and specificity of 96.4%. The positive predictive value was 50.7%, while the negative predictive value was 99.9%. In practical terms, the system rarely missed severe disease and almost never cleared a patient who truly needed referral. The false-positive burden was real, but in screening medicine that tradeoff can be acceptable when the alternative is missing lesions that threaten vision.
One sentence captures the clinical stakes: the clinically important miss rate for severe nonproliferative diabetic retinopathy or proliferative diabetic retinopathy was 0%.
That outcome is especially meaningful because some patients with severe disease were interpreted as having moderate diabetic retinopathy and were still referred to clinic. The algorithm therefore preserved the safety net that screening programs depend on. In diseases driven by progressive microvascular damage, the cost of a missed referral is not abstract; delayed treatment can allow macular edema, neovascularization, vitreous hemorrhage, and irreversible visual loss to accumulate before a patient ever reaches an ophthalmologist.
Sight-threatening diabetic retinopathy was also identified with high accuracy
Performance remained strong when the target broadened to sight-threatening diabetic retinopathy, defined as diabetic macular edema or severe nonproliferative diabetic retinopathy or proliferative diabetic retinopathy. Sensitivity for this endpoint was 95.9%, and specificity was 94.9%. Positive predictive value was 67.9%, and negative predictive value was 99.5%. Those values suggest the tool was not simply finding obvious advanced disease; it was also separating true pathology from the much larger pool of screened images that do not require urgent escalation.
For a large screening network, that balance has immediate operational consequences. High sensitivity protects patients with vision-threatening lesions, while high specificity limits the number of unnecessary in-person referrals that can flood retinal clinics and lengthen wait times. The algorithm’s behavior therefore speaks to two systems at once: bedside safety for individual patients and throughput for eye-care infrastructure.
The postdeployment setting is the important part. Performance measured inside a clinical service after screening 600,000 patients is harder to dismiss than an internal validation result from a curated dataset. Real-world screening inevitably includes variable pupil dilation, cataract, uneven image capture, and a spectrum of diabetic eye disease that may not mirror the prevalence in model development cohorts. Here, the algorithm still held up well enough to support its role as a triage tool.
Why this matters for diabetic eye screening programs
Diabetic retinopathy follows a biological sequence that begins with chronic hyperglycemia-induced microvascular injury and progresses toward capillary leakage, ischemia, and neovascular change. An artificial intelligence system in this setting does not need to replace ophthalmologists; it needs to sort high-risk eyes from lower-risk ones with enough reliability that time-sensitive referrals are preserved. ARDA appears to have done exactly that in this Indian screening program.
The prevalence data also help explain the test characteristics. Severe disease was relatively uncommon compared with the overall screened population, which naturally depresses positive predictive value even when sensitivity and specificity are excellent. In other words, many patients flagged for referral will not have the most advanced retinopathy, but the system still protects the small subgroup whose disease could progress quickly if ignored.
For clinicians, the more useful signal is the near-perfect negative predictive value. A screen that reliably rules out severe disease can reduce avoidable referrals and concentrate specialist attention on the patients most likely to need laser treatment, intravitreal therapy, or closer follow-up. That is where AI can become a force multiplier in regions where retina specialists are scarce relative to the burden of diabetes.
Postmarketing monitoring is now part of algorithm care
This study also argues for a shift in how AI tools are regulated and followed after launch. A model is not finished when it clears development testing; it becomes clinically meaningful only after it is embedded in a screening workflow and observed at scale. The data from Southern India support ongoing surveillance of algorithm performance, especially as disease prevalence, camera systems, and referral thresholds evolve over time.
The findings are encouraging, but they are not the final word. Cross-sectional adjudication shows how the system performed on a sampled set of images, not how vision outcomes changed over years of follow-up. The study also does not answer how many referrals led to completed ophthalmic care, treatment initiation, or preserved visual acuity. Those endpoints will determine how much benefit the algorithm delivers beyond detection alone.
Even so, the signal is difficult to ignore: in a large postdeployment cohort, ARDA detected severe diabetic retinopathy with very high sensitivity and specificity and identified every patient with severe nonproliferative diabetic retinopathy or proliferative diabetic retinopathy for referral. In screening medicine, where the main danger is missing the patient who cannot afford to be missed, that is the result that counts.