Found the Disease, Lost the Patient

Aravind Eye Care has deployed AI-powered retinal screening at rural vision centers across Tamil Nadu, using smartphone-based fundus cameras operated by trained technicians. The AI achieves sensitivity >90% and specificity >85% for referable diabetic retinopathy — technically sufficient for population screening. But a persistent gap undermines the system's public health impact: only 30–40% of patients identified as having referable disease actually present to an Aravind hospital for confirmatory examination and treatment. The screening technology works; the referral pathway does not. The bottleneck is not detection but follow-through, and the causes are structural — transport cost and distance, lost daily wages, family caregiving obligations, and the difficulty of understanding why an eye that doesn't hurt yet needs surgery.

India has an estimated 77 million adults with diabetes and approximately 18% have some degree of diabetic retinopathy. Early detection and laser treatment can prevent 95% of severe vision loss. AI screening promises to make early detection scalable in settings without ophthalmologists — but if 60–70% of screen-positive patients never reach treatment, the screening program identifies disease without preventing blindness. This is not merely a compliance problem: a screening program that detects disease but doesn't connect patients to treatment may be worse than no screening, because it creates a documented population of untreated disease — a known failure that erodes trust in the health system.

Aravind has tried phone-based follow-up (calls to remind patients of their referral), patient navigators at vision centers, and subsidized transportation. Phone follow-up increases attendance modestly (5–10 percentage points) but doesn't address the structural barriers — a patient who needs to take two buses and lose a day's wages still can't afford to come, regardless of how many reminder calls they receive. Patient navigators help with system navigation but can't solve the transport-cost-versus-wages calculation. Subsidized transportation addresses direct costs but not opportunity costs (lost wages, childcare). The most effective intervention Aravind has found is same-day treatment at the vision center — but most interventions for referable diabetic retinopathy require equipment and expertise available only at base hospitals, not at peripheral screening sites.

Decentralizing treatment capability to the screening point — bringing laser treatment or anti-VEGF injection capability to vision centers, even on a periodic circuit-rider basis — would eliminate the referral gap for a subset of treatable conditions. For conditions that genuinely require hospital-level care, the design challenge is reducing the total patient burden of accessing treatment: bundling eye care with other needed health services (so the trip serves multiple purposes), providing compensation for lost wages (not just transportation), or deploying mobile treatment units that bring hospital-level care to the community on a predictable schedule. Aravind's own analysis points toward integrating screening into existing community health infrastructure rather than creating standalone screening programs.

A service design team could map the full patient journey from screening detection to treatment completion, identifying every decision point where patients drop out and what intervention at each point would be most effective. An HCI team could design a patient communication tool that explains asymptomatic-but-threatening findings in terms meaningful to rural patients with low health literacy — the challenge of communicating "your eye will go blind in 2 years but feels fine now." A health systems team could model the economics and logistics of circuit-rider specialist visits to vision centers, determining what visit frequency and treatment menu would capture the highest proportion of referable cases.