AI-guided OCT triage cuts unnecessary diabetic macular oedema referrals by 45 percentage points in Hong Kong randomised trial

A team at The Chinese University of Hong Kong has inserted an artificial-intelligence system that reads optical coherence tomography scans as a second-stage filter in a territory-wide diabetic retinopathy screening programme, then tested it in a multicentre randomised non-inferiority trial. In results published in JAMA on 21 July 2026, false-positive referrals for diabetic macular oedema fell from 69.1 per cent under standard practice to 24.1 per cent with the system in place, while sensitivity remained at 100 per cent.

The system, developed at the university, combines automated image-quality assessment, detection of diabetic macular oedema and an explicit uncertainty flag that hands borderline scans back to a human reader.

The work ran in two stages. The first was a prospective silent-mode validation carried out between February 2020 and July 2023 in 603 patients with diabetes attending a tertiary hospital triage unit. Their mean age was 64.7 years, standard deviation 9.4, and 56.2 per cent were male. Of 1,200 scans, 86, or 7.2 per cent, were ungradable; of the 1,114 gradable scans, 49, or 4.4 per cent, were flagged as uncertain. Sensitivity was 98.8 per cent, 95 per cent confidence interval 94.5 to 100.0, and specificity 90.7 per cent, 95 per cent CI 88.7 to 92.4.

The second stage was a multicentre randomised non-inferiority trial that ran from September 2023 to April 2025, with follow-up complete in May 2025. It enrolled 276 patients with suspected diabetic macular oedema referred from Hong Kong's territory-wide screening programme. Of these, 137 were assigned to the intervention arm, in which referral depended on both the fundus-photograph screening report and the AI-OCT report, and 139 to the control arm, in which a fundus photograph alone triggered automatic referral. For ethical reasons every participant ultimately received specialist evaluation. The trial is registered with the Chinese Clinical Trial Registry as ChiCTR2300075087. Mean age in the trial population was 63.9 years, standard deviation 10.9, 54.7 per cent were male, and disease prevalence was similar in the two arms, at 30.9 per cent and 29.9 per cent.

The primary outcome was the false-positive referral rate, assessed against a pre-specified non-inferiority margin of 20 per cent. It was 24.1 per cent, 95 per cent CI 14.6 to 37.0, in the intervention arm and 69.1 per cent, 95 per cent CI 61.0 to 76.1, in the control arm, an absolute difference of minus 45 percentage points, 95 per cent CI minus 58.2 to minus 31.9, with P less than 0.001 for non-inferiority. Sensitivity for referral of diabetic macular oedema was 100.0 per cent in both arms. Specificity was 86.5 per cent, 95 per cent CI 79.3 to 92.9, with AI-OCT, against 0.0 per cent in the control arm. No case of diabetic macular oedema occurred among intervention participants who were not referred.

Senior and corresponding author Carol Y. Cheung is in the Department of Ophthalmology and Visual Sciences at The Chinese University of Hong Kong and holds US patent 12,361,697 B2 on the method. Co-authors include Tien Yin Wong of Tsinghua University and the Singapore Eye Research Institute, Sobha Sivaprasad of the NIHR Moorfields Biomedical Research Centre in London, Pheng-Ann Heng of the university's Department of Computer Science and Engineering, and Clement C. Tham and Chi Pui Pang, also of The Chinese University of Hong Kong. Recruiting sites were Hong Kong Eye Hospital, Prince of Wales Hospital, United Christian Hospital and Tseung Kwan O Hospital.

The authors conclude: "Compared with standard practice, incorporation of the AI-OCT system as a secondary screening tool was noninferior with respect to false-positive referral rates and was associated with a substantial reduction in potentially unnecessary DME referrals without compromising sensitivity." The paper appears in JAMA 2026;336(3):215 to 223, under DOI 10.1001/jama.2026.7025, with PubMed identifier 42295755.

The path the authors identify is integration into Hong Kong's territory-wide diabetic retinopathy screening pathway. They position AI-OCT as a second-stage filter rather than a replacement for fundus-photograph screening, and they note that diabetic retinopathy screening programmes worldwide are constrained less by detection than by the specialist capacity consumed by false positives.