03 September 2025
Research led by Moorfields Eye Hospital and University College London Institute of Ophthalmology has developed an artificial intelligence (AI) system for identifying early signs of eye damage in patients taking hydroxychloroquine.
The study, which was published in Ophthalmology Retina, highlighted that the technology was able to flag patients who would go on to develop retinopathy on average two and a half years before diagnosis.
Hydroxychloroquine is a medication commonly used to treat rheumatoid arthritis, lupus and other autoimmune conditions.
According to British Ophthalmological Surveillance Unit data from 2022, there are an estimated 77,000 long-term users of hydroxychloroquine in the UK.
The AI algorithm, which is called HCQuery, was trained on more than 8000 optical coherence tomography scans from 409 patients in the UK and US.
The algorithm’s effectiveness was demonstrated across different populations, including people who self-reported their ethnicity as Black, White and Asian.
Lead author, Dr Peter Woodward-Court, of Moorfields Eye Hospital, highlighted that the research could support a new approach to screening that would improve the care of people using hydroxychloroquine.
“Early detection would prevent irreversible vision loss while allowing patients to continue benefiting from this important medication for longer,” he said.
In the next phase of research, scientists will assess how the algorithm performs in a real-world setting and how the current care pathway could be improved to aid the earlier detection of hydroxychloroquine retinopathy.
Comments (1)
You must be logged in to join the discussion. Log in
Don Williams04 September 2025
The HCQuery study demonstrates how artificial intelligence can move beyond human limitations in hydroxychloroquine retinopathy screening. Using a convolutional neural network (EfficientNet-b4), the system ingests raw OCT B-scans and outputs a likelihood-of-retinopathy score (LRS), achieving external validation with sensitivity and negative predictive value of 1.00. Most strikingly, longitudinal signal analysis revealed HCQuery could flag risk almost three years before clinical diagnosis, effectively serving as a prognostic biomarker.
Deep learning differs from conventional algorithms by leveraging end-to-end feature learning, hierarchical pattern recognition without manual segmentation or handcrafted inputs mitigating observer variability and enhancing robustness across devices. This capacity to generalise across institutions and acquisition protocols points to genuine clinical scalability.
The role of AI here is not replacement but augmentation: automating triage, reducing false reassurance from unreliable fields and standardising interpretation where subtle outer retinal changes might be overlooked. With appropriate regulatory pathways, calibration across phenotypes and integration into EPR/PACS, HCQuery-style models could be embedded into community optometry as software-as-a-medical-device (SaMD).
The trajectory is clear here. AI-driven risk scoring is shifting from proof-of-concept to a clinical safety net, enabling earlier intervention and protecting patients before irreversible structural damage occurs.
ReportLike0