Deep Learning in the Detection and Prediction of Hydroxychloroquine Maculopathy

Deep Learning in the Detection and Prediction of Hydroxychloroquine Maculopathy

Hydroxychloroquine retinal toxicity affects a significant number of patients using this medication. Detection of toxicity is difficult in the early stages of the disease and depends on the subjectivity of the clinician who reads the tests (optical coherence tomography, autofluorescence and visual fields). Automating the reading of these diagnostic exams could lead to earlier detection of this pathology and reduce the burden associated with interpreting these exams in the ophthalmology service. The images that are usually taken in the screening and monitoring of hydroxychloroquine toxicity by will be collected - photography of the ocular fundus and optical coherence tomography with autofluorescence.

HCQ is one of the most prescribed drugs worldwide. Originally developed for the treatment and prevention of malaria, it was soon proven effective for several other non-related disorders, most commonly non-organ specific autoimmune diseases. The use of HCQ in patients with systemic lupus erythematosus (SLE), for example, is estimated at 50%, increasing to 90% in specialized centers. These numbers are likely to increase following the LUMINA study that found a survival benefit of SLE patients on HCQ. In recent years, HCQ indications have been expanding to other medical areas such as dermatological disorders and oncology. Adding to the growing list of clinical indications, there are other factors that increase HCQ prescription such as a favourable safety profile and the possibility of adjunctive use with primary therapies, leading to a growing confidence amongst physicians to start treatment.

But HCQ is not an innocuous drug. There are several HCQ related adverse effects, either from acute or chronic intake, in particular in the nervous and cardiovascular system. It is estimated that around 1.8-7.5% of patients with more than 5 years of HCQ therapy suffer from retinal toxicity, with prevalence increasing to 20% after 20 years. Other known risk factors for toxicity are a higher dose, kidney failure and concomitant tamoxifen use and it has been suggested that previous macular pathology and genetic factors should be taken in account when calculating disease probability. Due to increase in HCQ usage, HCQ maculopathy and subsequent screening has become a public health problem. Although toxicity is directly related to drug use, retinal degeneration might continue despite drug cessation, which adding to the increase of users empathises the need for early disease detection. HCQ maculopathy screening guidelines differ slightly from country to country and have been evolving over time. The American Academy of Ophthalmology recommends baseline examination with OCT, autofluorescence and visual fields with annual review after 5 years unless there are other risk factors. On the other hand, the Royal College of Ophthalmology currently suggests annual monitoring 5 years after drug therapy with OCT and Widefield FAF unless there are other risk factors.

Deep learning techniques are paving the way in image-centric specialities and promise to lower healthcare costs and increase accuracy when compared to current methods. In ophthalmology, systems are being developed for the detection of diabetic retinopathy, glaucoma and age-related macular degeneration with high sensitivity and sensibility. We believe that using RETINAI technology and a sequential analysis of HCQ toxicity patients, a higher prediction model could be achieved. Three blinded readers will validate data as controls or toxicity.

Inclusion Criteria:

• Patients with > 10 years of HCQ intake

Exclusion Criteria:

• Patients with ocular diseases that might mimic HCQ maculopathy or interfer with HCQ maculopathy screening