Accelerating clinical trial recruitment via AI-enabled pre-screening for geographic atrophy

Abstract

Purpose : Clinical trials are resource-intensive endeavours, notorious for their high costs and duration. Recent advancements in AI have positioned it to transform this, promising significant benefits for trial sponsors and patients. With a particular focus on geographic atrophy (GA), a subtype of AMD, we show that our AI system can effectively surface individuals from a general eye hospital population based on trial-specific criteria using OCT images alone. Capable of functioning independently or in conjunction with an electronic health record (EHR) search, we consider this a promising exemplar for modernising ophthalmic trial recruitment more broadly.

Methods : Using a broad retrospective AMD dataset from Moorfields Eye Hospital, London, United Kingdom (225,652 eyes, 113,858 patients), we validated Octane - a deep-learning model which processes OCT scans to generate segmentations of the retinal tissue and predictions of a range of diseases. The wide array of outputs from Octane was used to produce comprehensive reports on an individual’s eligibility based on common GA trial criteria (Figure 1). Validation was based on criteria from the HORIZON trial (NCT04566445) and performed by an optometrist with over 11 years of experience and specialist training in OCT interpretation. As a baseline approach, we conducted a search of the EHR letters for the term ‘geographic atrophy’.

Results : In predicting eligibility for the HORIZON trial, the Octane system alone surfaced the highest number of eligible patients (1,235) with a positive predictive value (PPV) of 68%, compared to 657 patients with a PPV of 38% using the EHR search. Of those identified using the EHR search, Octane achieved a PPV of 89% and identified a total of 611 patients (Figure 2). In total we surfaced patients according to four different trial criteria, showcasing the ability of the model to extract valuable insights about an individual's atrophy on a global level (such as the total area) and a local level (location of GA lesion relative to the fovea).

Conclusions : We demonstrate the potential for AI in streamlining recruitment for clinical trials, helping trial sponsors avoid the waste of resources associated with a prolonged recruitment process. Following regulatory approval, similar AI systems could also be used to highlight individuals who may benefit most from receiving the intervention, such as those in the more nascent stages of the disease.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.