Pilot study on the use of AI to identify and pre-screen patients for an interventional Geographic Atrophy clinical trial

Joel Pearlman¹, Roger Goldberg², Albert Edwards³, Margaret Chang¹, Mark Barakat⁴, Nicholas Fuerst¹, Luis Monsalve², Ryan Lebien³ ; Neali Austin⁵;Yasmin McQuinlan⁶; Dipti Rao⁶; Pearse Keane⁷; Franziska Bucher⁸*: Blaise Thomson^6*

1. Retina Consultants Medical Group; 2. Bay Area Retina Associates; 3. Sterling Vision; 4. : Retina Macula Institute of Arizona; 5. Ora Clinical; 6. Bitfount; 7. : Moorfields Eye Hospital ; 8. Boehringer Ingelheim; *joint last author

Purpose

Artificial Intelligence promises to revolutionize clinical trials via increased patient yields, reduced screen failure rates and accelerated timelines. Despite the promise, there is limited published research on the idea in real trial settings.

Methods

In Boehringer Ingelheim’s JADE trial (clinicaltrials.gov ID NCT06769048), we compare performance with and without use of an AI image analysis tool. 4 of 46 sites were installed with Bitfount’s pre-screening tool leveraging Altris AI’s geographic atrophy (GA) models. Within those sites, the main screening location had the tool while other locations did not (Bitfount runs locally to maximise privacy so only had access to local images). 

The automation analysed routinely collected OCTs to assess GA lesion size, distance of GA from fovea, largest lesion size, presence of choroidal neovascularization and age. Patients matching the imaging criteria for the trial were presented to sites in spreadsheets, sites checked medical records for further trial criteria and continued recruitment following their standard practices. 

Sites also continued recruiting using standard practices (both at locations with and without the tool). This meant some patients in each site were identified with the tool and some without. Sites and recruitment approaches were compared on patient yield and screen failure rate.

Results

Sites using the tool enrolled a mean of 4.5 patients, vs 1.9 patients for sites (n=42) without [p=0.009, Poisson exact test]. Of the 17 patients enrolled at sites with Bitfount, 8 were marked as having been specifically identified by the tool.

Within sites that had the tool available, screened patients that had been pre-screened with the tool had a screen failure rate of 38% (n=13), vs 69% that hadn’t (n=28) [p=0.059, one-sided Fisher test]. The screen failure rate for sites not using the tool was 59% (n=295). 

Conclusions

This pilot study showed that an AI analysis tool limited to image analysis was able to roughly double the number of subjects enrolled by participating sites. Further, the AI tool may have reduced screen failure rates. We expect that joining image analysis with review of inclusion and exclusion criteria available in medical records would further accelerate recruitment. The study also identified several hurdles in implementation - the approach requires sites to adjust their recruitment practices and sponsors need to invest in translating their inclusion/exclusion criteria for AI models.

‍

_Disclosures

_{Boehringer Ingelheim was given the opportunity to review the manuscript for medical and scientific accuracy as well as intellectual property considerations in relation to potential mention of Boehringer Ingelheim substances.}

‍