Q&A: How an AI platform cut screen failures and saved 4,000 hours in a retina practice

At the American Society of Retina Specialists (ASRS) 44th Annual Scientific Meeting in Montréal, Québec, Canada, Louie Cai, MD, presented findings from a prospective evaluation of an AI clinical screening platform for retina clinical trials. Cai, a vitreoretinal surgeon and cofounder of Cosign AI—a digital health company focused on accelerating clinical trial recruitment—found that when the platform was folded into daily clinic workflow, trial randomization rose 37.5% relative to baseline, the screen failure rate fell by roughly 13%, and the practice saved an estimated 4,000 hours of manual chart review over the study period.

In the conversation below with Ophthalmology Times, Cai discusses why manual prescreening is so labor-intensive, how the study was designed, and what broader adoption of AI-assisted screening could mean for clinician time, patient satisfaction, and the pace of drug development.

Ophthalmology Times: Can you walk us through the presentation itself—what you set out to study, what you found, and why it matters?

Louie Cai, MD: Manual prescreening for clinical trials is very labor intensive, taking as long as 30 minutes to an hour per patient per trial. This is very mundane and very hard, arduous work. We developed an AI platform that could screen patient charts in real time to check their documentation and their imaging for eligibility for clinical trials.

So in this study, we evaluated, prospectively, how it affected clinical trial screening efficiency as well as the randomization efficiency. So we broke a single retina practice into 2 groups: 11 doctors using it and 11 doctors not using it. In the first 3 months, nobody was using it—It was our baseline period. In the next 3 months, we found that the doctors who did use our product had a 37.5% increase relative to their own baseline in randomization, as well as a 13% or so decrease in the screen failure rate.

We also found using a model has saved about 4,000 hours of human labor. That would have had to be done if they were not using the platform.

Ophthalmology Times: Can you go into the development of the program and what baseline you used to measure whether it was succeeding or failing?

Louie Cai, MD: Yeah. So I think the baseline is not using any software. So manual review, which is what happens in a lot of retina practices today. Many practices rely on their clinical research coordinators to preview the appointment list a day ahead of time and click through every single patient to see if they would be a good candidate.

We found that 87% of patients don't qualify for any clinical trial. They're just there for a routine checkup. And in those 13% that do, they have to then spend another 30 minutes to make sure that all the criteria match. And that can be a lot of work every single day, very repetitive work. And I think this is the perfect task for something that can be mundane and routine to be performed by an AI platform.

Ophthalmology Times: Looking ahead, if this were implemented across clinics, what would that mean for clinician time, patient satisfaction, and clinical trials?

Louie Cai, MD: For clinicians, they don't have to guess: clinicians spend about 5 to 10 minutes with the patient, and they don't have the time to go through all the criteria with them. And so they may accidentally send patients who are not ideal candidates for clinical trials.

What that does is delay their care. Patients end up finding out later down the road that they are actually not qualified for the trial, for whatever reason. Then they have to go back to the clinic and then be treated as a standard-of-care patient. And that can take weeks. So the care was delayed because of the lack of information.

Ultimately, as a whole, clinical trials and drug development can get accelerated and more efficient. And so instead of a year-long enrollment period, maybe it'll just be a 6-month enrollment period. And people can also screen their patients before the trial even starts. And so that you can have an idea of how many patients that you would likely be able to recruit at a given clinic.

Ophthalmology Times: Would this be applicable to any step of a trial?

Louie Cai, MD: Yeah. So we're primarily focusing on the clinic side of things. So as soon as phases 1, 2, and 3 reach the clinic side and need to get patients enrolled, that's where we can help with a prescreening platform like this. And many different platforms exist. But I think this area, as a field, is very amenable to this kind of technology to help.