Retina specialists are distinctively efficient. We’re used to managing dozens of patients while synthesizing tremendous amounts of data to inform decision-making. Image interpretation is a core strength and a key value we bring to our patients.
Most of our optical coherence tomography interpretation is qualitative. Is there fluid? Is there more fluid? Or less? Is the pigment epithelium detachment enlarging? Is the atrophy expanding?
This approach appears to be on the edge of a sea-change. Swiss imaging expert Marion Munk, MD, PhD, reported at EURETINA 2023 that she routinely uploads clinical images into an artificial-intelligence system that quantifies multiple OCT-based biomarkers, including fluid volumes and areas of geographic atrophy. She uses AI to guide daily management.
Such analyses should expedite our clinical flow. Software that can determine and quantify disease states will allow us to focus on more sophisticated management decisions and patient communication. Right?
Not so fast. The Solow paradox, named for economist Robert Solow, suggests that despite advances in information technology, a counter-intuitive delay or even negative impact on productivity growth can follow. For example, in the United States through the 1970s and 1980s, despite rapid developments in computing and information technologies, productivity gains slowed measurably. Absolute gains only emerged 10 to 20 years later when the U.S. economy began to reap the benefits of these foundational technologies.
This paradox has been observed in radiology,1 for example, in situations where AI has been implemented with high sensitivity at the expense of specificity. Radiologists then may face the dilemma of disagreeing with AI and either hedging, potentially leading to unnecessary interventions or having to over explain their reasoning for fear of medical-legal consequences.
In retina, how will our efficiency change when AI detects fluid in an OCT volume scan of a patient with intermediate age-related macular degeneration that you’re convinced doesn’t have neovascularization?
There’s also the likely scenario that AI-assisted population screening will meaningfully increase our patient volumes. I recently comoderated a session that included a talk on an iPhone camera-based AMD screening tool. The threshold for patient referral was intermediate AMD regardless of symptoms or signs of advanced AMD. My comoderator, Clare Bailey, MD, made an astute point that left the audience silent: Most health-care systems will not be able to see all of these patients.
An analogous situation may emerge for patients with diabetic retinopathy identified through the many AI-enabled systems that have been cleared by or are being reviewed by regulatory agencies globally.
As we continue to integrate AI into our clinics, we must be ready to adapt to the associated, and sometimes paradoxical, practical implications and challenges. RS
REFERENCE
1. Jha S. Algorithms at the gate—radiology’s AI adoption dilemma. JAMA. 2023;330:1615-1616.
 |
