What are the benefits and challenges of introducing AI-based tools in clinical practice?
Click Here to Manage Email Alerts
Click here to read the Cover Story, "AI in retina moving toward real-world practice."
AI in retina moving toward real-world practiceAI has multiple benefits
The Notal OCT Analyzer is the AI segmentation algorithm that analyzes each of the daily OCT scans obtained by patients as part of a neovascular age-related macular degeneration remote-monitoring paradigm.
In a small pilot study of 15 patients with neovascular AMD performing the Notal Home OCT daily for 3 months, more than 2,300 scans required review. The first important message is this: No human will want to review the high throughput of data coming from daily OCT imaging. No human could do so with a high degree of fidelity in a sufficiently timely manner to generate alerts that retinal fluid had returned or worsened. Thus, the overwhelming output of clinical data our field will generate in the future will necessitate the use of AI.
The Fluid Monitor (RetInSight) has an AI segmentation algorithm that has EMA approval in Europe, with FDA approval being sought in the U.S. Presentations by European retina specialists using this technology in their clinics are impressive. Fluid quantification measured in nanoliters of fluid volume is validated to allow for expert level interpretation of the OCT at every office visit. This is the second important message: AI technology will provide data analytics that are superior to what the average retina specialist can do on a busy, long day. The first patient of the day will receive the same accurate, reproducible and reliable OCT analysis as the 100th patient that day without the potential intrusion of human error. The bar for AI validation is high, and I will bet on it over humans in the long run.
Finally, AI will allow for scientific research that informs us about the diseases we treat in ways that were not possible before. Scientists at Roche have trained and validated an AI algorithm that quantifies both number and volume of hyperreflective foci (HRF) in patients with diabetic macular edema. At the 2023 Association for Research in Vision and Ophthalmology meeting, Schulthess and colleagues presented data comparing clearance of HRF from the central macula of patients with DME treated with anti-VEGF monotherapy vs. dual-targeted therapy of anti-VEGF and anti-Ang2, showing better, faster clearance with the latter treatment. Would a human have been able to review 49 OCT B-scans and calculate HRF volume? This is the third important message: AI’s promise to outperform the human eye and human brain is here — to the benefit of our profession.
- References:
- Liu Y, et al. Ophthalmol Retina. 2022;doi:10.1016/j.oret.2022.02.011.
- Maunz A, et al. Automated segmentation of hyperreflective foci in diabetic macular edema shows greater volume reduction by faricimab vs aflibercept in phase 3 YOSEMITE and RHINE. Presented at: Association for Research in Vision and Ophthalmology meeting; April 23-27, 2023; New Orleans.
Nancy M. Holekamp, MD, is director of retina services at Pepose Vision Institute in St. Louis.
AI presents unique challenges
Implementing AI in ophthalmology clinics presents unique challenges, particularly in the realm of data integration and interoperability.
One significant issue is the absence of universally accepted standards, akin to the Fast Healthcare Interoperability Resources, which enable seamless communication between clinical information systems and ophthalmic imaging picture archiving and communication systems. This gap hinders the effective exchange and utilization of valuable patient data and imaging, which are crucial for the accurate diagnosis and treatment of eye diseases.
The integration of AI in ophthalmology clinics is further complicated by current FDA regulations, which necessitate a strict 1:1 pairing between AI models and specific imaging devices. This regulation implies that an AI system can only be used if the clinic possesses the exact camera model for which the AI was developed and approved. Such a requirement significantly limits the versatility and applicability of AI tools in diverse clinical settings, especially in clinics that may not have the resources to acquire or upgrade to compatible imaging equipment. This constraint not only hampers the widespread adoption of AI in ophthalmology but also restricts the potential benefits it can offer in enhancing diagnostic accuracy and patient care.
Moreover, there is a notable scarcity of head-to-head studies comparing different AI models in ophthalmology. Such comparative studies are essential for clinicians to understand the expected performance of AI systems in real-world scenarios. Additionally, these studies are crucial in addressing and mitigating concerns related to bias, particularly with respect to race and ethnicity. The lack of comprehensive research in this area raises questions about the fairness and inclusivity of AI tools, as biases in AI algorithms can lead to disparities in diagnosis and treatment outcomes among diverse patient populations. To build trust and ensure equitable health care, it is imperative to conduct rigorous evaluations of AI systems, examining their performance across various demographic groups. This will not only enhance the understanding and acceptance of AI among clinicians but also ensure that these advanced tools contribute positively and fairly to patient care in ophthalmology.
Aaron Y. Lee, MD, MSCI, is an associate professor in the department of ophthalmology at the University of Washington in Seattle.
Collapse
Click Here to Manage Email Alerts