AI in Healthcare: AI in Diabetic Retinopathy Exams & Benefits of Human Interpretations
Like many other industries, artificial intelligence (AI) has great potential to transform medicine. In fact, diagnostic eye imaging is one area where AI shows promise in allowing earlier, accurate diagnoses. Utilizing AI in healthcare creates the potential to make a diagnosis of diabetic retinopathy quicker and more efficient, but there is also potential for significant limitations if proactive measures are not taken into account.
New advances in artificial intelligence and retinopathy
In recent years, AI technology in healthcare has been a driver in innovation, developed to diagnose various retinal diseases that previously were not as effectively detected. Currently, there are two platforms that the FDA has cleared for autonomous identification of diabetic retinopathy from fundus photos through leveraging AI in healthcare software. In short, this allows for the autonomous AI to make an immediate determination without input from a physician – expediting the process and reducing the potential limitations or errors.
Limitations of AI in Healthcare
While advances in ophthalmic AI technology are promising, there are limitations to take into account. Let’s dive into some of the most prominent ones in eyecare & diagnoses of diabetic retinopathy-
Results are formatted as either “refer” or “no refer.”
Results that are labeled as “refer” are defined as having more than mild retinopathy. However, the specific stage is not defined. Any presence of diabetic retinopathy should be documented to prevent future vision loss. Still, there is an urgency behind late-stage retinopathy that is not communicated with AI solutions’ results.
“No refer” results are defined as having detected no DR or mild DR. While this is a proper diagnosis for a patient with no DR, it’s a missed opportunity to identify and monitor cases of mild DR before it worsens.
Limited Equipment Options
The FDA clearance for autonomous read programs is contingent upon using one of two specific tabletop cameras, leaving no room for flexibility in hardware selection.
“One of the challenges we’ve seen with autonomous read solutions is that when it’s applied to a camera, it wasn’t developed on, the accuracy diminishes,” explained Ron Gross, MD.
Many healthcare organizations today opt to purchase handheld cameras for two reasons: price and portability. These organizations rely heavily on taking the camera to the patient; mobility is a crucial feature handheld cameras were designed for. Handheld fundus camera prices are also a fraction of tabletop cameras, with pricing starting around $16,000. That said, these easy, portable, and inexpensive devices are not approved to be used in an autonomous read solution.
Higher-quality images are a necessity.
A quality image is one where the macula, optic nerve, and inferior arcades can be easily identified. Not to mention, AI solutions require a higher quality image to produce a diagnosis confidently. If the image does not adequately meet the AI’s gradable requirements, it will be marked as ungradable. As a result, gradability rates among AI solutions are lower than those where a human grader provides the diagnosis.
In fact, IRIS maintains a network of 120+ board-certified ophthalmologists and retina specialists who grade exams for the IRIS Reading Center (IRC). Though high-quality images are always preferred, IRC physicians can also rely on years of medical expertise and practice to return a diagnosis on some images that AI solutions would have deemed ungradable.
Benefits of human graders
Ability to notate additional pathologies
DRE solutions that utilize human physicians for the diagnostic component can offer insight into the patient’s health outside of diabetic retinopathy. It may be surprising to learn how many diseases or health indicators can be discovered through retinal imaging! While the prevalence of diabetic retinopathy is growing, it’s not the only sight-threatening disease that can be detected through retinal imaging.
Other leading causes of blindness, such as glaucoma, macular degeneration, and wet/dry AMD, can also be noted when organizations choose a DRE program with human grading physicians. For example, the IRIS solution allows the grading physician to diagnose all stages of diabetic retinopathy (mild – proliferative) as well as all stages of macular edema (mild – severe) and notate additional “suspected” conditions.
With an AI solution, “None of the other suspected conditions that a human grader can notate would be recognized or reported,” Louis Morrow, IRIS VP of Sales.
Here’s more on what AI in healthcare technology can overlook –
- Risk-Adjusted Factor (RAF) diagnoses are missed. Even within the diabetic retinopathy spectrum, an AI solution does not explicitly identify patients with mild diabetic retinopathy. Though treatment may not be necessary for these patients, a diagnosis indicates to the primary care provider that the patient’s diabetes is not adequately controlled.
“Of the 250,000+ cases of DR diagnosed using the IRIS Solution, mild DR diagnoses are the most prevalent, and all of those diagnoses would have been missed using an AI solution,” explained Ron Gross, MD.
All stages of DR are RAF-eligible diagnoses, including mild. The absence of a mild DR diagnosis will leave physicians unable to adjust risk scores appropriately.
- Select the equipment that works for your organization. Human physicians can diagnose disease from retinal images regardless of the fundus camera used to capture them. A benefit that enables organizations to shop around for a camera that will help them achieve the highest level of success. DRE providers like IRIS offer clients a variety of fundus camera options ranging in price and portability.
Choose a solution that comes with human graders.
The use of AI in healthcare has become increasingly present, but there are still many instances where a human can do the job better. The IRIS Solution recognizes the importance of the human grader element through the IRIS Reading Center (IRC). A team of 120+ board-certified interpreting physicians makes up the IRC, issuing diabetic retinopathy and macular edema diagnoses for IRIS clients. The IRC physicians can denote other suspected conditions outside of DR and macular edema, such as wet/dry AMD, glaucoma, tumors, melanoma, and more.
The promise of an instant diagnosis from autonomous read programs can be an intriguing concept. And, while the IRIS Solution is not instantaneous, clients are still impressed with the quick turnaround time. They feel confident that the diagnosis is accurate and specific because a board-certified ophthalmologist or retina specialist issued it. Though results are instant, an AI Solution will never be able to create a sense of urgency on behalf of a patient.
Not to mention, IRC physicians sometimes come across images that display an emergent health situation (eminent stroke, tumor, etc.). In those instances, they have contacted the right people directly to ensure the patients received immediate care – a priceless benefit of using a solution with human physicians rather than relying solely on healthcare AI technology.
AI in Healthcare Conclusion
While AI in healthcare holds intriguing promise in the world of eye care, the current landscape has many crucial limitations. Choosing a solution with human grading physicians means you’re doing right by your patients, improving quality scores, and saving money on equipment costs.
Want to learn more about how you can integrate a blended approach, leveraging both human and AI healthcare solutions into your practice? Contact us.
Get started with IRIS today.
Want to know if IRIS is right for you? Schedule a one-on-one consultation with our team. We’re here to help.