(Image credit: AdobeStock/maaramore)
Reviewed by Nitish Mehta, MD
Cutting-edge advances such as artificial intelligence (AI) and machine learning (ML) are part of the retinal imaging process, according to Nitish Mehta, M.D., Ph.D., a retinal specialist at New York University (NYU) Langon Eye Center in New York City. It is said that it is becoming a department. He is thrilled with both the technology itself and the description of how it works.
AI: concepts
It was originally defined as “the notion that any aspect of learning or characteristic of intelligence can be described so precisely that it can be simulated by a machine”.1 AI systems in medicine can be designed to learn from clinical and imaging data in annotated datasets and assist in treatment, Mehta said.
But AI comes in many flavors, he notes. Machine learning, as he defines it, is “a set of statistical techniques that allow you to do more than you would normally do with a simple statistical model” and is most relevant to retinal research. Also, Deep His Learning, a subset of His ML that is particularly good at learning from images (computer vision), demonstrates its utility in analyzing retinal images. “These tools may give us more insight and predictive power,” he said.
make good use of tools
It is well known that there are not enough diabetics in the United States to see an eye doctor in a timely manner for a variety of reasons. To streamline screening, programs have been developed for remote analysis of retinal images by trained human readers, e.g., individuals who should come to the clinic for face-to-face assessment and management of diabetic retinopathy. Identify
Many institutions and health systems, including New York University Langone Health, have such programs. However, AI leaders could replace human leaders, saving time and money. AI models have demonstrated the ability to classify images as consistently as, and perhaps more consistently than, human readers. Investments in AI readers could reduce the burden on retinal screening programs and enable their expansion.
A natural next step would be to extend beyond screening to diagnostic grading. AI is already being used to segment fluid volume in diabetic macular edema (DME) and neoangiogenic age-related macular degeneration (AMD), so the algorithm can be used to differentiate between disease stages (mild, moderate, and severe diabetic retinal degeneration). There is reason to hope that it will be possible to determine the severity of the disease (AMD), potentially determining the severity of the disease in stages.
Taking this further, could AI itself teach observers new ways to classify patients based on retinal images and clinical outcomes? This is an active area of research, but the holy grail of AI in this field. is probably a guide for treatment and prognosis.
For example, Mehta notes that clinical trials provide doctors with a ton of recommendations and predictions. Could an AI system synthesize such data to provide a patient prognosis? Are there features we haven’t yet used that might be of clinical value, such as small changes in pixelation of retinal images? can be used to predict the direction of spread of geographic atrophy in patients with advanced AMD?
Finally, Mehta predicts synergies between AI screening algorithms and the future introduction of optical coherence tomography at home, detecting disease activity without manual supervision in home OCT, and predicting seamless outcomes in converted patients. We envision an AI-assisted platform that can provide timely identification in From dry AMD to wet AMD. These are only theories at the moment, but ophthalmologists hope they will become practical in the future.
AI timeline
Fundus photography screening is the most advanced application in terms of development and implementation. The FDA has approved two of his devices, IDx-DR and EyeArt, for use in primary care to identify diabetic retinopathy requiring referral. Automated screening could prioritize these patients earlier or more frequently than standard methods, potentially improving outcomes across the population. However, the ability of AI to grade diseases and recommend treatments is still in the research stage.
Claim
In 2021, the American Medical Association released a new Current Procedural Terms (CPT) code that will allow clinicians to bill government and private insurance companies for their use of AI services. CPT Code 92229 refers to retinal imaging performed for disease detection and point-of-care reporting through automated analysis. As of late 2022, the Centers for Medicare and Medicaid Services have established national pricing for this code.2 “The hope is that payers will continue to embed this code so we can see this… AI tool,” Mehta said. [being used]… in the practice of retinal imaging. ”
Impact of AI on Outcomes
Mehta also gave examples of how AI can benefit patients. If a person with retinal vein occlusion has been treated with monthly anti-VEGF injections for about 6 months, a doctor may consider treatment if necessary. Clinical trial data suggest that, on average, these patients receive fewer injections later in treatment. However, there may be nuances in the data that AI/ML models may reveal.
Mehta described a New York University study in which clinical data and images from the COPERNICUS and GALILEO trials were fed into the ML algorithm. The researchers wanted to know if there were any clinical or imaging biomarkers in his first six months that could predict outcome during the on-demand treatment period.
“Results from the ML model suggest that patients with greater central subfield thickness (thicker retina) at baseline or week 4 may be more likely to receive two or more injections during the later PRN treatment phase. was interesting in that it was high (from week 24 to week 52),3 He said.
The ML model provides actionable clinical information to help physicians tell patients with thicker retinas at baseline that they are more likely to need more injections in year 1 than in year 2 and beyond I was able to. “As this research continues to advance, we may see lessons learned from AI being translated into clinical practice,” added Mehta. [in terms of treatment guidance]”
technical limit
However, as Mehta makes clear, the effectiveness of these models is determined by the data you provide and the questions you ask. For example, if the model is trained on a particular patient subset and then the model is used on another subset, the results may not be the best. This was most famously demonstrated by Google Groups, who had very poor performance and predictive results when applying a recently developed diabetic retinopathy AI screening program abroad.Four
Another concern concerns the explainability of the model, or how the results are derived, and Mehta called for transparency regarding model development. There are also regulatory issues that may affect patient outcomes. Cost-effectiveness and how the system is integrated into the clinical workflow are additional considerations. Finally, ethical issues are very important in AI, Mehta said, as well as ensuring that biases are not introduced into the dataset that could affect real patients.
future
Mehta believes AI image processing reveals previously unpredicted features, so doctors’ ability to answer isn’t necessary. “Our retinal imaging contains a large set of biomarkers. Perhaps there are things we haven’t seen yet,” he said, noting that AI imaging will enable interprofessional Effective collaboration could also be possible, he added. “It’s what they need that might help them treat their patients by uncovering neurodegenerative disease and cardiovascular risk.”
For example, the difficult clinical question of choosing between medication and surgery can be guided by AI-based outcomes. “And most importantly, the route a patient takes to our retinal clinic will be guided by his AI-based screening platform,” concluded Mehta.
