Written by Alind Gupta, Cytel; Haridarshan Patel, Horizon Therapeutics; and Jason Simeone, Cytel
Ophthalmology is well-suited to using artificial intelligence (AI) methods because clinical decisions often rely on complex data-rich information from medical images of the eye and patient health status. AI has revolutionized image data analysis over the last decade and is promising to improve healthcare delivery and clinical decision-making while reducing healthcare costs. In fact, some AI-based diagnostic platforms for early detection of diabetic retinopathy have received FDA clearance and have been introduced in resource-constrained healthcare settings worldwide to screen for early signs of disease and to accelerate patient access to correct therapies.1,2,3
How do AI methods work for imaging data?
The most common machine learning model for image classification is convolutional neural networks (CNN). A CNN is a type of neural network that learns a model to map images to labels, such as “severe disease” or “mild disease,” using clinician-labeled training data as input. The goal is that this trained model will then be able to accurately predict the label of new medical images once deployed in the clinic.
The CNN takes the data inputs and performs three operations:
-
- Convolution — Convolution is a mathematical operation that assists the computer to automatically learn important features of the image that may be helpful for prediction, such as curvature of the retina or choroidal thickness, without humans having to specify these features manually. It also enables the model to learn patterns related to the translational change in the position of image features, such as the location of the retina within the image.
- Pooling — This step down samples the result from convolutions, making them smaller and more accessible for the computer to store and process. This step also serves to remove noise from the image and keep only the most important information — a compact representation of the features of the image.
- Prediction — The final step uses as input this compact representation and learns parameters, similar to coefficients in a regression model, that can maximize accuracy on the training set.
To test this model, the CNN is provided with images from a test dataset that it has not encountered before, and its predictions are compared to those of radiologists or clinicians.
Potential uses for AI in ophthalmology
AI can help improve screening, early diagnosis, and monitoring in community practices to support specialty referrals and accelerate patient access to treatment to prevent disease complications and long-term adverse outcomes such as blinding. AI tools can also help inform prognosis to stratify patients by risk of adverse outcomes by predicting disease severity and long-term outcomes. Optical coherence tomography (OCT) is typically performed annually and provides non-invasive high-resolution images that enable early detection and staging of rare eye conditions, such as thyroid eye disease (TED), based on measurements of eye muscles and choroid tissue,4 and can improve monitoring of progression and treatment response over time. AI may be well-poised to effectively learn from subtle changes in the eye, and the ease and speed of OCT increases the practicality of using these images for AI models. Alternatively, orbital imaging with MRI or CT scans can provide detailed images of the eye socket structures and surrounding soft tissues and help assess changes in the eye muscles and eye fat; this can be used to identify inflammation and detect tissue expansion due to swelling, as well as monitor treatment response.
For rarer ophthalmologic conditions with a risk of misdiagnosis, incorporating data from imaging, patient health status and history, and lab tests or biomarkers in AI models may help improve specificity of diagnoses and risk stratification without requiring more expensive testing modalities or experienced subspecialty clinicians. For example, TED may be misdiagnosed as more prevalent conditions such as conjunctivitis due to the complex and heterogenous presentations of the disease. AI may predict this rare condition better and faster and improve time to prescription of effective therapies that prevent vision loss. Finally, AI can help recommend treatments by identifying patients who would respond well to specific therapies. Ultimately, such AI models will need to be incorporated into clinical decision support systems or imaging software that complement clinical workflows with extensive clinical validation and support of stakeholders. Cytel and Horizon are currently exploring a study that utilizes these methods to help patients with TED receive an accurate diagnosis earlier, increasing patient access to appropriate therapy.
Final takeaways
AI has the potential to revolutionize image data analysis and is a promising technologic innovation that could improve healthcare delivery and clinical decision-making while reducing healthcare costs. To take advantage of these benefits, rigorously trained AI models should be incorporated to complement clinical workflows.
Interested in learning more? Cytel’s Real-World and Advanced Analytics team will be at ISPOR Europe 2023 at Booth #C3-049. Click below to book a meeting with our experts:
Dr. Alind Gupta, PhD, is Research Principal in the Real-World and Advanced Analytics team at Cytel and adjunct lecturer in the Department of Epidemiology at the Dalla Lana School of Public Health at the University of Toronto. His work focuses on comparative effectiveness research, issues of generalizability of evidence, and machine learning.
Dr. Hari Patel, PharmD, PhD, is an outcomes researcher who specializes in generating real-world evidence to assist patients, providers, and payers in making informed decisions. With more than 12 years of experience in the pharmaceutical industry, Dr. Patel presently leads the Health Economics and Outcomes Research team at Horizon Therapeutics. Along with creating effective evidence-generation teams, he also conducts cutting-edge research utilizing synthetic control arms, trajectory analyses, and comparative effectiveness studies.
Dr. Jason Simeone, PhD, MS, is a Senior Director and leads Cytel’s North American Real-World Evidence team. He is a pharmacoepidemiologist, with extensive experience in conducting studies to generate real-world evidence using US claims and EMR data, as well as data sources globally. His research has focused on medication safety, effectiveness, burden of illness, and treatment patterns in a wide range of therapeutic areas, and he is particularly interested in the generation of real-world evidence for disease risk prediction and improvement of patient outcomes.
References
- Abràmo MD, Lavin PT, Birch M, et al. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. NPJ Digit Med 2018;1:39. 10.1038/s41746-018-0040-6; FDA permits marketing of artificial intelligence-based device to detect certain diabetes-related eye problems.
- Assessment of EyeArt as an Automated Diabetic Retinopathy Screening Tool, ClinicalTrials.gov.
- Gulshan, Varun, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA22 (2016): 2402-2410.
- Laplant J, Caplan C, De La Rosa S, Hubbell M. Applications of optical coherence tomography in thyroid eye disease. Investigative Ophthalmology & Visual Science. 2019 Jul 22;60(9):1889.
Read more from Perspectives on Enquiry and Evidence:
Sorry no results please clear the filters and try again
Global Planning to Local Execution Market Success
Writing a Successful Study Protocol for Real-World Evidence Studies
A Better Way to Track Medication Adherence
