Artificial intelligence (AI) trained to identify warning signs in retinal images and clinical data can forecast the likelihood and timing of high-risk individuals, often referred to as "glaucoma suspects," progressing to actual glaucoma development, according to research published online in the British Journal of Ophthalmology.
The researchers suggest that, pending further refinement with larger study populations, this AI tool may become a valuable diagnostic resource for healthcare professionals. While recent advancements in AI have led to the development of algorithms for enhanced glaucoma detection, none have previously leveraged clinical indicators to anticipate disease progression among individuals at high risk.
Glaucoma stands as a leading global cause of blindness, presenting a challenge for physicians in identifying whether individuals displaying early signs of optic nerve damage, but lacking the typical diagnostic hallmark of elevated intraocular pressure (IOP), will eventually develop glaucoma and face the risk of vision loss.
In an effort to employ AI in bridging this gap, the researchers conducted an analysis of clinical data related to 12,458 eyes displaying initial signs of glaucoma suspicion.
From this cohort, their focus turned to 210 eyes that eventually developed glaucoma and 105 eyes that did not, all of which underwent regular monitoring at intervals of 6 to 12 months over a span of at least seven years.
During the monitoring period, red flag indicators in retinal images, combined with 15 critical clinical attributes, were employed to generate a series of "predictive" combinations. These combinations were subsequently input into three machine learning classifiers, algorithms designed for the automatic categorization of data.
The clinical attributes encompassed variables such as age, gender, intraocular pressure (IOP), corneal thickness, retinal nerve layer thickness, blood pressure, and body weight (BMI).
All three algorithms demonstrated exceptional performance and consistently achieved highly accurate predictions of glaucoma progression, as well as the timing of progression, with accuracy rates ranging from 91% to 99%.
Among the predictive clinical features, the three most crucial were the initial IOP, diastolic blood pressure (the second number in a blood pressure reading, measuring arterial pressure between heartbeats), and the average thickness of the retinal nerve fiber layer.
At the outset of the monitoring period, participants had an average age of 55, spanning a range from 33 to 76. While the baseline age did not prove to be a significant predictive factor, the researchers observed that those who eventually developed glaucoma had a notably lower average age compared to those who did not.
The researchers acknowledge several limitations in their study. For instance, the AI training data was derived from a relatively limited dataset, and the study exclusively encompassed individuals with normal intraocular pressure (IOP) who had not undergone any glaucoma treatment during the monitoring period.
"The current results, thus, demonstrate only that the built model works well for a limited range of patients,” the researchers warned. Still, "Our results suggest that [deep learning] models that have been trained on both ocular images and clinical data have a potential to predict disease progression in [glaucoma suspect] patients. We believe that with additional training and testing on a larger dataset, our [deep learning] models can be made even better, and that with such models, clinicians would be better equipped to predict individual [glaucoma suspect] patients' respective disease courses,” they concluded.
Ahnul Ha et al, Deep-learning-based prediction of glaucoma conversion in normotensive glaucoma suspects, British Journal of Ophthalmology (2023). DOI: 10.1136/bjo-2022-323167