Tremendous growth of biomedical data and the rapid development of IT-related technologies are contributing to significant advances in the field of medicine. With the help of deep learning and machine learning, clinicians can distinguish normal and abnormal patient data, assisting in the early prediction and diagnosis of medical conditions more efficiently. According to a study published in the British Journal of Ophthalmology, deep-learning models can be used for better prediction of conversion to normal-tension glaucoma (NTG) in normotensive glaucoma suspect (GS) patients.
The study conducted by Dr. Ha and colleagues reviewed 12,458 eyes with GS. The study included 210 eyes that were followed up for a minimum of 7 years. During this follow-up period, the intraocular pressure (IOP) was consistently below 21 mm Hg. Out of these eyes, 105 showed NTG conversion, while 105 did not. The study employed a convolutional autoencoder to extract features from two fundus images: the optic disc photography and the red-free retinal nerve fiber layer (RNFL) photography. In addition, 15 clinical features, including age, sex, IOP, spherical equivalent, central corneal thickness, axial length, average circumpapillary RNFL thickness, systolic/diastolic blood pressure, and body mass index, were used to predict NTG conversion. In this study, three machine-learning classifiers (XGBoost, Random Forest, and Gradient Boosting) were used with different feature combinations to predict NTG conversion.
The study showed that all three algorithms provided high accuracy in predicting NTG conversion. The areas under the curve ranged from 0.987 to 0.994, with the best performance achieved by XGBoost trained with both clinical features and fundus images. The XGBoost algorithm also provided the best prediction performance for time to NTG conversion, with a mean squared error of 2.24. The most important clinical features for predicting time-to-conversion were baseline IOP, diastolic blood pressure, and average circumpapillary retinal nerve fiber layer thickness.
Seo and Cho conducted a similar study on the classification of early NTG and GS using Bruch’s membrane opening-minimum rim width and RNFL. The study found that the Deep Neural Network (DNN) model had outstanding diagnostic performance, just like the other machine-learning models, in classifying either early NTG or GS. The DNN model had the highest AUC (0.966) compared to other machine-learning models (0.927–0.947).
The study findings highlight the potential of deep learning models as valuable tools in assessing and forecasting the progression of GS patients, specifically in predicting the conversion to NTG. These findings represent a significant advancement in the field of glaucoma diagnosis and may lead to improved patient care and outcomes.
References
- Ha A, Sun S, Kim YK, Jeoung JW, Kim HC, Park KH. Deep-learning-based prediction of glaucoma conversion in normotensive glaucoma suspects. Br J Ophthalmol. 2023 Nov 2;bjo-2022-323167.
- Seo S byul, Cho H kyung. Deep learning classification of early normal-tension glaucoma and glaucoma suspects using Bruch’s membrane opening-minimum rim width and RNFL. Sci Rep. 2020 Nov 4;10(1):19042.