An Artificial Intelligence-Based Prognostic Model for Prediction of Functional Glaucoma Progression from Clinical and Structural Data

Abstract

Purpose : Several pieces of information are available for clinicians to monitor glaucoma progression; integration of various sources of information for prognostic models is an unmet need in glaucoma diagnostics. We designed a deep learning-based prognostic model incorporating available clinical and structural data for predicting functional glaucoma progression.

Methods : We included 2,077 eyes (1,176 patients) with ≥5 24-2 visual fields (VF) and ≥3 years of follow-up. VF mean deviation (MD) rates of change were estimated with linear regression. VF progression was defined as a confirmed negative MD slope with p<0.05 at the final follow-up. A convolutional neural network pre-trained on ImageNet was designed to predict VF progression using clinical data, baseline disc photographs, and OCT-derived global and sectoral retinal nerve fiber layer, and macular thickness measurements. The following baseline clinical/demographic data were put into the DL model: gender, ethnicity, age, intraocular pressure, central corneal thickness, and VF MD and pattern standard deviation. A separate deep-learning model was trained for every combination of the clinical/demographic data and the three structural imaging modalities.

Results : Average (SD) baseline MD and number of VF exams were –3.6 (5.1) dB and 12.6 (8.5). 637 eyes (31%) deteriorated. The mean (SD) follow-up time for stable and progressing eyes was 7.8 (4.9) and 10.4 (5.0) years. Table 1 shows model performance utilizing different structural modalities along with demographic/clinical data for the prediction of VF progression. The best-performing model was the one using baseline ODP, and RNFL and macular OCT measurements (AUC= 0.863; 95% CI: 0.793-0.934) (p <0.015 for all comparisons to simpler models except comparison to the combined ODP and macula OCT; p=0.199). Figure 1 displays the ROC curve for the top 3 AI models.

Conclusions : Our newly designed deep learning model is able to combine baseline demographic and clinical data with widely available structural measurements and provides clinically relevant information for the prediction of glaucoma progression many years later.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.