An Artificial Intelligence-Based Prognostic Model for Prediction of Functional Glaucoma Progression From Clinical and...

Integration of various sources of information for prediction of disease progression is an unmet need in glaucoma diagnostics. We designed a deep learning-based prognostic model incorporating clinical and structural data for forecasting functional glaucoma progression and compared its performance to clinicians.

Retrospective, comparative cohort study of prognostic accuracy.

We included 1599 eyes (908 patients) with definite or suspected glaucoma with ≥5 24-2 visual fields (VF) and 3 or more years of follow-up.

VF mean deviation (MD) rates of change were estimated with linear regression. Sequential MD rates of change were estimated with each series spanning only 5 years of follow-up. VF progression was declared when four sequential statistically significant negative MD slopes were observed, and slope for the entire follow-up was significant. A convolutional neural network pretrained on ImageNet was designed to predict VF progression using baseline clinical and demographic data, disc photographs, and optical coherence tomography-derived global and sectoral retinal nerve fiber layer and macular thickness measurements. In addition, average intraocular pressure and treatment information during follow-up were put into the model. The same data for a subset of patients was provided to two clinicians to independently predict future progression. The model was validated on a separate cohort of eyes in which optical coherence tomography imaging was done with a different device (291 eyes).

Model's area under receiver operating characteristic curves (AUC), accuracy, and area under the precision and recall curves.

Average (SD) baseline MD and number of VF exams were -3.5 (4.9) dB and 10.1 (4.7). 399 eyes (25%) deteriorated. The best-performing model incorporated baseline disc photographs, and retinal nerve fiber layer and macular thickness: AUC, 0.839 (0.771-0.906), accuracy, 76.0% (62.0%-85.0%), and area under the precision and recall curves, 0.558 (0.385-0.733). Deep learning model significantly outperformed clinical graders (AUC : 0.629 [0.531-0738], P < .001 and 0.680 [0.584-0.776], P = .001, for grader one and two, respectively). Model performance was similar on the validation cohort (AUC: 0.754 [0.671-0.837], and accuracy: 77% [71%-82%], respectively, P = .122). The model performed well when predicting fast-progression, defined as MD rate <-1.0 dB/y (AUC: 0.869 [0.792-0.947]).

Our newly designed deep learning model can combine baseline demographic and clinical data with widely available structural measurements and provide clinically relevant information for the prediction of glaucoma progression.