Forecasting glaucoma from multimodal data using deep learning models

Abstract

Purpose : To assess the applicability of deep learning (DL) models to forecast glaucoma development based on multiple data modalities.

Methods : We developed a DL model to forecast glaucoma development from 3271 eyes of 1636 patients who participated in the Ocular Hypertension Treatment Study. The input data consisted of various combinations of fundus photographs, visual fields (VFs), clinical factors (such as intraocular pressure, central corneal thickness, and cup-to-disc ratio), and demographics (age, gender, race) from the baseline visit. We investigated the model based on FVCD (fundus + VF + clinical + demographic), FV (fundus + VF), FCD (fundus + clinical + demographic), VCD (VF + clinical + demographic), F (fundus), V (VF), and CD (clinical + demographic). Fundus features were extracted by a VGG19 architecture while VF and other input data were scaled and then concatenated before feeding to the multilayer perceptron (MLP) classifier (Fig.1). The models for single modality data were either a VGG19 (F modal) or artificial neural network (ANN) (V and CD modal). We evaluated the models by computing accuracy, area under the receiver operating characteristic curve (AUC), and weighted F1 score.

Results : On average, eyes developed glaucoma after 7.5 years. The FVCD and FCD model achieved the highest and comparable performance with AUC, accuracy and weighted F1 score of 0.87 (95% CI: 0.84-0.89), 79% (75-81%), 0.81 (0.79-0.83) in FVCD, and 0.86 (0.82-0.90), 78% (73-82%), 0.81 (0.77-0.84) in FCD, respectively. While the AUC and weighted F1 score in the other models were as follows: F-0.74 (0.68-0.79), 0.69 (0.66-0.73); V-0.68 (0.65-0.71), 0.69 (0.66-0.71); CD-0.77 (0.72-0.81), 0.74 (0.71-0.78); FV-0.81(0.78-0.84), 0.77 (0.75, 0.79), VCD-0.68 (0.61-0.74), 0.68 (0.64-0.73). The performance of FVCD and FCD was significantly higher than the other models (p<0.05 in AUC and weighted F1 score) (Fig.2).

Conclusions : The model integrating information from fundus, clinical and demographic factors provided a highly promising outcome for forecasting glaucoma from the baseline visit. This approach may prove useful for predicting glaucoma development in glaucoma research and clinical practice.

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

 

Figure 1. Flowchart of the multimodal model. MLP: multilayer perceptron.

 

Figure 2. Performance comparison of models based on single and multimodality data. F-fundus, V-visual field, CD-clinical factor and demographics, FV-F+V, FCD-F+CD, VCD-V+CD, FVCD-F+V+CD.