Abstract
Purpose :
Artificial intelligence (AI) models for predicting glaucoma outcomes have typically relied on single data modalities, such as data exclusively sourced from electronic health records (EHR) or imaging. Since clinical glaucoma evaluation entails integration of patient information in various forms, we hypothesize that combining structural data from retinal nerve fiber layer optical coherence tomography (RNFL OCT) scans with EHR data would improve prediction performance. The goal of this study was to develop multimodal AI models to predict which glaucoma patients progress to require surgical intervention using EHR and RNFL OCT data.
Methods :
Glaucoma patients from an academic center (2008-2022) were identified via encounter diagnosis codes. Structured EHR information included demographic and eye exam data (visual acuity, intraocular pressure, central corneal thickness, spherical equivalent from both eyes). RNFL OCT data included average RNFL thicknesses, cup-to-disc ratios, rim and disc areas, and quadrant thicknesses from both eyes. Employing TabNet, a novel deep learning architecture for tabular data, we integrated RNFL and EHR data to predict progression to glaucoma surgery in the coming year using data from 1.5 years prior. We compared the multimodal model to single modality models with either EHR or RNFL input features. Evaluation metrics on a held-out test set included area under the receiver operating curve (AUROC), recall (sensitivity), and precision (positive predictive value).
Results :
Of the cohort of 1472 glaucoma patients, 367 (29.9%) progressed to glaucoma surgery. The RNFL and EHR fusion model predicted patients’ progression to surgery with AUROC 0.827 (Figure), recall 0.897, and precision 0.388. In comparison, the single modality RNFL only model had AUROC 0.613, recall 0.948, and precision 0.240, while the EHR only model exhibited AUROC 0.742, recall 0.590, and precision 0.400.
Conclusions :
Innovative deep learning techniques to fuse EHR with RNFL OCT data may enhance the ability of AI models to predict glaucoma progression versus models trained solely on EHR or RNFL inputs. Future work could achieve further improvements in prediction performance by integrating additional data modalities, including visual fields and free-text clinical notes.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.