Faricimab treatment response prediction in diabetic macular edema (DME) using deep learning on optical coherence tomography (OCT) and clinical data

Abstract

Purpose : To predict 1-year best-corrected visual acuity (BCVA) from baseline (BL) OCT and clinical data in patients with DME treated with faricimab using deep learning.

Methods : Eligible patients enrolled in YOSEMITE (NCT03622580) and RHINE (NCT03622593) clinical trials, treated with faricimab every 8 weeks or personalized treatment interval arms, and documented with Heidelberg OCT scans were included. Patient data were split into development data (N=621) with 5-fold cross-validation (CV) and test data (N=151) via anticlustering using Diabetic Retinopathy Severity Scale, hemoglobin A1C, BCVA, and central subfield thickness at BL. RETFound foundation model was explored with OCT images only and multimodality data. B-scans within the center 1mm area of the OCT volume were independently fed into the RETFound model to predict 1-year BCVA (average BCVA letter score across weeks 48, 52, and 56). Fine-tuning protocols¹ were followed with minor adaptations to the regression task to minimize the mean absolute error (MAE) loss. Test data performance was averaged across models trained in CV (Fig 1). For comparison, a ResNet model pretrained on ImageNet 1K was fine-tuned and underwent similar hyperparameter tuning and evaluation. For the multimodality approach, clinical data (variables listed above, sex, age, diabetes type, and treatment-naïve label) were concatenated to an intermediate representation learned by the model from images, and a prediction head was formed by a multilayer perceptron with 1 hidden layer.

Results : The multimodality approach had the highest performance: MAE, 5.95; R², 0.415. Using only images, RETFound demonstrated a comparable or slightly better performance than a standard pretrained ResNet. The complete set of performances are summarized in Table 1.

Conclusions : RETFound demonstrated competitive performance to a pretrained ResNet model. Addition of clinical tabular data improved performance, suggesting OCT images alone may not provide enough information to form the best predictor of 1-year BCVA treatment outcome for DME. Further explorations of other multimodal approaches are needed to understand the contributions from each data modality and improve performance robustness.
1. Zhou Y et al. Nature. 2023;622:156-63.

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

 

Fig 1. Hyperparameter selection based on average performance across validation folds

 

Table 1. Performance summary of experiments