Predicting Visual Improvement After Macular Hole Surgery: A Combined Model Using Deep Learning and Clinical Features

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Predicting visual acuity (VA) improvement after macular hole (MH) surgery is crucial for patient outcomes.
• High-definition optical coherence tomography (HD-OCT) and clinical data are key factors in assessing surgical success.

Purpose of the Study:

• To evaluate the feasibility of deep learning (DL) methods for predicting VA improvement post-MH surgery.
• To assess the combined predictive performance of DL on HD-OCT B-scans and clinical features.

Main Methods:

• A deep learning convolutional neural network (CNN) was trained on pre-operative HD-OCT B-scans.
• Clinical features were modeled using logistic regression.
• A hybrid model combined DL predictions with clinical data to predict VA increase (≥15 ETDRS letters at 6 months).

Main Results:

• The clinical features model achieved an AUROC of 80.6 and F1 score of 79.7.
• The CNN model alone showed an AUROC of 72.8 ± 14.6 and F1 score of 61.5 ± 23.7.
• The hybrid model yielded an AUROC of 81.9 ± 5.2 and F1 score of 80.4 ± 7.7, showing no significant improvement over the clinical model alone.

Conclusions:

• Both HD-OCT based DL models and clinical data models demonstrate good predictive performance for VA improvement after MH surgery.
• Fusing DL predictions with clinical data did not significantly enhance the overall predictive accuracy.