The Use of Artificial Intelligence to Predict Visual Outcome in Patients With Macular Pucker
Stijn Cornelis1, Ivo De Clerck1,2, Bart De Clerck3
1University Hospitals Leuven, Leuven, Belgium.
Purpose:
To assess the current role of artificial intelligence (AI) to predict postoperative best-corrected visual acuity (BCVA) after macular pucker surgery using optical coherence tomography (OCT) data.
Methods:
A systematic search was conducted of PubMed, EMBASE, Web of Science, Scopus, and the Cochrane Library. Data extraction included study design, demographics, AI model characteristics, and performance metrics. Study quality was assessed using the Prediction Model Risk of Bias Assessment + Artificial Intelligence tool, and results were synthesized narratively.
Results:
Six studies involving 2728 patients (65-71 years) were analyzed. Convolutional neural networks and gradient boosting decision trees were applied to predict postoperative BCVA. The convolutional neural networks ResNet-101 and Inception-ResNet-V2 outperformed human grading, with ResNet-101 achieving the highest area under the receiver operating characteristic curve (0.97). Interpretability techniques identified ectopic inner foveal layers (EIFL) and irregular inner nuclear layer-outer plexiform layer as key predictors.
Conclusions:
Deep learning models demonstrate potential in predicting postoperative outcomes from preoperative OCT scans after macular pucker surgery. Advanced architectures such as ResNet-101 and Inception-ResNet-v2 outperform simpler models and expert grading. However, clinical translation depends on model generalizability, interpretability, real-time application, and inclusion of functional data. Future research should emphasize multicenter validation to ensure broader applicability.
